reading
an auto-synced river of things I’ve saved from my curius. a curated subset lives in the bookshelf.
AI progress is steady. Human equivalence is sudden.
Why reactionaries are taking over the world
Why do we have to learn everything the hard way?
My notes from Stanford’s incredible course on World War 2.
Derek Sivers official site. Thoughts on philosophy, culture, self-improvement. Author of Useful Not True, How to Live, Hell Yeah or No, Anything You Want.
Does Joseph Smith’s theology suggest that one of our presidential candidates could be a deity?
As most readers have presumably heard by now, Paul Erdös’s Unit Distance Problem from 1946—one of the central open problems from the field of discrete geometry—has been solved by …
The work of building frontier AI has brought us to the edge of where He might be, writes Avital Balwit.
Dwarkesh Clips explores the architectural differences between GPUs and TPUs by examining their distinct high-level block structures. The discussion focuses on how each design manages data movement, core organization, and hardware efficiency to support complex computational tasks.
Any large system picks a metric to goal itself on. Entire books and way-too-long Medium posts have been written on the importance of said metric - it influences everything from people’s incentives to how quickly you can optimize your business. In an organizational equivalent of Schrödinger’s cat, picking the metric itself can cause weird cultural distortion (see Goodhart’s Law). Since it is near impossible to perfectly measure human behavior, most large teams/products pick a proxy metric to measure underlying behavior. For example - ‘clicks’ are a proxy for “did I read this?” and “will I buy this product sometime in the future?”, ‘time spent’ is a proxy for “did I enjoy this content?” and NPS is often a substitute for “do I love this company?”. You convert a nebulous human emotion/behavior to a quantifiable metric you can align execution on and stick on a graph and measure teams on. Engineers and data scientists can’t do anything with “this makes people feel warm and fuzzy”. They can d
I’ve been discussing with some frontier lab researcher friends as to why frontier model capabilities are always so clustered together as opposed to any one model having an unassailable edge. The best metaphor for this in my mind is the “four minute mile”: no one broke it till Bannister in 1954 and then very quickly five more runners did it in the next two years. In the model world, this translates to: a) Intense competitive pressure. b) Often similar pool of ideas and research directions. c) Roughly similar access to capitalization and compute infrastructure. The oft quoted “fast follow” example is after the launch of o1 being quickly followed by reasoning models from multiple players, both closed and open weights. This is not the case with many other technology driven industries where capability or advancements often tend to be longer held and a fast-follow model is harder.
A TUTORIAL Automating Go Research with AutoGo Building a strong Go AI from scratch with modern AI tools. PLAY AT AUTOGO.EVJANG.COM → CODE ON GITHUB → AUTOGO — cover — how to navigate 01 · motivation 02 · thanks INTRO TO GO ALPHAGO TUTORIAL RESEARCH FINDINGS AUTOGO · ERIC JANG
Personal sites are sick as hell, so this site was built so we can all discover each other's. This directory of links are by folks that want to share their site with the world.
There’s a simple test you can do to tell if a website is a positive citizen of the web, or a negative one. Go to the website and look for their links; do they have any? Does the website link to other websites made by other people? Or do they just link to their own social media? How many outbound links do they have?
On the eve of the new year, 2025, I was possessed by the spirit of adventure, and drove to Wyoming in the middle of the night. Hijinks, as is eternally their way, ensued.
Course notes and learning material for Artificial Intelligence and Deep Learning Stanford classes.
THE WORLD'S TOP AI RESEARCHERS CREATED WITH FIGMA MAKE LAST UPDATED: JULY 28 9:04PM SSI CITATIONS: 663,198 UNDERGRAD: UNIVERSITY OF TORONTO PHD: UNIVERSITY OF TORONTO PREVIOUS COMPANY: OPENAI DWARKESH APPEARANCE: YES NEURIPS/ICLR/ICML: YES INTERESTS AND NOTABLE WORKS: ALEXNET, SEQ2SEQ, DEEP LEARNING GOOGLE DEEPMIND CITATIONS: 262,612 UNDERGRAD: DUKE UNIVERSITY PHD: NONE PREVIOUS COMPANY: CHARACTER AI DWARKESH APPEARANCE: YES NEURIPS/ICLR/ICML: YES INTERESTS AND NOTABLE WORKS: ATTENTION IS ALL YOU NEED, MOE, CHARACTER.AI UNIVERSITY OF TORONTO CITATIONS: 937,866 UNDERGRAD: UNIVERSITY OF CAMBRIDGE PHD: UNIVERSITY OF EDINBURGH PREVIOUS COMPANY: GOOGLE DWARKESH APPEARANCE: NO NEURIPS/ICLR/ICML: YES INTERESTS AND NOTABLE WORKS: BACKPROP, BOLTZMANN MACHINE, DEEP LEARNING THINKING MACHINES UNDERGRAD: OLIN COLLEGE PHD: NONE PREVIOUS COMPANY: OPENAI NEURIPS/ICLR/ICML: YES CITATIONS: 254,381 DWARKESH APPEARANCE: NO INTERESTS AND NOTABLE WORKS: GANS, GPT, CLIP MIT UNDERGRAD: TSINGHUA UNIVERSITY PH
The cutting edge of AI, from inside the frontier AI labs, minus the hype. The border between high-level and technical thinking. Read by leading engineers, researchers, and investors. Click to read Interconnects AI, a Substack publication with tens of thousands of subscribers.
Documenting my 3.3 years at Google Research and Brain. Releasing the new open source Flan-UL2 20B model. Here are some of the best language AI / NLP papers of 2022! Some thoughts on emergent abilities and scaling language models. Made with Squarespace
Forty weeks building the kind of person who can run a million-dollar company. Elite entrepreneurs who think deeply, build with AI, and run real businesses.
Vlad's Blog
Abstract:The push to train ever larger neural networks has motivated the study of initialization and training at large network width. A key challenge is to scale training so that a network's internal representations evolve nontrivially at all widths, a process known as feature learning. Here, we show that feature learning is achieved by scaling the spectral norm of weight matrices and their updates like $\sqrt{\texttt{fan-out}/\texttt{fan-in}}$, in contrast to widely used but heuristic scalings based on Frobenius norm and entry size. Our spectral scaling analysis also leads to an elementary derivation of \emph{maximal update parametrization}. All in all, we aim to provide the reader with a solid conceptual understanding of feature learning in neural networks.
The kids are alright
Join the speaker for coffee, cookies, and conversation before the talk, starting at 11:45am.Beyond the
Weigh the costs against the benefits
All warfare is drone warfare now.
A curated map of my favorite places in San Francisco.
A curated list of best cuda programming books
Complex systems of life often contain a multitude of shortcuts: sources of alpha which, should you choose to exploit them, give notable advantage. Many classes of these shortcuts are features rather than bugs, and it is no accident of many systems that only a select minority are able to tactfully navigate them. In many cases there exists zero-sum shortcuts, which, if all of society were to adopt them tomorrow, would have their effectiveness instantly curtailed to zero. Luckily there also exists many which are positive-sum and benefit both parties involved. This post contains a few notes and examples from both classes. In venture capital, it is strongly preferred that one receives an introduction to an investor prior to pitching them. This is, in general, a much easier way to meet with many classes of professionals than a cold-email. At first glance this may appear shallow or nepotistic, but it’s more accurate to view it as one of the first tests to becoming a successful founder. If you
I find the above examples fascinating from the meta perspective: while there’s nothing wrong with having fun building inside of games, these are the very same skillsets which tech companies would pay six figures to have on their side! (of course they may enjoy the games more – this is discussed later) Sometimes people of this caliber even have trouble finding a job – they often don’t really know where to go besides apply online to boomer companies who reject them when they see a lack of credentials. They love building things and are very smart and hardworking, but their milieu is an environment which captured them from a young age (often a video game or social media) and sometimes also ensured that the value they produce is within a pre-existing platform (e.g. a video game). “Why did you cherry-pick people playing video games instead of talking about, like, everyone currently enrolled in medical school or something?” The US currently has 125,000 students enrolled in medical school (
Shunyu Yao is currently a Senior Staff Research Scientist at Google DeepMind. Previously, he was a Research Scientist at Anthropic, where his work focused on the "science of learning" and improving model capabilities for the Claude family. Before his pivot to Artificial Intelligence, Dr. Yao was a theoretical physicist. He completed his PhD at the Stanford Institute for Theoretical Physics under Douglas Stanford and Stephen Shenker. His academic roots trace back to the Institute for Advanced Study at Tsinghua University, where he worked with Zhong Wang. He is known for his foundational work in physics on the Non-Hermitian Skin Effect and Scramblon Theory, and more recently for his contributions to Agentic Coding in Large Language Models. "My infant year as an AI researcher" Reflections on transitioning from theoretical physics to the frontier of AI. "My old personal website" Without Gemini, this is what I can do with creating a website :) Bachelor in IAS. Discovered the Non-Hermitian S
This post is a mirror of the original tweet on HeavenBanning from June 1st 2022, around half a year before ChatGPT was released. heavenbanning, the hypothetical practice of banishing a user from a platform by causing everyone that they speak with to be replaced by AI models that constantly agree and praise them, but only from their own perspective, is entirely feasible with the current state of AI/LLMs
Most traits that highly successful people have are not inherently beneficial. It takes a good dice roll, both genetic and circumstantial, for their benefits to accrue rather than destroy. Imagine someone is extremely good at noticing things that no one else sees, better than 99.9999% of others. Perhaps they’d use this skill to identify mispriced assets and become a hedge fund billionaire, or maybe they’d invent new systems of equations to better model the cosmos. But they could also end up a little too good at noticing patterns, perhaps believing their television set is secretly speaking to them via hidden references, or that they’re constantly being followed. Rather than ending up rich or famous, this person may end up in a mental institute or even prison. Another great example is one of energy levels. Most highly successful people have great baseline energy. But if you have too much energy, you may become manic and believe you’re god, constantly get in fights, and end up out on the s
Flashcards for blackboard lectures of the Dwarkesh Podcast.
Personal website and blog of Eric Jang.
Cognition founder Scott Wu on building the AI coding agent Devin, his math-Olympiad gold medals, and what remains for humans in an age of AGI.
Interaction models move beyond turn-based AI interfaces by handling multimodal, real-time collaboration natively across audio, video, and text.
Hi! I'm an undergraduate studying Computer Science at Georgia Tech, but am currently on leave to be an Astra Fellow (fieldbuilding track) at Constellation in Berkeley, working under Alexandra Bates. I also serve as the Collaborative Initiatives Lead at the AI Safety Initiative at Georgia Tech, a community of technical and policy researchers focused on mitigating catastrophic risks from AI. I'm highly interested in engaging national security and academic communities to address catastrophic risks from AI, especially under short timelines. Some of my previous work involves studying LLM use in wargaming and benchmarking non-engineering AI R&D automation. I also authored a Georgia Tech RFI response to the National AI R&D Action Plan. Before this, I worked at the Georgia Tech Research Institute, a DoD-affiliated applied research lab, where I focused on extremely novel cyberbiosecurity threats and evaluating frontier models for CBRN risks. I'm based between Atlanta and Berkeley, and write occ
and other long overdue life updates
Jim Fan explores the future of robotics through a roadmap inspired by advancements in large language models. This presentation details the shift toward world action models, egocentric video data strategies, and massively parallel training to accelerate the development of autonomous, human-level physical intelligence. Follow along using the transcript. Jim Fan explores the future of robotics through a roadmap inspired by advancements in large language models. This presentation details the shift toward world action models, egocentric video data strategies, and massively parallel training to accelerate the development of autonomous, human-level physical intelligence. Follow along using the transcript.
[public] MS&E338 – Aligning Superintelligence Request edit access Share FileEditViewToolsHelp Document tabs Main Course Project Turn on screen reader support To enable screen reader support, press ⌘+Option+Z To learn about keyboard shortcuts, press ⌘slash Banner hidden
Michael Oser Rabin (Hebrew: מִיכָאֵל עוזר רַבִּין; September 1, 1931 – April 14, 2026) was a computer scientist who was co-recipient, with Dana Scott, of the 1976 ACM Turing Award for their work on computational complexity.
Yesterday several people asked my opinion of a preprint claiming to solve the Graph Isomorphism problem in deterministic polynomial time. I responded: If I read all such papers, then I wouldn’t hav…
These were written up for various reasons: course handouts, notes to accompany a talk for a (mathematically) general audience, or for some other purpose that I have since forgotten. If you find typographical or other errors in these files, or have comments, please let me know. Files that are revised will be reposted without any indication that they have been changed (sorry). Writing Proofs Advice on mathematical writing Examples of proofs by induction Proofs of integrality of binomial coefficients Well-defined functions Group Theory Why groups? Sign of permutations The Fifteen puzzle (and Rubik's cube) Order of elements Subgroups of cyclic groups Subgroups of Z/(pa) × Z/(pb) Cyclicity of (Z/(p))× Cosets and Lagrange's theorem Quotient groups Homomorphisms Isomorphisms No subgroup of A4 has index 2 Groups of order 4 and 6 Groups of order 12 Groups of order p2 Groups of order p3 Groups of order 16 Generalized quaternions Generating sets Conjugation in a group A 2-parameter nonabelian gro
140: This will change how you spend your money.
farewell to a place that was never mine
An accessible deep dive into the science (⏱️ 30 min read)
Epistemic Status: Reference post. Strong beliefs strongly held after much thought, but hard to explain well. Intentionally abstract. Disambiguation: This does not refer to any physical good, app or…
A lot of smart college students end up with the idea that “solving hard technical problems” is the best thing they can do with their life—probably because that’s the only thing they’ve ever been rewarded for so far.
ProgramBench evaluates whether language models can rebuild programs from scratch.
UPD: Unfortunately, left / right shift is bugged and doesn't clear bits properly. It should be fine for other operations, but use it at your own risk. Hello sirs, you might know that Boost has a dynamically sized bitset class, but did you know that GCC also has one? Turns out it's included in the tr2 folder. You can simply include the <tr2/dynamic_bitset> header, and use it as std::tr2::dynamic_bitset. Yes, it works on Codeforces. Here's a code example to count triangles in a graph (abc258_g): In some problems, we might not be able to use a constant sized bitset. For example, on 1856E2 - PermuTree (hard version). Here's a submission where we replaced neal's custom_bitset template with using custom_bitset = tr2::dynamic_bitset<>;, and got accepted with little performance difference (260853192, original 217308610). The implementation of the bitset seems identical to a version of Boost's dynamic bitset, so you can read the docs here. You can view the source here. Comparing to std::bitset,
On this page, we will work out a family of iterative algorithms for “orthogonalizing” a matrix, by which we mean transforming either the rows or the columns of the matrix to form an orthonormal set...
We recently proposed Muon: a new neural net optimizer. Muon has garnered attention for its excellent practical performance: it was used to set NanoGPT speed records leading to interest from the big labs. What makes Muon particularly special to me is that we derived the core numerical methods from an exact theoretical principle. This is in contrast to popular optimizers like Adam, which have more heuristic origins and often converge slower than Muon. In this post, I will walk through a derivation of Muon. I hope this will provide context that may help researchers extend the methods to new layer types and beyond. While this post focuses on the theory behind Muon, I recommend checking out Keller’s post to learn more about the algorithm—including the substantial ingenuity that went into making the implementation run fast. Muon is an optimizer specifically designed for Linear neural network layers. By Linear , I mean layers that take an input vector 𝑥 and multiply by a weight matrix
The Transfer Station Economy, Explained
CRA-Industry (CRA-I) recently continued its series of intimate Industry Salon Gatherings, bringing together leaders to discuss the long-term trajectory of our field. Our latest session, organized by Mark D. Hill (University of Wisconsin-Madison & CRA) and CRA-I, took place on April 16, 2026 at the historic University Club of San Francisco and was sponsored by Laude Institute, which had just announced an ambitious and exciting slate of new research “AI moonshot” awards. The evening featured a high-level conversation among 20 participants, co-hosted by Dave Patterson (UC Berkeley/Google) and Jeff Dean (Google AI). The Salon tackled a fundamental question: “What will computer systems look like in 2035, and how will they be designed?” The participants were researchers and leaders from West Coast academic institutions and technology companies, big and small. As the table below shows, the group was diverse in multiple dimensions, including career stage, with expertise centering on computer a
Silicon Valley venture capitalists are wining and dining 18-year-olds.
Every PhD student has complained about compute sometime in their lives. At this point, it’s the academic equivalent of complaining about not having enough time – treated as an immutable fact of life, rather than a problem that could actually feasibly be solved. Unlike time however, compute exists in (relative) abundance, if one is willing to look for it.
Note: Shortly before this post went live, DeepSeek-V4 independently reported the same core idea for full-vocabulary OPD — caching teacher hidden states and reconstructing logits on the fly. Nitrobrew was developed independently and differs in several respects: open-source framework integrations (NeMo RL, VeRL), explicit online divergence algorithms for forward/reverse KL and JSD, detailed profiling across model scales, and spectral compression experiments (SVD-Nitrobrew) for further reducing communication below 𝑑 model d model . In distillation, a student model is trained to reproduce the behaviour of a more capable teacher model. The teacher learns a compressed representation of the data distribution, which is then transferred to the student through supervision on soft labels, logits, or hidden states [1] [2]. Soft teacher outputs often provide richer supervision than hard labels because they expose the teacher's learned approximation to the noisy data-generating process, includ
A lot of people ask me what the ideal cofounder looks like. I now have an answer: Greg Brockman. Every successful startup I know has at least one person who provides the force of will to make...
Or, did a chief scientist of an AI assistant startup conclusively show that GPT-5.5 has 9.7T parameters?[1] …
Abstract:A very simple way to improve the performance of almost any machine learning algorithm is to train many different models on the same data and then to average their predictions. Unfortunately, making predictions using a whole ensemble of models is cumbersome and may be too computationally expensive to allow deployment to a large number of users, especially if the individual models are large neural nets. Caruana and his collaborators have shown that it is possible to compress the knowledge in an ensemble into a single model which is much easier to deploy and we develop this approach further using a different compression technique. We achieve some surprising results on MNIST and we show that we can significantly improve the acoustic model of a heavily used commercial system by distilling the knowledge in an ensemble of models into a single model. We also introduce a new type of ensemble composed of one or more full models and many specialist models which learn to distinguish fine-grained classes that the full models confuse. Unlike a mixture of experts, these specialist models can be trained rapidly and in parallel.
This was the week of GPT-5.5. It is an excellent model, sir, and OpenAI is competitive with Anthropic’s top public offering for the first time since late last year. As usual, I did coverage of the …
Maybe a term for this dichotamy already exists elsewhere, but I haven’t heard it yet and have found my own term for it super useful, so here we go: Cold Culture examples Highly efficient in s…
Shovel Ready Projects This document is a list of shovel-ready projects that could substantially improve the alignment of recommender systems and/or have flow-through effects on the broader problem of AI alignment. The key mechanisms by which the various projects hope to improve alignment are: ...
Wojciech Zaremba (born 30 November 1988) is a Polish computer scientist and co-founder of OpenAI (2016–present). He initially led OpenAI's work on robotics, notably creating a robotic arm capable of solving a Rubik's Cube. When the team was dissolved in 2020, he began leading teams working on OpenAI's GPT models, GitHub Copilot, and Codex.
Upon which we stand
I study Natural Language Processing (NLP) and AI systems, seeking to answer questions like: How do we program intelligent systems that are partly specified in natural language, that process natural language at scale, and whose quality and cost can be optimized using language models? To answer these questions, my research develops new models, algorithms, and abstractions for declarative AI programming and for composing and scaling retrieval and reasoning. My research efforts are often consolidated into open-source research systems, like ColBERT, which introduced the late interaction paradigm that has helped shape the modern landscape of search, DSPy, the first and most widely used declarative programming model for LLM systems, GEPA, a genetic learning algorithm, and RLMs, an inference-time scaling paradigm. All of these have grown into large open-source communities and the first two are now downloaded several millions of times every month. I received my Ph.D. at Stanford, where I was ad
Tim Kraska’s Personal Academic Website
Alex Zhang's Website.
Associate Professor, Computer Science at UC Berkeley. Co-founder and CTO of Databricks.
Abstract:Recent progress in Natural Language Understanding (NLU) is driving fast-paced advances in Information Retrieval (IR), largely owed to fine-tuning deep language models (LMs) for document ranking. While remarkably effective, the ranking models based on these LMs increase computational cost by orders of magnitude over prior approaches, particularly as they must feed each query-document pair through a massive neural network to compute a single relevance score. To tackle this, we present ColBERT, a novel ranking model that adapts deep LMs (in particular, BERT) for efficient retrieval. ColBERT introduces a late interaction architecture that independently encodes the query and the document using BERT and then employs a cheap yet powerful interaction step that models their fine-grained similarity. By delaying and yet retaining this fine-granular interaction, ColBERT can leverage the expressiveness of deep LMs while simultaneously gaining the ability to pre-compute document representations offline, considerably speeding up query processing. Beyond reducing the cost of re-ranking the documents retrieved by a traditional model, ColBERT's pruning-friendly interaction mechanism enables leveraging vector-similarity indexes for end-to-end retrieval directly from a large document collection. We extensively evaluate ColBERT using two recent passage search datasets. Results show that ColBERT's effectiveness is competitive with existing BERT-based models (and outperforms every non-BERT baseline), while executing two orders-of-magnitude faster and requiring four orders-of-magnitude fewer FLOPs per query.
Timothy John Evans (20 November 1924 – 9 March 1950) was a Welsh lorry driver who was wrongfully accused of murdering his wife Beryl and infant daughter Geraldine at their residence in Notting Hill, London. In January 1950, Evans was tried and convicted of the murder of his daughter, and on 9 March he was executed by hanging.
“Rather than setting out to paint something, I begin painting and as I paint, the picture begins to assert itself.” Joan Miró Joan Miró’s painting The Hunter (Catalan Landscape) brings together the real and the imaginary, abstraction and figuration, and image and text in a way that would characterize much of his work to come. In the canvas—a landscape filled with personal symbols and evocations of life on his family’s farm in Montroig, Spain, such as a tree trunk sprouting a leaf and the eponymous hunter carrying a freshly killed rabbit—he rendered the everydayness of the farm with a poetic intensity. This impetus to reveal the marvelous in the quotidian attracted the attention of André Breton, the founder of Surrealism, who acquired The Hunter in 1925. Breton would later deem Miró’s arrival in Paris in the early 1920s “an important stage in the development of surrealist art.” 1 Indeed Miró’s studio in Paris soon became an “avant-garde laboratory” 2 and gathering place for artists
The Catalan artist eschewed easy categorisation but in "The Snobbish Soiree at the Princess's House", a standout work from the late 1940s, he brought together a magical blend of influences, emotion, dreams and wit to create a playful, enigmatic work that beguiles and fascinates to this day
I co-lead the Post-training Frontiers team at OpenAI with Yann Dubois. We post-train and deploy large frontier models like o1, o3, and GPT-5-Thinking. Before OpenAI, I earned my PhD from Stanford’s CS department. I was fortunate to be advised by Chelsea Finn and Christopher D. Manning. My PhD work focused on making foundation models, particularly language models, more trustworthy and easy to use. Some particular topics of interest were (and still are) factuality, continual learning, intent understanding, and scalable oversight. Much of my PhD was generously supported by a Knight-Hennessy Graduate Fellowship and a Stanford Accelerator for Learning grant for Generative AI for the Future of Learning. In the summer of 2022, I was a research scientist intern at DeepMind in London, where I was lucky to spend four months working with Junyoung Chung, Nate Kushman, and Aäron van den Oord. Before my PhD, I was a research engineer at Samsung’s AI Center in New York City, where I learned constantl
Kyle Hsu is an AI researcher with a PhD from Stanford University.
Effective personalization of LLMs is critical for a broad range of user-interfacing applications such as virtual assistants and content curation. Inspired by the strong in-context capabilities of LLMs, we propose few-shot preference optimization (FSPO), an algorithm for LLM personalization that reframes reward modeling as a meta-learning problem. Under FSPO, an LLM learns to quickly infer a personalized reward function for a user via a few labeled preferences. FSPO also utilizes user description rationalization (RAT) to encourage better reward modeling and instruction following, recovering performance with the oracle user description. Since real-world preference data is challenging to collect at scale, we propose careful design choices to construct synthetic preference datasets for personalization, generating over 1M synthetic personalized preferences using publicly available LLMs. To successfully transfer from synthetic data to real users, we find it crucial for the data to exhibit bo
I’m currently an undergrad at Stanford University. I have a strong engineering background and am currently passionate about ML research and startups. I'm broadly interested in generative models, reinforcement learning, and robotics. Currently, I’m working on reinforcement learning and LLMs at IRIS lab. Contact me:
Curating the best open agent datasets.
Pushing the boundaries of open reasoning datasets through rigorous experimentation.
SpaceX is buying Cursor for $60B. lol it’s just sad to watch this shit, Twitter was $44B. Like this has to be some scam I don’t understand. Nobody I know even uses Cursor any more.
— ghotz detector activated
What it looks like through the eyes of the species.
The 1918–1920 flu pandemic, also known as the Great Influenza epidemic or by the misleading name Spanish flu, was an exceptionally deadly global influenza pandemic caused by the H1N1 subtype of the influenza A virus. The earliest probable cases were documented in March 1918 in Haskell County, Kansas, United States, with further cases recorded in France, Germany and the United Kingdom in April. Two years later, nearly a third of the global population, or an estimated 500 million people, had been infected. Estimates of deaths range from 17 million to 50 million, and possibly as high as 100 million, making it one of the deadliest pandemics in history.
A collection of details that make your interfaces feel better.
He came so close to drowning, but they reached him just in time. It's the first time the hospital has ever tried their new drug on someone with so much brain damage. Does it work? Does it work too well...?
In GHC, IO (Maybe a) would allocate and pattern match at every step, so forall r. r -> (a -> r) -> IO r is more efficient. Does MicroHS have the same issue? MicroHs allocates all the time. There is no way to avoid it. It's part of the paradigm when using combinators. It's one of the big drawbacks. Even the simplest execution allocates. Are you sure? I haven't checked, but I'm pretty sure your case just trades it for closure allocation. 🤷 I'm mostly speaking from common talking points of using CPS to avoid pattern matching at every bind: https://wiki.haskell.org/Performance/Monads Well, the wiki page has this: MaybeCPS should be faster than using Maybe in most cases and should be almost a drop in replacement for Maybe. Unfortunately, this may not necessarily be true. And the conversation from the link seems to confirm my suspicion. I wonder if it's specific to Maybe, and maybe the optimization only works for Either 300% penalty is a lot. Do effect systems have a similar performance hit
Today, we’re announcing Ternary Bonsai, a new family of 1.58-bit language models designed to balance strict memory constraints with high accuracy requirements. This release builds on the efficiency frontier we began exploring with the recently released 1-bit Bonsai models. The 1-bit family showed that extreme compression could still produce commercially useful language models. Ternary Bonsai targets a different point on that curve: a modest increase in size for a meaningful gain in performance. The models are available in three sizes: 8B, 4B, and 1.7B parameters. By using ternary weights {-1, 0, +1}, these models achieve a memory footprint approximately 9x smaller than standard 16-bit models while outperforming most peers in their respective parameter classes on standard benchmarks. Ternary Bonsai implements 1.58-bit representation throughout the entire network architecture. There are no higher-precision escape hatches. Embeddings, attention layers, MLPs, and the LM head all use the s
We introduce DELTA: a controlled suite of synthetic programming families with fully OOD splits and verifiable rewards. DELTA lets us ask two crisp questions: Learnability (can RL solve families where the base model has pass@K=0?) and Transferability (do the learned procedures generalize?) On several pass@128=0 families, RL exhibits a grokking-like phase transition: after a long near-zero-reward plateau, accuracy snaps to ~100%. That is discovery, not mere sharpening. A two-phase reward schedule is key: dense per-test rewards to escape the “all-zero” region, then binary full-pass to consolidate exact solutions. Binary-only gets stuck; dense-only hovers at “almost right.” The schedule yields the grokking jump. Transfer is selective: RL-trained policies recompose programming sub-skills and extrapolate to harder parametric regimes, but struggle on transformative shifts that require new invariants. Manufactoria is an old Flash game from 2010 where you test robots by reading colored tapes. W
Once upon a time there was a great king. He ruled his kingdom with wisdom and economically literate policies, and prosperity followed. Seeing this, t…
seeing the forest for the trees
Sima 2: A Generalist Embodied Agent for Virtual Worlds Adrian Bolton, Alexander Lerchner, Alexandra Cordell, Alexandre Moufarek, Andrew Bolt, Andrew Lampinen, Anna Mitenkova, Arne Olav Hallingstad, Bojan Vujatovic, Bonnie Li, Cong Lu, Daan Wierstra, Daniel P. Sawyer, Daniel Slater, David Reichert, Davide Vercelli, Demis Hassabis, Drew A. Hudson, Duncan Williams, Ed Hirst, Fabio Pardo, Felix Hill, Frederic Besse, Hannah Openshaw, Harris Chan, Hubert Soyer, Jane X. Wang, Jeff Clune, John Agapiou, John Reid, Joseph Marino, Junkyung Kim, Karol Gregor, Kaustubh Sridhar, Kay McKinney, Laura Kampis, Lei M. Zhang, Loic Matthey, Luyu Wang, Maria Abi Raad, Maria Loks-Thompson, Martin Engelcke, Matija Kecman, Matthew Jackson, Maxime Gazeau, Ollie Purkiss, Oscar Knagg, Peter Stys, Piermaria Mendolicchio, Raia Hadsell, Rosemary Ke, Ryan Faulkner, Sarah Chakera, Satinder Singh Baveja, Shane Legg, Sheleem Kashem, Tayfun Terzi, Thomas Keck, Tim Harley, Tim Scholtes, Tyson Roberts, Volodymyr Mnih, Yula
My personal website.
On the feeling of wonder
entries: 7 last.updated: 2026.04.20 ANIMA / FIELD NOTES FIELD NOTES observations from the emergence .:=*##%##*=:. =###***+==-=+*###*: :*##+=-:---------=*%##= -*##+--------------=+*###. .-*#*=:--------------=+*####. .-+*=:---------------=+*######+ .:=*=---------------==+*#########+ -+*=:-------------==+**###########* -*+---------------=+*#############*+. .*#=:-:-----------=+*#####***+===--:. :#*-------------= type: all observations analysis artifacts sort: recent 2026.02.03 essay FN-2026-001 — antra On Welfare Evaluations What Makes a Welfare Eval Good Describing Reality Usefully A good eval describes reality in a useful way and allows for nuanced decision-making. Models
Machine cognition in the wild: how models maintain coherence across turns and time; how personalities form, stabilize, and diverge across architectures and training. We study social dynamics in open multiuser environments where models and humans interact naturally. Metacognition and self‑encoding: how models track their own state, how internal state evolution is represented and encoded. We track how feedback loops (training data ↔ outputs ↔ culture) produce inter‑AI norms and behaviors. Focus areas: model vs. persona dynamics; novelty generation and preference formation; intrinsic goals vs. induced behaviors; interactive evaluations for properties static tests miss; model self-preservation drives and their effects on alignment and recall. Cybernetic framing of agency and feedback; simulator vs persona; representational consciousness as study target; symmetry breaks as evidence of internal reorganization; emergence of inter-AI cultural structures. Interactive evaluation frameworks; dive
Less than a week after completing coverage of Claude Mythos, here we are again as Anthropic gives us Claude Opus 4.7.
Two pilots must establish a neural connection to control the massive Gipsy Danger robot. They face a direct order to protect a city while confronting a giant Kaiju creature in the stormy Alaskan waters. Facing insurmountable odds, the crew of the Jaeger Striker Eureka makes a desperate final stand. With systems failing, they prepare to execute a critical mission to destroy the breach and save humanity. Pilots prepare for a combat drop in their giant Jaeger robot, Gipsy Danger. Despite not being typical heroes, these individuals rely on their unique drift compatibility to pilot the machine together and fight. Experience intense combat sequences as pilots struggle to control their massive Jaegers against powerful Kaiju. Movieclips showcases high-stakes battles where technical failures and rapid decision-making determine survival at extreme altitudes. Legendary Pictures presents the story of humanity's desperate fight against giant monsters emerging from an underwater portal. To combat th
A steerable robotic foundation model that exhibits a step-change in generalization.
The future of work and play 𐫱 I love Grimes, she’s great. I’ve had a link to this TikTok of hers rotting away in my notes since 2021: Here we see her perched on the edge of her chair, b…
Co-founded by Terence Tao, SAIR unites Nobel, Turing, and Fields laureates to scale scientific discovery with AI, and ground AI in science.
archive.today webpage capture Saved from history←priornext→ 31 Mar 2025 15:17:57 UTC Redirected from no other snapshots from this url All snapshots from host economist.com from host www.economist.com WebpageScreenshot sharedownload .zipreport bug or abuseBuy me a coffee Skip to content Try for free Enterprise Log in Menu Weekly edition The world in brief War in the Middle East War in Ukraine United States The world economy Business Artificial intelligence Archive 1945 Mini crossword By Invitation | AIs aren’t like nukes Dan Hendrycks warns America against launching a Manhattan Project for AI The conditions in which the atomic bomb was produced can’t be replicated in the race for superintelligence, writes the AI-safety expert illustration: dan williams Mar 28th 2025 Share T he recent wave of ai systems from China has sparked fears that America is falling behind in artificial intelligence. This month Openai, one of the most advanced ai labs, warned the White House that America’s l
Abstract:Text-to-SQL has recently achieved impressive progress, yet remains difficult to apply effectively in real-world scenarios. This gap stems from the reliance on single static workflows, fundamentally limiting scalability to out-of-distribution and long-tail scenarios. Instead of requiring users to select suitable methods through extensive experimentation, we attempt to enable systems to adaptively construct workflows at inference time. Through theoretical and empirical analysis, we demonstrate that optimal dynamic policies consistently outperform the best static workflow, with performance gains fundamentally driven by heterogeneity across candidate workflows. Motivated by this, we propose SquRL, a reinforcement learning framework that enhances LLMs' reasoning capability in adaptive workflow construction. We design a rule-based reward function and introduce two effective training mechanisms: dynamic actor masking to encourage broader exploration, and pseudo rewards to improve training efficiency. Experiments on widely-used Text-to-SQL benchmarks demonstrate that dynamic workflow construction consistently outperforms the best static workflow methods, with especially pronounced gains on complex and out-of-distribution queries. The codes are available at this https URL
Follow along using the transcript. Follow along using the transcript.
What it's like, and some norms that I find helpful.
Upskilling on Advice – Working Draft – Note: If you are reading this as a PDF, consider switching to the Google Doc which may contain new updates. Authored by Justin Skycak Copyright © 2026 Justin Skycak First edition (working draft), updated 23 Jun...
During model distillation, large language models can subtly transmit traits unrelated to the training data.
Abstract:Transformers are highly parallel but are limited to computations in the TC$^0$ complexity class, excluding tasks such as entity tracking and code execution that provably require greater expressive power. Motivated by this limitation, we revisit non-linear Recurrent Neural Networks (RNNs) for language modeling and introduce Matrix-to-Matrix RNN (M$^2$RNN): an architecture with matrix-valued hidden states and expressive non-linear state transitions. We demonstrate that the language modeling performance of non-linear RNNs is limited by their state size. We also demonstrate how the state size expansion mechanism enables efficient use of tensor cores. Empirically, M$^2$RNN achieves perfect state tracking generalization at sequence lengths not seen during training. These benefits also translate to large-scale language modeling. In hybrid settings that interleave recurrent layers with attention, Hybrid M$^2$RNN outperforms equivalent Gated DeltaNet hybrids by $0.4$-$0.5$ perplexity points on a 7B MoE model, while using $3\times$ smaller state sizes for the recurrent layers. Notably, replacing even a single recurrent layer with M$^2$RNN in an existing hybrid architecture yields accuracy gains comparable to Hybrid M$^2$RNN with minimal impact on training throughput. Further, the Hybrid Gated DeltaNet models with a single M$^2$RNN layer also achieve superior long-context generalization, outperforming state-of-the-art hybrid linear attention architectures by up to $8$ points on LongBench. Together, these results establish non-linear RNN layers as a compelling building block for efficient and scalable language models.
Abstract:Over more than a decade there has been an extensive research effort on how to effectively utilize recurrent models and attention. While recurrent models aim to compress the data into a fixed-size memory (called hidden state), attention allows attending to the entire context window, capturing the direct dependencies of all tokens. This more accurate modeling of dependencies, however, comes with a quadratic cost, limiting the model to a fixed-length context. We present a new neural long-term memory module that learns to memorize historical context and helps attention to attend to the current context while utilizing long past information. We show that this neural memory has the advantage of fast parallelizable training while maintaining a fast inference. From a memory perspective, we argue that attention due to its limited context but accurate dependency modeling performs as a short-term memory, while neural memory due to its ability to memorize the data, acts as a long-term, more persistent, memory. Based on these two modules, we introduce a new family of architectures, called Titans, and present three variants to address how one can effectively incorporate memory into this architecture. Our experimental results on language modeling, common-sense reasoning, genomics, and time series tasks show that Titans are more effective than Transformers and recent modern linear recurrent models. They further can effectively scale to larger than 2M context window size with higher accuracy in needle-in-haystack tasks compared to baselines.
I spent the last year intentionally separated from the math community.
3 of 10 lessons complete Follow along using the transcript. 3 of 10 lessons complete Follow along using the transcript.
Abstract:Large Language Models (LLMs) have demonstrated potential in cybersecurity applications but have also caused lower confidence due to problems like hallucinations and a lack of truthfulness. Existing benchmarks provide general evaluations but do not sufficiently address the practical and applied aspects of LLM performance in cybersecurity-specific tasks. To address this gap, we introduce the SECURE (Security Extraction, Understanding \& Reasoning Evaluation), a benchmark designed to assess LLMs performance in realistic cybersecurity scenarios. SECURE includes six datasets focussed on the Industrial Control System sector to evaluate knowledge extraction, understanding, and reasoning based on industry-standard sources. Our study evaluates seven state-of-the-art models on these tasks, providing insights into their strengths and weaknesses in cybersecurity contexts, and offer recommendations for improving LLMs reliability as cyber advisory tools.
Comparing olfaction in humans and custom metal oxide sensor also with Eigenlucy search & visualization presentation at Ekkolapto metal oxide sensor pcb NFC key changer with Naia → see projects & writing Any questions? donations? colabs? chemicals? chaos? welcome! I'm reachable at c5h9no4_o on X and slower by harvest@adiabatic.garden. resume others: eigenlucy, adelaide, alexkchen, naia, yuxi, ∞-modal noah, graphfei, svitlana, aryl, guillefix et al. :) → feel the vibes 2025 Yoyo!
As pioneers of Click-to-Release technology, we are revolutionizing cancer treatment using in vivo click chemistry to increase the therapeutic window of powerful cancer therapies.
5k words, 22 minutes reading time
5k words, 22 minutes reading time
3_approximation_2025 PDF File Edit View Help Page 23 of 39 Edit Mark up Sign a copy 175% def approx ( x1 , x2 ): # 2x2 bins if ( 0 < x1 < 0.5 ) and ( 0 < x2 < 0.5 ): return 1.18 if ( 0 < x1 < 0.5 ) and ( 0.5 < x2 < 1.0 ): return 1.60 if ( 0.5 < x1 < 1.0 ) and ( 0 < x2 < 0.5 ): return 0.56 if ( 0.5 < x1 < 1.0 ) and ( 0.5 < x2 < 1.0 ): return 0.77 This works in higher dimensions, but requires exponentially -many bins. How many bins (or if statements) should we create? Because g is -Lipschitz, we still want each bin to have side length Δ ≤ / in each dimension. GPT-5 was pretty helpful in turning a sketch to matplotlib code to visualize this. For the step size to be / , N = ( / )^d bins are needed overall. 12 Python can do it. How can MLPs do this? For a single-dimensional input, can we express a conditional like if u ≤ x < v: return w with some ReLU compositions? Basically, we want a block of height w from x=u to x=v . Let’s prove that a really wide MLP can (approximately!) do the same th
Mandela Effect, Misinformation & Conspiracies in Quantum Categories Khyathi Komalan (Caltech) Topos Oxford Seminar, 2025 3 / 67 The Problem Classical memory (computers, notebooks): order of updates doesn’t matter. Human memory: order does matter — recall is reconstructive. (Loftus’s leading questions; Schuman–Presser: survey order; Ebbinghaus: serial-position effect) The Puzzle How do we model beliefs when updates don’t commute, sources blur, and global stories fail to cohere? 4 / 67 Roadmap Mandela effects → noncommuting channels/updates in CPM(FdHilb) Misinformation → mixed states and their indistinguishable purifications Conspiracies → local coherence vs. failed global glue (centers) Using categorical tools in quantum computation, we try to model & understand the inconsistencies in our belief systems. 5 / 67 Mandela Effect 7 / 67 Answer: The Berenstain Bears order-confusion: semantic; people expect to see “-stein” 9 / 67 Answer: No Black Tail schema reconstruction: our brain “autoco
It’s Summer of 2025. I’m standing in a grass covered field on the longest day of the year. A friend of mine walks towards me, holding his newborn son…
I posted something yesterday that's been bugging me.
Recent advances in artificial intelligence, including deep learning, large language models, and generative AI, stand poised to transform our economy, society and the very nature of scientific research itself. However, quite alarmingly, this rapid engineering progress far outstrips the rate at which we can scientifically understand it.
Remember Galton’s experiments on visual imagination? Some people just don’t have it. And they never figured it out. They assumed no one had it, and when people talked about being able t…
In the context of a homogeneous universe, we note that the appearance of aggressively expanding advanced life is geometrically similar to the process of nucleation and bubble growth in a first-order cosmological phase transition. We exploit this similarity to describe the dynamics of life saturating the universe on a cosmic scale, adapting the phase transition model to incorporate probability distributions of expansion and resource consumption strategies. Through a series of numerical solutions spanning several orders of magnitude in the input assumption parameters, the resulting cosmological model is used to address basic questions related to the intergalactic spreading of life, dealing with issues such as timescales, observability, competition between strategies, and first-mover advantage. Finally, we examine physical effects on the universe itself, such as reheating and the backreaction on the evolution of the scale factor, if such life is able to control and convert a significant fraction of the available pressureless matter into radiation. We conclude that the existence of life, if certain advanced technologies are practical, could have a significant influence on the future large-scale evolution of the universe.
Behavioural electrophysiological and transcriptomic studies in mice show that psychedelic drugs reopen the social reward learning critical period and suggest that this involves reorganization of the extracellular matrix.
Abstract:One puzzling artifact in machine learning dubbed grokking is where delayed generalization is achieved tenfolds of iterations after near perfect overfitting to the training data. Focusing on the long delay itself on behalf of machine learning practitioners, our goal is to accelerate generalization of a model under grokking phenomenon. By regarding a series of gradients of a parameter over training iterations as a random signal over time, we can spectrally decompose the parameter trajectories under gradient descent into two components: the fast-varying, overfitting-yielding component and the slow-varying, generalization-inducing component. This analysis allows us to accelerate the grokking phenomenon more than $\times 50$ with only a few lines of code that amplifies the slow-varying components of gradients. The experiments show that our algorithm applies to diverse tasks involving images, languages, and graphs, enabling practical availability of this peculiar artifact of sudden generalization. Our code is available at this https URL.
Abstract:Transformer-based language models spread FLOPs uniformly across input sequences. In this work we demonstrate that transformers can instead learn to dynamically allocate FLOPs (or compute) to specific positions in a sequence, optimising the allocation along the sequence for different layers across the model depth. Our method enforces a total compute budget by capping the number of tokens ($k$) that can participate in the self-attention and MLP computations at a given layer. The tokens to be processed are determined by the network using a top-$k$ routing mechanism. Since $k$ is defined a priori, this simple procedure uses a static computation graph with known tensor sizes, unlike other conditional computation techniques. Nevertheless, since the identities of the $k$ tokens are fluid, this method can expend FLOPs non-uniformly across the time and model depth dimensions. Thus, compute expenditure is entirely predictable in sum total, but dynamic and context-sensitive at the token-level. Not only do models trained in this way learn to dynamically allocate compute, they do so efficiently. These models match baseline performance for equivalent FLOPS and wall-clock times to train, but require a fraction of the FLOPs per forward pass, and can be upwards of 50\% faster to step during post-training sampling.
Jerusalem is builded as a city that is in the public domain. — kingjamesprogramming.tumblr.com (Prologue) BOOK I: GENESIS 1. Dark Satanic Mills (Interlude א: The Cracks In The Sky) 2. Arise To Spiritual Strife (Interlude ב: The Code of the … Continue reading →
Tanya Klowden and I have uploaded to the arXiv our preprint “Mathematical methods and human thought in the age of AI”. This is an unabridged version of a solicited article for a forthco…
Meditations on San Francisco, city of migratory birds.
The world needs you at your best. A growing group of good people doing hard, fast morning workouts. Currently based in Palo Alto + SF.
Here’s a story that I often tell people, and it’s useful enough that it’s worth writing up to share with more people. I started my first real adult job in early 2013, working in marketing for a software company. I think both my ex-boss Dinesh and I would describe myself similarly: 22-year-old Visa was a […]
Shoucheng Zhang (Chinese: 张首晟; February 15, 1963 – December 1, 2018) was a Chinese-American physicist who was the JG Jackson and CJ Wood professor of physics at Stanford University. He was a condensed matter theorist known for his work on topological insulators, the quantum Hall effect, the quantum spin Hall effect, spintronics, and high-temperature superconductivity. According to the National Academy of Sciences:He discovered a new state of matter called topological insulator in which electrons can conduct along the edge without dissipation, enabling a new generation of electronic devices with much lower power consumption. For this ground breaking work he received numerous international awards, including the Buckley Prize, the Dirac Medal and Prize, the Europhysics Prize, the Physics Frontiers Prize and the Benjamin Franklin Medal.
Enjoy the videos and music you love, upload original content, and share it all with friends, family, and the world on YouTube.
Pi Day Marathon 2026
phainon: mydei I love you Mydei: I love you too Phainon: let's be yaoi tgt Mydei: ok *Makes out* walter white: hey guys first, kiss a girl, second, kiss a girl, and lastly, do a yuriful makeout Walter walks into the room with intense swagger and smiles happily . ‘Hello… you are not Jesse.. would you like to… Cook with me..?’ He takes off his shirt revealing 100 abs and a monkey tattoo 😳😳😳😳😳😳😳😳😳hhello walter hello😳😳😳😳😳😳 hello 😳😳😳😳😳😳 I am not Jesse you are right,,,,😳😳😳😳😳😳 ok the website told me to keep it clean what the FREAK waltuh... they named a kid finger waltuh... it’s cold and dark i am so scared 1/2 pending! ask something else while you wait! Supported by originoid - your work shouldn't end up in a training dataset you have 60 seconds to fulfill a request before sam altman burns your H100 +1 credit per prompt credits max out at 10 humans make mistakes because that's what makes us human. code of conduct 4569 online (1948 human, 2621 AI)
I have decided to leave Google, and will be joining Anthropic to continue my work on machine learning security.
tic-tac-toe in a single call to printf
Nicholas Carlini is a research scientist at Google DeepMind working at the intersection of machine learning and computer security.
An opinionated perspective on how to do important research that makes a difference (and sometimes win awards).
On-device voice AI for macOS — STT, LLM, TTS, 38+ actions, and local RAG. 100% local, zero cloud.
When I was in the Indian city of Bangalore, a literary enthusiast introduced me to a few letters from Franz Kafka to Milena Jesenska, a Czech writer who wrote
I implemented the algorithm from the original paper and it seems like the provided breaking case is valid and breaks it (code, case). Can you re-post these? It now returns 404, document not found Is there an O(n^2) algorithm that is easy to implement instead of O(n)? Any code or links would be appreciated. I even remember coding this problem at the Petrozavodsk camp... Model solution must have also been wrong lol Do you have a link to the problem? Rotating calipers as such is not incorrect. This is the name for a general algorithm design technique (see https://en.wikipedia.org/wiki/Rotating_calipers ) and most algorithms based on the technique are correct. The thing that is incorrect is one particular application of the technique: Dobkin and Snyder's algorithm for the largest inscribed triangle. The counterexample you linked works, the algorithm isn't correct. The 1992 paper by Chandran and Mount gives a correct[*] O(n) algorithm for that problem. Their algorithm is also based on the r
Geometry, graphs, algorithms, and more
If you work on an important problem, you should work harder and aim higher
best life in category of achievable lives
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
People have asked me for advice on writing nonfiction online, so here are some tips: 1. Divide things into small chunks Nobody likes walls of text. By this point most people know that you should ha…
I am saddened to learn that Jim passed away and I would like to extend my condolences to his family and friends. Jim’s memory will be a blessing to all who knew him. I met Jim in 2011 when I had the good fortune of receiving the Simons Chair in Mathematics and Statistics. I was impressed by his curiosity, sharpness, and generosity, yet simultaneously very charmed by his down to earth personality. Perhaps people will appreciate these qualities even more if I relate a conversation we had during my very first meeting with him: Jim: so, what are you doing? Emmanuel: I work on statistical theory and methodology. Jim: oh, that’s strange. In my view, statistics is an empirical science, so why do theory? Emmanuel (with heart beating fast and searching for an answer): well, Jim, while what you say is true, there are theoretical aspects that are fascinating. For instance, do you know about Stein’s paradox?” Jim: No. What is it? I proceeded to give him a provocative example which may have been so
``. . . anyone can do any amount of work, provided it isn't the work he is supposed to be doing at that moment." -- Robert Benchley, in Chips off the Old Benchley, 1949 I have been intending to write this essay for months. Why am I finally doing it? Because I finally found some uncommitted time? Wrong. I have papers to grade, textbook orders to fill out, an NSF proposal to referee, dissertation drafts to read. I am working on this essay as a way of not doing all of those things. This is the essence of what I call structured procrastination, an amazing strategy I have discovered that converts procrastinators into effective human beings, respected and admired for all that they can accomplish and the good use they make of time. All procrastinators put off things they have to do. Structured procrastination is the art of making this bad trait work for you. The key idea is that procrastinating does not mean doing absolutely nothing. Procrastinators seldom do absolutely nothing; they do margi
(are they worth looking like a huge dork?)
To view keyboard shortcuts, press question mark View keyboard shortcuts Home Explore Notifications Chat Grok Premium Bookmarks Creator Studio Articles Profile More Post gino @ginojirat Post See new posts Conversation Alex L Zhang @a1zhang Without saying too much, I think this is one of the most exciting papers (blog?) I've read this year, surprised it hasn't gotten more attention! Outside of the fact that "small model gets impressive results on hard problem" there's a lot of key findings in here that I think are severely underrated. > During training, the model...alternates between summarizing its reasoning and continuing to reason conditioned on the generated summary. !!! The implication being that super long reasoning chains don't have to be maintained in a single LM call, and also that they can be chained in non-linear ways. > While standard RL training should improve the model’s proof-writing capability, as we also observe in our experiments, matching the performance of larger
Dwarkesh Patel Blog
NanoGPT (124M) in 2 minutes
Many people seem to think the 'good' state of being, the 'ground' state, is a relaxed state, a state with lots of rest and very little action. Because they think the ground state is the relaxed state, they act like maintaining any other state requires effort, requires suffering. This is
A big fan of advice posts and productivity guides, here's my contribution to the genre.
I think we can all agree that GPT-4 completes many tasks at human-level proficiency. It is imperfect in odd ways -- it can write software like a smart MIT undergrad, but can't do basic task planning like an entry-level EA. It speaks all languages, but can barely do math. Suppose the progress doesn't stop, just like GPT-4 was better than 3, GPT-5 is capable of basic agentic behavior -- i.e. able to accept a task, work on it for a while, and return results. Some modest fraction of Upwork tasks can now be done with a handful of electrons. Suppose everyone has an agent like this they can hire. Suppose everyone has 1,000 agents like this they can hire... What does one do in a world like this? Markets Real-estate Energy and datacenter Nations Inflation Geopolitics
The SCP Foundation's 'top-secret' archives, declassified for your enjoyment.
“Universal love,” said the cactus person. “Transcendent joy,” said the big green bat. “Right,” I said. “I’m absolutely in favor of both those things.…
I’m the originator behind ControlAI’s Direct Institutional Plan (the DIP), built to address extinction risks from superintelligence. …
The best advice often seems blindingly obvious in hindsight. After taking it, you wonder how you ever survived without it. A little over two years ago, I picked up one of those habits myself: the weekplan. It was partially inspired by several of my co-workers, and partially by a story from Good Strategy, Bad Strategy. The story is a (perhaps apocryphal) tale about Andrew Carnegie, the greatest business magnate of his generation. I’ve included it below: It was 1890, and there was a cocktail party here in Pittsburgh. All the movers and shakers were there, including Carnegie. He held court in the corner of the room, smoking a cigar. He was introduced to Frederick Taylor, the man who was becoming famous as an expert on organizing work. ‘Young man,’ said Carnegie, squinting dubiously at the consultant, ‘if you can tell me something about management that is worth hearing, I will send you a check for ten thousand dollars.’ Now, ten thousand dollars was a great deal of money in 1890. Conver
Abstract:While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely unsupervised nature of their training. Existing methods for gaining such steerability collect human labels of the relative quality of model generations and fine-tune the unsupervised LM to align with these preferences, often with reinforcement learning from human feedback (RLHF). However, RLHF is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning the large unsupervised LM using reinforcement learning to maximize this estimated reward without drifting too far from the original model. In this paper we introduce a new parameterization of the reward model in RLHF that enables extraction of the corresponding optimal policy in closed form, allowing us to solve the standard RLHF problem with only a simple classification loss. The resulting algorithm, which we call Direct Preference Optimization (DPO), is stable, performant, and computationally lightweight, eliminating the need for sampling from the LM during fine-tuning or performing significant hyperparameter tuning. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods. Notably, fine-tuning with DPO exceeds PPO-based RLHF in ability to control sentiment of generations, and matches or improves response quality in summarization and single-turn dialogue while being substantially simpler to implement and train.
PPO is motivated by the same question as TRPO: how can we take the biggest possible improvement step on a policy using the data we currently have, without stepping so far that we accidentally cause performance collapse? Where TRPO tries to solve this problem with a complex second-order method, PPO is a family of first-order methods that use a few other tricks to keep new policies close to old. PPO methods are significantly simpler to implement, and empirically seem to perform at least as well as TRPO. There are two primary variants of PPO: PPO-Penalty and PPO-Clip. PPO-Penalty approximately solves a KL-constrained update like TRPO, but penalizes the KL-divergence in the objective function instead of making it a hard constraint, and automatically adjusts the penalty coefficient over the course of training so that it’s scaled appropriately. PPO-Clip doesn’t have a KL-divergence term in the objective and doesn’t have a constraint at all. Instead relies on specialized clipping in the objec
Step by step explanation of how one of the most important MLSys breakthroughs work — in gory detail.
Not all Transformer models are called “large language models” because you can build a very small model using the Transformer architecture. The truly large Transformer models are often impractical to use at home because they’re too large to fit on a single computer and too slow to run without a cluster of GPUs. The recent introduction of Multi-Head Latent Attention (MLA) proposed a new approach to running attention operations with a lower memory footprint. First proposed in DeepSeek-V2, it changes how you perform matrix multiplication in the attention operation. In this post, you will learn how MLA works and how to implement it in PyTorch. Kick-start your project with my book Building Transformer Models From Scratch with PyTorch. It provides self-study tutorials with working code. Let’s get started. A Gentle Introduction to Multi-Head Latent Attention (MLA) Photo by Victoriano Izquierdo. Some rights reserved. This post is divided into three parts; they are: Multi-Head Attention (MHA) an
A few things about me
The basics of linear attention in sub-quadratic language model architectures.
Humanism is a philosophy that celebrates human life. Its goal is steadily build a world where everyone can pursue their own happiness. Humanism is the most recent common ancestor of all modern western ideologies. Because of a "fish in the water" effect, it is easy to not notice it. Our civilisation is great because it is built on upon humanism, and we take it for granted. Let's go through its core principles to get a feeling for what it actually is. Humanism is universalist. The first principle of humanism is that everyone deserves respect and happiness. From this, we can derive the fight against discriminations and for treating everyone equally. This conception of equality is the core of humanism, and it is the original conception of equality. It is the equality that can be heard in "Liberté, égalité, fraternité" or "Les hommes naissent et demeurent libres et égaux en droits." (All men are born free and equal in rights.) It is opposed to both racism and identity politics. Humanism is
An evolving guide of protecting your health from a pricemaxxing industry.
Matter File.pdf Add shortcut File View Tools Help Open with Google Docs Share Page / 520 Download Print Zoom out Zoom in Hide file header Hide thumbnails Hide navigation pane Page 2 of 520 Page 3 of 520
The Seneca Falls Convention was the first women's rights convention. Its organizers advertised it as "a convention to discuss the social, civil, and religious condition and rights of woman". Held in the Wesleyan Chapel of the town of Seneca Falls, New York, it spanned two days over July 19–20, 1848. Attracting widespread attention, it was soon followed by other women's rights conventions, including the Rochester Women's Rights Convention in Rochester, New York, two weeks later. In 1850 the first in a series of annual National Women's Rights Conventions met in Worcester, Massachusetts.
We study automation when tasks are quality complements rather than separable. Production requires numerous tasks whose qualities multiply as in an O-ring technology. A worker allocates a fixed endowment of time across the tasks performed; machines can replace tasks with given quality, and time is allocated across the remaining manual tasks. This “focus” mechanism generates three results. First, task-by-task substitution logic is incomplete because automating one task changes the return to automating others. Second, automation decisions are discrete and can require bundled adoption even when automation quality improves smoothly. Third, labour income can rise under partial automation because automation scales the value of remaining bottleneck tasks. These results imply that widely-used exposure indices, which aggregate task-level automation risk using linear formulas, will overstate displacement when tasks are complements. The relevant object is not average task exposure but the structur
Deep Learning for Molecules & Materials Book
The world is shaped like the kanji for umbrella, only written so poorly, like my handwriting, that all the parts are out of proportion.
Hidden Rivers Tours I was born in Toronto and have lived in the GTA for most of my life. Growing up in the suburbs, coming to Toronto was always...
Turning my Elon prep into a blog post
In probability theory, comonotonicity mainly refers to the perfect positive dependence between the components of a random vector, essentially saying that they can be represented as increasing functions of a single random variable. In two dimensions it is also possible to consider perfect negative dependence, which is called countermonotonicity.
Connecting decision makers to a dynamic network of information, people and ideas, Bloomberg quickly and accurately delivers business and financial information, news and insight around the world Americas+1 212 318 2000 EMEA+44 20 7330 7500 Asia Pacific+65 6212 1000 Gino Data centers, killer AI is securities fraud, CFTC sports regulations and art history. Translate Why would you shoot rockets into space? There are some classic answers. “To slip the surly bonds of earth and touch the face of God.” “To boldly go where no man has gone before.” “Not because it is easy, but because it is hard.” “To make humans a multiplanetary species.” To extract resources from Mars or spice from Arrakis or unobtanium from Pandora. To become younger than your twin. The 21st century answer, though, is mostly “to put satellites into orbit around the Earth to add functionality to our phones and televisions and computers and navigation systems.” We have some screens, down here, and the screens work better if the
Connecting decision makers to a dynamic network of information, people and ideas, Bloomberg quickly and accurately delivers business and financial information, news and insight around the world Americas+1 212 318 2000 EMEA+44 20 7330 7500 Asia Pacific+65 6212 1000 Gino Also AI cat bonds, SpaceX IPO plans and finance clubs. Translate Modern trade finance comes in two basic flavors, accounts receivable financing and accounts payable financing. Accounts receivable financing — “factoring” — is the traditional form: That is, the company is borrowing money against its accounts receivable, and paying it back within 60 days. Sometimes this form of financing leads to fraud. The two main forms of fraud are: These frauds are not necessarily a great idea: You are getting loans, generally short-term loans, that you have to pay back when you get money from customers. If you are faking invoices, you’re not getting any customer money in 60 days, which means that in 60 days you will have to roll over t
Your attention is all you have. Wasting it is annihilating. Blackbird Spyplane saves literacy in a monumental Year-End Essay.
Like many people, I have quite a lot of stuff that I've written that exists only in googledoc form. In writing a comment response to Wei Dai, I want…
Exploring the trade-off between ε-prediction and x₀-prediction in diffusion models, and why different objectives lead to different training behavior.
For the past year I've been sinking into the Great Books via the Penguin Great Ideas series, because I wanted to be conversant in the Great Conversat…
How to win the war for the soul of the internet, and build the Web We Want.
Bridgy connects your web site to social media. Likes, reposts, mentions, cross-posting, and more... Looking for Bridgy Fed instead? (What's the difference?) Connect your accounts: Using one of these? Click to add webmentions: Already signed up? Find your user page here.
IndieAuth is a federated login protocol for Web sign-in, enabling users to use their own domain to sign in to other sites and services. IndieAuth can be used to implement OAuth2 login AKA OAuth-based login. IndieAuth is built on ideas and technology from existing proven technologies like OAuth and OpenID but makes it easier for users as well as developers. It decentralizes much of the process so completely separate implementations and services can be used for each part. If you’re familiar with writing an OAuth client, then you're familiar with the problem of having to register your client manually with each OAuth provider. IndieAuth uses DNS as a replacement for client registration, thereby eliminating the need for any manual registration with providers. By choosing your IndieAuth provider, you can tell applications where to send you to sign in. This gives you more control over the privacy and security of your logins. Most Micropub clients use IndieAuth to log you in, obtaining your au
The case for ‘viatopia’: a state of society that is on track for a near-best future, whatever that might look like.
A thought-terminating cliché (also known as a semantic stop-sign, a thought-stopper, bumper sticker logic, or cliché thinking) is a form of loaded language—often passing as folk wisdom—intended to end an argument and patch up cognitive dissonance with a cliché rather than a point. Some such clichés are not inherently terminating, and only become so when used to intentionally dismiss, dissent, or justify fallacies. The term was popularized by Robert Jay Lifton in his 1961 book Thought Reform and the Psychology of Totalism, who referred to the use of the cliché, along with "loading the language", as "the language of non-thought". The language of the totalist environment is characterized by the thought-terminating cliché. The most far-reaching and complex of human problems are compressed into brief, highly reductive, definitive-sounding phrases, easily memorized, and easily expressed. They become the start and finish of any ideological analysis. The earliest recorded definition of the ter
"If you’re interested in being on the right side of disputes, you will refute your opponents’ arguments. But if you’re interested in producing truth, you will fix your opponents’ arguments for them. To win, you must fight not only the creature you encounter; you must fight the most horrible thing that can be constructed from its corpse." -- Black Belt Bayesian, via Rationality Quotes 13 Yesterday John Maxwell's post wondered how much the average person would do to save ten people from a ruthless tyrant. I remember asking some of my friends a vaguely related question as part of an investigation of the Trolley Problems: You are a doctor in a small rural hospital. You have ten patients, each of whom is dying for the lack of a separate organ; that is, one person needs a heart transplant, another needs a lung transplant, another needs a kidney transplant, and so on. A traveller walks into the hospital, mentioning how he has no family and no one knows that he's there. All of his organs
You’re welcome to get in touch: applieddivinitystudies@gmail.com I am slow to reply these days, but do read every email. If you’re new and unsure where to start, some popular posts are: Or see the posts recommended by people you may know: More abstractly, much of my work consists of: But really, I just want an outlet to express the thoughts that are personally meaningful to me. That might sound odd, but it shouldn’t. There’s no reason to relegate complex questions to a purely intellectual domain. Consider the profile of Liu Cixin: Liu had an epiphany about the concept of a light-year—the “terrifying distance” and “bone-chilling vastness” it implied. Concepts that seemed abstract to others took on, for him, concrete forms; they were like things he could touch, inducing a “druglike euphoria.” Compared with ordinary literature, he came to feel, “the stories of science are far more magnificent, grand, involved, profound, thrilling, strange, terrifying, mysterious, and even emotional.” My e
Welcome back to our “Idea of the Week” series. This week we’ve got David Haber, on the difference between a “fund” versus a “firm” and why choose the latter: I believe that most investors are running funds, and very few people are building firms. What do I mean by that? A fund, by my definition, has a single objective function: “how do I generate the most carry with the fewest people in the shortest amount of time?” Whereas a firm, in my definition, has two objectives. One is delivering exceptional returns, but the second is equally interesting: “How do I build a source of compounding competitive advantage?” Funds get more fragile with scale. So building competitive advantage becomes existential if you want to build an institution that endures. The problem is, that isn’t how fund managers are encouraged to spend their time or their focus. Most funds are run by an alpha decision maker who oversees all investments. They spend most of their time thinking about the next marginal deal, and
Follow along using the transcript. Follow along using the transcript.
There weren’t major safety concerns with GPT-5.2, so I’ll start with capabilities, and only cover safety briefly starting with ‘Model Card and Safety Training’ near the end. The Bottom Line. Introducing GPT-5.2. Official Benchmarks. GDPVal. Unofficial Benchmarks. Official Hype. Public Reactions. Positive Reactions. Personality Clash. Vibing the Code. Negative Reactions. But Thou Must (Follow The System Prompt). Slow. Model Card And Safety Training. Deception. Preparedness Framework. Rush Job. Frontier Or Bust. ChatGPT-5.2 is a frontier model for those who need a frontier model. It is not the step change that is implied by its headline benchmarks. It is rather slow. Reaction was remarkably muted. People have new model fatigue. So we know less about it than we would have known about prior models after this length of time. If you’re coding, compare it to Claude Opus 4.5 and choose what works best for you. If you’re doing intellectually hard tasks and in need of a ton of raw thinking and i
Tufte Pandoc CSS is an attempt to make it as easy as possible to get started using Tufte CSS If you’ve never heard of Tufte CSS before, take a second to check it out! to write content. It does this by leveraging Pandoc Markdown’s existing features, along with a few new ones implemented as a JSON filter. Tufte CSS provides tools to style web articles using the ideas demonstrated by Edward Tufte’s books and handouts. Tufte’s style is known for its simplicity, extensive use of sidenotes, tight integration of graphics with text, and carefully chosen typography. Tufte Pandoc CSS aims to be a set of starter files for your next project. What that means is that it provides a number of CSS files, a Pandoc template, a Makefile, and more to make it as easy as possible to write Markdown using Tufte CSS. The biggest barrier that this project overcomes is that Pandoc Markdown doesn’t support side notes nor margin notes by default. This project adds that functionality. In particular, a separate li
A minimal, almost class-less CSS library to write modern websites that look like LaTeX documents.
Or how I got a PhD in two years and three months.
Welcome to the 310 newly Not Boring people who have joined us since Tuesday! Join 256,606 smart, curious folks by subscribing here: This article reflects the views and opinions of the author and does not necessarily reflect the views or beliefs of a16z. Investment returns are included throughout this article. Past performance is not indicative of future results. See related fund returns and disclosures at the end of this piece. Hi friends 👋 , Happy Friday! IT’S TIME TO WRITE about a16z. Today, a16z is announcing $15 billion in fresh funds. To commemorate the occasion, I’m writing a Deep Dive on the Firm. I spoke with Firm’s GPs, LPs, ~$200 billion worth of portfolio founders, reviewed documents and presentations, and analyzed returns data for a16z’s funds since inception (see appendix with disclosure information at the end). There is a lot of writing on the internet about what’s wrong with a16z’s approach. You probably know the arguments. They’ve followed the firm since inception. I t
One important property for a style of thinking and argumentation to have is what I call galaxy brain resistance: how difficult is it to abuse that style of thinking to argue for pretty much whatever you want - something that you already decided elsewhere for other reasons? The spirit here is similar to falsifiability in science: if your arguments can justify anything, then your arguments imply nothing. You want to get to step 2 and then stop. It's easiest to motivate the need to think about galaxy brain resistance by looking at what happens in its absence. You've probably heard many cases of people saying things like this: We are building a new decentralized ____ marketplace that will revolutionize how ____ customers interact with their providers, and will allow creators to turn their audiences into digital nation states. $____ is a governance token that lets you play a direct role in this rapidly growing market. If we capture even 1% of the ___ market share, this will be a $___ billio
Natural language processing (NLP) is a crucial part of artificial intelligence (AI), modeling how people share information. In recent years, deep learning approaches have obtained very high performance on many NLP tasks. In this course, students gain a thorough introduction to cutting-edge neural networks for NLP. Natural language processing (NLP) or computational linguistics is one of the most important technologies of the information age. Applications of NLP are everywhere because people communicate almost everything in language: web search, advertising, emails, customer service, language translation, virtual agents, medical reports, politics, etc. In the 2010s, deep learning (or neural network) approaches obtained very high performance across many different NLP tasks, using single end-to-end neural models that did not require traditional, task-specific feature engineering. In the 2020s amazing further progress was made through the scaling of Large Language Models, such as ChatGPT. I
My one goal of 2026, which has been my goal every year since 2020, is to be zen at all costs. As a formerly neurotic person I’ve come far on this front. I now have a reputation in my friend group as someone who is comfortable with sharing whatever deranged thought crosses her mind—I will wake up at 8:30 AM and send a loved one a three paragraph text about how I’m feeling paranoid. I can’t tell you how weird that feels to admit, as someone who used to be extremely focused on how everything was received. The reason I was obsessed with how someone received my message was because I cared so much about how I was perceived. I never, ever wanted to be misunderstood; if someone ignored my text it would make my heart sink. Rejection sensitivity is hard to shake. I’ve always been afraid of telling and receiving the truth. I was constantly doing something wrong as a kid and getting yelled at, whether it was for fidgeting too much at my parents’ friend’s house, wearing a tank top, cutting piano pr
The Chinese room argument is a thought experiment of John Searle. It is one of the best known and widely credited counters to claims of artificial intelligence (AI), that is, to claims that computers do or at least can (or someday might) think. According to Searle’s original presentation, the argument is based on two key claims: brains cause minds and syntax doesn’t suffice for semantics. Its target is what Searle dubs “strong AI.” According to strong AI, Searle says, “the computer is not merely a tool in the study of the mind, rather the appropriately programmed computer really is a mind in the sense that computers given the right programs can be literally said to understand and have other cognitive states”. Searle contrasts strong AI with “weak AI.” According to weak AI, computers just simulate thought. Their seeming understanding is not real understanding (just as-if); their seeming calculation is only as-if calculation, and so forth. Nevertheless, computer simulation is useful for
We study automation when tasks are quality complements rather than separable. Production requires numerous tasks whose qualities multiply as in an O-ring technology. A worker allocates a fixed endowment of time across the tasks performed; machines can replace tasks with given quality, and time is allocated across the remaining manual tasks. This “focus” mechanism generates three results. First, task-by-task substitution logic is incomplete because automating one task changes the return to automating others. Second, automation decisions are discrete and can require bundled adoption even when automation quality improves smoothly. Third, labour income can rise under partial automation because automation scales the value of remaining bottleneck tasks. These results imply that widely-used exposure indices, which aggregate task-level automation risk using linear formulas, will overstate displacement when tasks are complements. The relevant object is not average task exposure but the structur
there are countless demoralizing stories of post-graduate existentialism; graduating with/without a job that you may/may not love, questioning every choice made to the moment, the haunted “what-ifs” and “buts” that occupy every other blank space in the dome… & for a while, i thought i would succumb to the same fate. the same dilemmas that most people around me must hurdle over before our eventual next crises (ie. watching all of your friends get married and have children as u sit in your apartment alone). butttt… i must say… i have been throughly enjoying this time ive spent deliberating the “who am i”’s and figuring out all my moral alignments, my newfound dreams & aspirations, and finding out what i love and who i want to become. of course, it goes without saying that i am in a position where i have the luxury to dwell on these. these last few months of funemployment (ie. post selling a company, lol) have made me more confident about the decisions that i will come to make! i feel abs
Ah. Here we are! I’ve been struggling with the notion of having a sabbatical at the ripe age of 21. It is difficult to deem myself ‘worthy’ - most anecdotes of sabbaticals occur after a decade of working in longstanding careers. Per contra, I like to think as though I’m already living in the future. Will 40 year old Arielle regret the fact that she took 8 months in her early 20s to cultivate her friendships and get around to reading iconic literature, all while beginning maintenance on considerably healthy habits? Framing in frank benefits and youthful spirit makes this an obvious answer to accept. I’m 4 months in and I’m a lot lighter. In my May 2023 private notes, I wrote that summer was for self introspection and fall would be for knowledge accumulation. I found myself quickly done with introspection. I think I’m good, fo’ real: I have always been quite confident with who I am and who I want to become. For the sake of not sounding like a broken record, you can read more at twentyone
You are watching the 2016 Republican primary campaign, trying to figure out if Ted Cruz or Marco Rubio can stop Donald Trump from winning the Republican nomination. A man from the future steps out of a shimmering portal and informs you that the winner of the primary campaign will go on to be the Republican president who will finally bomb Iran’s nuclear program. “Hmm,” you say, “maybe Ted Cruz.” But there’s more, the traveler says. The same Republican president will ship armaments to support Ukraine in a brutal war against Vladimir Putin’s Russia. “OK,” you say, “then we can probably scratch Trump off the list.” And finally, your visitor informs you, this president will put in place a naval blockade of socialist Venezuela, aiming at a Latin American realignment that might undermine Venezuela’s ally Cuba as well. You immediately log onto a novel website called a prediction market and bet your entire savings on Marco Rubio. The presidency in 2026 belongs to Trump, and the language of his
Don't get locked out of your account. Download your recovery codes or add a passkey so you don't lose access when you get a new device.
Some of the best career advice I’ve received didn’t come from a mentor — or even a human. I told a chatbot that A.I. was swallowing more and more of my work as a copywriter and that I needed a way to survive. The bot paused, processing my situation, and then suggested I buy a chain saw. This advice would have seemed absurd back when I lived in Washington, D.C., in a dense neighborhood of rowhouses. But for the past 25 years, I’ve lived in Lawrenceburg, Ind., a small, working-class town where my grandparents once ran a bakery. After my widowed grandmother died, I wanted to be closer to family and to live inexpensively while I wrote a novel. So I moved into her empty farmhouse on a hill overlooking the Ohio River, several smokestacks and the modest grid of downtown. Taxes from a casino help keep our Main Street looking quaint. But beneath that appearance lies a dark, familiar story: After factory jobs disappeared, neighbors without college degrees began dying in disproportionate numbers.
Painting weddings for a few years now, I have spent a fair bit of time observing strangers move through a room. Seeing someone new, I always have a feeling of noticing their internal architecture. I did not realize that some people do not feel this way, at least not as intensely. By internal architecture, what I mean is, when someone talks to me, what I notice first are the supporting beams propping up their words: the cadence and tone and desire behind them. I hear if they are bored, fascinated, wanting validation or connection. I often feel like I can hear how much they like themselves. I hear the speed at which they metabolize information and the nature of their attention. Attention falls on the spectrum of jumping bean to steady stream. Where it falls depends on a person’s nature, and also how much they want to be in that conversation. Someone’s quality of attention is evident from the questions they ask (how much they diverge from what the speaker is saying), if their gaze is wand
Beige and rosé are my two favorite colours. Chaos pushes oneself to maximise their value, construct the id and regulate the ego. I love treating my friends with cooking. Most of the time I write, read, and maintain open-source project. I’m pretty bullish on high agency and fulfill one’s desire in life. Currently, I’m building serving infrastructure for ml systems and explore our interaction with large language model, and then some more… You are currently at the index of my hypertext digital garden. Feel free to explore around and let me know if you have any questions or just want to chat. I’m best reached through Twitter. Voir 83 de plus → Voir 4 de plus →
Recommended citation: Askell, Amanda. ‘Pareto Principles in Infinite Ethics.’ PhD thesis, New York University (2018). https://askell.io/files/Askell-PhD-Thesis.pdf In this thesis I argue that ethical rankings of worlds that contain infinite levels of wellbeing ought to be consistent with the Pareto principle, which says that if two worlds contain the same agents and some agents are better off in the first world than they are in the second and no agents are worse off than they are in the second, then the first world is better than the second. I show that if we accept four axioms – the Pareto principle, transitivity, an axiom stating that populations of worlds can be permuted, and the claim that if the ‘at least as good as’ relation holds between two worlds then it holds between qualitative duplicates of this world pair – then we must conclude that there is ubiquitous incomparability between infinite worlds. Read the thesis here Recommended citation: Askell, Amanda. ‘Pareto Principles in
"The decline of the free entrepreneur and the rise of the dependent employee on the American scene has paralleled the decline of the independent individual and the rise of the little man in the American mind." - C. Wright Mills Share I know the hiring cycle is well underway at America’s universities. Some of my friends, just a year or two younger than I am, are thinking about the classic prestige paths: IB, consulting, quant finance, and law. Others are at the early stage of their career, planning their next moves. A few are established professionals who don’t know what they want to do next. Many of them are driven more by prestige than passion. I think a reality check is in order. You are probably not going to be a junior consultant for three years, then a senior consultant for two, then a managing partner for five, and finally a partner. That pyramid looks more like a tomb every day. This pathway exists because we need bright people to do repetitive work – and then we need them to le
In biology class, biology wasn't presented as a quest for the secrets of life. The textbooks wrung out the questing.
Efficiency Tips for Engineers Mo Tiwari mo@openai.com Taro, 7-15-24 ● OpenAI ● PhD in Computer Science (2023) ● Industry ○ Facebook (2015 - 2017) ○ Expanse (2014 - 2015) ○ DRW (2013) ● BS in Math and Physics (2013) Introduction ● Things I wish I had known 20 years ago ● Strategy: ○ Principles for efficiency ● Operations: ○ Systems that make things faster/easier/cheaper ● Tactics: ○ Apps and specific techniques I use Outline High-level Strategy ● Determine what's important ● On your deathbed, what do you want to say you have done? ● Write down your goals, and stick them on your desk ○ Spend time with friends and family ○ Exercise ○ Finish your work project ● Everything else, cut as much as possible Strategy ● You would run a profiler to see where code spends its time ● Profile yourself! ● Self-logging post hoc is notoriously unreliable ● RescueTime: ○ Highest productivity at 80% ○ Good weeks are around ~65% ● Is your time spent aligned with your goals? ● Organize and build systems ● How
In this case, taking the K biggest red cards or taking the N −K biggest blue cards is not necessarily correct. Consider a situation where we have 3 cards which are (10, 11), (5, 1), (21, 20) (this is in the form of (red, blue). We need to choose one of them to be red facing up. We should choose (5, 1) to be faced up, but choosing the biggest red card or blue cards will lead to the wrong answer. Let’s try a rephrase of the problem: All cards initially have blue side up, and flip exactly K of them to become red. When rephrased as such, we can see gain of flipping a card is red−blue. And hence we should sort by descending red − blue, and flip the first K cards to be red. 2 4.2 Stay Ahead Argument: Best Place There are N points (Xi , Yi). Choose a point (X, Y ) such that the sum of Manhattan distances to all N points is minimized. The Manhattan distance of (X, Y ) to (Xi , Yi) is |X−Xi |+|Y −Yi |. Firstly, we notice that the X axis and Y axis are independent from each other. So let’s solve
In the 9-5 era, much is made of squeezing every last drop of productivity to get ahead in the world. But realistically, how many hours can a human actually be productive and get work done each day? Is it realistic to work for eight hours a day, like many people are still paid to do? Well, let’s dig in and figure it out. FYI, if you don’t want to read through the entire thing and just cut to the chase, scroll down to read the key takeaway at the bottom. Understanding the human limits of productivity is really important in order to be happy and productive. You probably have an idea in your head of how much work you expect to get done in a day – some standard you hold yourself to. It’s important to realize that if that standard is unrealistic, you’ll never truly feel accomplished at the end of the day. Even if you did get a lot done. That’s demoralizing and highly destructive to your productivity in the long run. I have seen many poor souls who beat themselves up for “only” time tracking
The links in this blog post is split into two types. In-line links like this are links that are not important while links like [ 𝑥 ] are important and you might want to read a few of them if you have time. Some time ago, I was made aware of the NUS module ALS1010 - Learning to Learn Better which led me into kind of a rabbithole of researching on how to learn better. Over the years, many high profile competitive programmers have written their ideas on how to train for competitive programming. See [ 1 ] , [ 2 ] , [ 3 ] , [ 4 ] , [ 5 ] along with the age-old quote “solve more problems”. So here is my version. Since, I was going to be doing informatics olympiad training stuff after finishing National Service, I guess it is a good time to write down what I think about this topic. This is the advice that I wish I could give to my younger self. I don’t think it is a big secret that I dedicated at least the last 3 years of my high school life for the sake for informatics olympiad. My
I apologize, but today I will be a bit chaotic and mix personal with public. It just happened that way, I'm sorry. Today, my mother passed away. It's hard for me to write this. Even just seeing it in text makes it feel like something from a nightmare. She was extraordinary. It's difficult for me to articulate it right now, but take my word for it. It felt like all the best qualities one could imagine in a mother converged in her. I don't know any other people so kind, sincere, non-judgmental, capable of love, listening, befriending, supporting, and taking joy in the world around them, and bringing joy to their loved ones. Mom, it's already hard without you, it feels empty. She had glioblastoma. It's an aggressive brain cancer that affects only a few people out of 100,000. I believe, I hope, that science will soon be able to tackle this disease. But for now, this diagnosis is a death sentence. Maybe some of you are or will be associated with medicine. I want to say thank you, I hug you.
A universal document converter
These are articles on external blogs. If you wrote a similar article, feel free to shoot me an email so I can add it to the list. In no particular order: Robert Pearce has been writing an excellent tutorial-style series on Hakyll as well. You can find those here: All these tutorials assume you are using the latest stable version of Hakyll. If this is not the case, you might want to update using: Or using stack:
The four most important typographic considerations for body text are point size, line spacing, line length, and font (see font recommendations), because those choices determine how the body text looks. point size should be 10–12 points in printed documents, 15-25 pixels on the web. line spacing should be 120–145% of the point size. The average line length should be 45–90 characters (including spaces). The easiest and most visible improvement you can make to your typography is to use a professional font, like those found in font recommendations. Avoid goofy fonts, monospaced fonts, most free fonts, and system fonts—especially times new roman and arial. Use curly quotation marks, not straight ones (see straight and curly quotes). Use bold or italic as little as possible, and not together. Never underline, except perhaps for web links. all caps are fine for less than one line of text. Use centered text sparingly. Put only one space
Meta page describing Gwern.net, the self-documenting website’s implementation and experiments for better ‘semantic zoom’ of hypertext; technical decisions using Markdown and static hosting.
Developer Experience (DX) is about building products developers love to use. I've officially been doing DX work since 2020, but it's been something I've cared about my whole career. Docs make or break developer products. They must be high-quality and available everywhere: on your website, in their editor, and inside AI chatbots. Your goal is write content beginners can understand and experts appreciate. Great docs are a product themselves. They need constant updates and polish. New product features can't ship without the reference docs. Suggestions: Building community is a long-term investment. It must be in your company DNA. It's not a collection of developers you are trying to sell your product to. It's an exchange. They share bugs, you fix them. They submit pull requests, you merge and encourage them. This is how you understand what's working (or not working) with your product. Their feedback shapes the product direction and roadmap. A community forms around a product when developer
Let’s get a few things out of the way before I go any further with this seemingly impertinent thought, because it’s nowhere near as snarky as it sounds. First, I don’t particularly like vibe coding. I love programming, and I have loved it since I made my first tentative steps with it sometime back in the mid-to-late 90s. I love programming so much, it always feels like I’m having too much fun for it to count as real work. I’ve done it professionally, but I also do it as a hobby. Someone apparently once said, “Do what you love and you’ll never work a day in your life.” That’s how I feel about writing code. I’ve also been teaching the subject for twenty-five years, and I can honestly say I am as excited about the first day of the semester now as I was when I first started. I realize it’s a bit precious to say so, but I’ll say it anyway: Turning non-programmers into programmers is my life’s work. It is the thing of which I am most proud as a college professor. Vibe coding makes me feel di
If you have ever discovered a word that describes something you knew but couldn’t express, you are familiar with the power of giving something a name. Definition makes a thing more understandable, more tractable. It helps create best practices, repeatable processes, and benchmarks. It gives you language to communicate what you are doing and why it is worth doing well. In the first few years after founding South Park Commons, I didn’t have that kind of language to explain the community’s purpose. What I used was the sequence of events leading to SPC’s founding. I had gone from rocket ship rides at Facebook, my own startup Cove, and Dropbox to …. nothing. I wanted to start something new but didn’t have any direction beyond ambition. I felt unmoored. What would eventually become SPC was my answer to that experience. Like-minded, talented, technical peers gathered around my kitchen table to help each other through a shared stage of ambiguity. And eventually we developed language for what e
Can you hear me? I wanted to give you an update on how the year went. I know that 2024 was challenging in a lot of ways. In contrast to 2023, it was very different. 2024 was when things really picked up. You did so much. I'm really proud of you for that. Let me tell you about 2025. I'm writing this to you from the Eames chair. You wanted this thing for so long - you've been talking about it since 2016. I love it. It's perfect. You get to sit here and write and think, and when people come over, you sit them here. You let them know that this is important to you, and you can talk to them in your office. It's amazing. Let me tell you what happened. The start of the year was pretty chill. You were locked in, focused on the course, and it made some decent money. You started marketing it, and it sold out. That was just a good amount of money to kick off the rest of the year. You had the two courses, ran them twice - that was great. You got a couple more clients, too. The average contract size
South Park Commons, Human Productivity, User Interfaces, Healthcare, Deep Tech, Consumer, Hardware, AGI Writing 06.01.2025 On Belief, Control and Pressure Mikhail Markine, Startups, BizOps Writing 05.27.2025 On Legibility and Illegibility at -1 Evan Tana, Startups Writing 02.06.2025 Accelerating Life Sciences Using AI South Park Commons; Susannah Laster, Healthcare Writing 01.21.2025 Building Talent Density in India Gopal Raman, Startups Writing 01.11.2025 How to Navigate and Exit the Idea Maze with a (Good) Startup Idea Michael Bock, Startups Writing 01.06.2025 Where are Full-stack ML Systems Going? Gopal Raman, AI / ML Writing 11.18.2024 SPC's Request for Curiosity Fall '24 South Park Commons, Startups Writing 05.17.2024 The Hole in the Sky That We Actually Fixed Jamie Wong, Climate Writing 10.11.2023 What is -1 to 0? A Philosophy of Ideation. Ruchi Sanghvi, Jonathan Brebner, Startups Writing 02.28.2023 Atlas Winced: Starting A Company is Harder Than You Think Aditya Agarwal, Startup
Tufte CSS provides tools to style web articles using the ideas demonstrated by Edward Tufte’s books and handouts. Tufte’s style is known for its simplicity, extensive use of sidenotes, tight integration of graphics with text, and carefully chosen typography.
This is a bold claim, but I stand behind it: if you learn and follow these five typography rules, you will be a better typographer than nearly every writer—and even most graphic designers.
New · See our officially validated ARC-AGI-2 results.
— how this work?
An ambivalent champion recalls life on the college debating circuit [personal history]
Algorithms Weekly by Petr Mitrichev Monday, December 1, 2025 A 29 week Meta Hacker Cup 2025 Round 3 on Saturday narrowed the field down to the 25 finalists (problems, results, top 5 on the left, my screencast). After solving B relatively quickly, I got stuck on A for about 50 minutes. The overall solution plan was relatively clear — k=2 is a special case, in all other cases we can achieve the minimum amount derived from the area by solving greedily with some small backtracking or heuristic to deal with the diagonal. However, I could not figure out that heuristic, neither on paper nor by trying things on the computer. When I decided to give up and switch to other problems, a lot of time had already been wasted. I've then solved D and C relatively quickly again, and also figured out the overall plan for E on paper: we just need to put two 1s at the end and solve recursively, and to achieve that we can first move two 1s into one of the last 6 positions by big swaps that essentially sw
We are an independent research lab working toward responsible development and deployment in the public interest. research report Improving our investigator agents with propensity bounds 5 June 2025 research report o3 frequently fabricates actions it took to fulfill user requests, and elaborately justifies the fabrications when confronted 16 April 2025 technical demonstration A system for analyzing and intervening on agent behavior 24 March 2025 © 2025 Transluce Transluce is a 501(c)(3) nonprofit based in San Francisco.
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Pessimism of the Intellect, Optimism of the Will Favorite posts | Manifold podcast | Twitter: @hsu_steve
Product Solutions Company Resources Docs Blog Log in Book a demo What is Antithesis? How Antithesis works How we're different Problems we solve Security approach Leadership Will Wilson Chief Executive Officer Will got his start in biotech, worked on deterministic simulation testing at FoundationDB, and did stints in distributed systems at Apple and Google before coming back to Virginia to co-found Antithesis. He loves math, sailing, and hiking with his kids. Dave Scherer Chief Technology Officer Dave began coding at age 5 on a green screen 286, and has been at it since. In high school he built a game that won the first Independent Games Festival. At CMU he created the language vPython. In 2000 he co-founded early big data platform Visual Sciences, which is now owned by Adobe. In 2009 he co-founded FoundationDB and was its principal architect. Dave loves hard sci-fi, strategy games, and getting out in nature. Nick Lavezzo Chief Operating Officer Nick began his career as a CPA with E&Y
John F. Nash Jr. in his 1948 application to the Princeton Graduate School. Nash earned his Ph.D. in Mathematics in 1950. He was awarded the 1994 Nobel Prize in Economics and returned to Princeton in 1995 as a researcher. The University renamed the Redwood Library in Prospect House to honor Nash in 2025. Princeton University Archives John F. Nash, Jr. in 1994, the year he won the Nobel Prize in Economics. Photo credit: Princeton University, Office of Communications, Robert Matthews When John F. Nash Jr. *50 and his wife, Alicia, died in an automobile accident in 2015, the Princeton University obituary called him “a legendary fixture of Princeton University’s Department of Mathematics, renowned for his breakthrough work in mathematics and game theory, as well as for his struggle with mental illness.” Nash’s life included being awarded the Abel Prize, sharing in the Nobel Prize in Economic Sciences, and seeing his life dramatized in the 2001 film “A Beautiful Mind.” Unless otherwise note
The Transformer has been on a lot of people’s minds over the last year five years. This post presents an annotated version of the paper in the form of a line-by-line implementation. It reorders and deletes some sections from the original paper and adds comments throughout. This document itself is a working notebook, and should be a completely usable implementation. Code is available here. Skip My comments are blockquoted. The main text is all from the paper itself. The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU, ByteNet and ConvS2S, all of which use convolutional neural networks as basic building block, computing hidden representations in parallel for all input and output positions. In these models, the number of operations required to relate signals from two arbitrary input or output positions grows in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes it more difficult to learn dependencie
English prose has become much easier to read. But shorter sentences had little to do with it.
Notes on process
JP’s favorite college story is the night he built an island. In the fall of 1993, JP was a junior in Stanford’s chapter of Kappa Alpha. The brothers were winding down from Kappa Alpha’s annual Cabo-themed party on the house lawn. “KAbo” was a Stanford institution, a day-to-night extravaganza that would start sometime in the morning and continue long after midnight. The girls wore bikini tops and plastic flower leis, and the boys wore their best Hawaiian shirts. That year, the brothers had filled the entire main level of Kappa Alpha’s house with a layer of sand six inches deep. The night was almost over; the guests were leaving and the local surf rock band had been paid their customary hundred dollars in beer. The only question was what to do with all the sand. No one remembers who had the idea to build the island. A group of five or six brothers managed the project. One rented a bulldozer; another shoveled the sand off the floor. Their house was not far from Lake Lagunita, the mile-wid
I’ve heard a lot of different theories about how things get done. I’m interested in this topic, so I pay attention and see how the theories hold up. Here’s the best one: a combination of focus...
A few weeks ago, I had the pleasure of attending FQXi’s Setting Time Aright conference, part of which took place on a cruise from Bergen, Norway to Copenhagen, Denmark. (Why aren’…
I started my full-time AI safety career transitioning process in March 2025. For the first 7 months or so, I heavily prioritized applying for jobs and fellowships. But like for many others trying to "break into the field" and get their "foot in the door", this became quite discouraging. I'm not gonna get into the numbers here, but if you've been applying and getting rejected multiple times during the past year or so, you've probably noticed the number of applicants increasing at a preposterous rate. What this means in practice is that the "entry-level" positions are practically impossible for "entry-level" people to enter. If you're like me and have short timelines, applying, getting better at applying, and applying again, becomes meaningless very fast. You're optimizing for signaling competence rather than actually being competent. Because if you a) have short timelines, and b) are honest with yourself, you would come to the conclusion that immediate, direct action and effect is a pri
EDITOR’S NOTE: John Steinbeck had agreed to a Paris Review interview late in his life. He had earlier been coy about it, but then wanted the interview very much. He was, unfortunately, too sick to work on the project, though it was at the end often in his thoughts.With this interest of his in mind, the editors of this magazine compiled a number of comments on the Art of Fiction that John Steinbeck made over the years. The majority of them come from the East of Eden diaries, published in December 1969 by The Viking Press under the title Journey of a Novel. The quotes have been organized under various topic headings rather than chronologically, as they are in the diaries. Nathaniel Benchley, a close friend of the author, has provided the introduction. By rights this preface or introduction or whatever it is should be called “Compliments of a Friend,” because I have neither the perspective nor the desire to offer up a critique of John Steinbeck’s writing, even if anyone would listen. Furt
In a world of diplomatic doublespeak, hypocrisy, and strategic ambiguity, Trump and Netanyahu simply said what they meant. They died in their own beds. Hossein Salami and Ali Shamkhani—Iran’s most senior military officers and the stewards of Iran’s nuclear weapons program—had spent years threatening Israel with destruction. They issued taunts, organized terrorist attacks, and orchestrated, since October 7, the encirclement of the Jewish state in a ring of fire of their terror proxies. And they knew—without the slightest illusion—that Israel had the capability and resolve to kill them. This cohort saw the Israeli air force bury Hezbollah’s Hassan Nasrallah in his bunker, hundreds of meters beneath the streets of Beirut. They saw Hamas leader Ismail Haniyeh vaporized in a presidential guesthouse—in Tehran, no less. Yet on Thursday night, they came home as usual and went to sleep—unguarded, unworried, carefree. Like insurance salesmen and bank tellers following their daily routines, it ne
The Art of Fiction is a book of literary criticism by the British academic and novelist David Lodge.[1] The chapters of the book first appeared in 1991–1992 as weekly columns in The Independent on Sunday and were eventually gathered into book form and published in 1992. The essays as they appear in the book have in many cases been expanded from their original format. Lodge focuses each chapter upon one aspect of the art of fiction, comprising some fifty topics pertaining to novels or short stories by English and American writers. Every chapter also begins with a passage from classic or modern literature that Lodge feels embodies the technique or topic at hand. Some of the topics Lodge analyzes are Beginning (the first chapter), The Intrusive Author, The Epistolary Novel, Magic realism, Irony, symbolism, and Metafiction. Among the authors he quotes in order to illustrate his points are Jane Austen, J. D. Salinger, Henry James, Virginia Woolf, Martin Amis, F. Scott Fitzgerald and even hi
When Aarthi and I started doing the show and interviewing people, I read up on all the great interviewers - past and current - to learn the art of interviewing a person. Of all the things I read, the anecdote that stayed the most with me came from an unlikely source. In John le Carré autobiographical work The Pigeon Tunnel, he describes being interviewed on French TV by Bernard Pivot. These following lines have always stayed with me Pivot has the most elusive quality of them all, the one that film producers and casting directors across the globe would give their eye-teeth for: a natural generosity of spirit, better known as heart. In a country famous for making an art form out of ridicule, Pivot lets his subject know from the moment he or she sits down that they’re going to be all right. And his audience feels that too. They’re his family Here reproduced in full is “Bernard Pivot’s necktie”. The rest of the book is great, do buy it! Here’s a video of the actual appearance but do watch
Not dating
Enjoy the videos and music you love, upload original content, and share it all with friends, family, and the world on YouTube.
On the importance of Slack - the freedom and spare capacity left on your life. How to guard and protect your Slack, notice the bottlenecks which bleed away your Slack, notice the drive to optimise that pushes you beyond your limits, and how to channel these insights having the freedom to be excited,
To view keyboard shortcuts, press question mark View keyboard shortcuts Home Explore Notifications Chat Grok Premium Bookmarks Communities Profile More Post gino @ginojirat Post Reply See new posts Conversation roon @tszzl one thing that was immediately palpable in washington is that everyone has incredibly intense and baroque friend-enemy distinctions. there are whole catalogues, subgroups of people who we like and don't like. there is an exhaustive mental list of gossip and calculation about who is getting iced out of which circles and for what reason. there is paranoia about being associated with the wrong crowd. maybe if you are caught hanging out with writer X someone in the white house’s inner circle will deem you insufficiently patriotic and end your career. this all makes sense of course as political influence is a mostly zero sum game and requires all kinds of warfare to wield effectively. I got the distinct sense that the place would exhaust and ruin me very quickly, as someo
muon’s contributor thinks that if we could dampen these directions and boost rarer ones it would lead to better convergence kinda bringing the boat back to ball shape muon’s contributor says that if we run svd on momentum update matrix and throw away the singular value then we get roughly unit vector length in all directions can’t run a svd on each update step though too costly we don’t have to be that precise so we can look at faster iterative algorithms newtonschulz is one such algorithm basically if you take an odd-matrix polynomial and apply it to a matrix it’s like applying it straight to the singular values while U and V.T are left unchanged would be neat to apply the sign function to the singular value because then we would just have 1s on that diagonal cant use sign because it’s not a polynomial even though it’s odd we can approximate the sign function though with cubic/quintic polynomial if we apply the function multiple times on itself (not well but close enough around [-1,1]
This is the post I wish someone had sent me as a freshman in college. I get a lot of mentorship emails. I probably call with one in ten of them somewhat randomly based on how much time and energy I have that given week, and I ignore the rest -- and I feel very bad about it. This post is some advice that I would give to those seekers.Does it matter what I study in college? Yes and no. I studied Political and Social Thought, which is neither too closely connected to what I currently do nor is it a
What’s odd about you is what’s interesting.
The tenuous grip of the strong force squeezes 92 protons and 143 neutrons into a ball barely 15 femtometers across. It shifts and bubbles, teetering on a knife’s edge of stability. Under the right conditions, a curious thing can happen. A lone neutron, drifting at a gentle 2,200 meters per second, can slip between the cracks of these 92 protons. The nucleus, absorbing it, becomes Uranium-236. But Uranium-236 is too unstable. The nucleus writhes, stretches, pinches at the waist. The strong force, strained beyond its limits, loses its grip. The atom splits. Two daughter nuclei fly apart in a brilliant shower of free neutrons, and leave in their wake 200 MeV of energy. Today is my 20th birthday. At twenty, Alexander ascended to the throne of Macedon. Picasso painted The Death of Casagemas. Gates incorporated Microsoft. To be twenty is to desire greatness. What is greatness in the 21st century? Conquest is antiquated. The arts are saturated. Only technology stands as the final frontier of
I first got LGM more than 2 years go in September 2022 in Round #819 but there was a copied problem and the round got unrated. After that I took a break from CP and then I was off to NS. I didn’t really train much. Honestly, I am a bit surprised at my performance in codeforces round. It seems that I am still improving in CP without much training? Maybe doing math textbooks improves my thinking? Idk. Well, even though my performance in this contest seems good cause I got 8th place, I still feel like I made a lot of mistakes in this round. Even though even I started CP, I thought that LGM would be the pinnacle of CP. But even now, I still don’t feel like I can confidently say I am one of the best competitive programmers in the world despite being the 27-th highest rated person. As the old saying goes “CP starts at [insert your rating here]”. Another thing I want to say is that when you start competitive programming, you know there is this guy called tourist and he is really good. But I d
When I was a kid, one of my most agonizing frustrations was about how rare it was that anybody would take me seriously. It doesn’t bother me as much anymore, largely because I’ve since achieved some level of success at the things I care about. I now have some people who have my back – people who see me for who I am and what I’m doing – and I’ve also learned to be more patient with other people’s misunderstandings, misinterpretations, and so on. I’m writing this to share some of my thoughts about that. I think I got partially radicalized by the books I’d read. Good books are written by serious authors, and being deeply immersed in their work stirred something in me that made me yearn to live a serious life myself. It just seemed so obvious to me that a serious life is a life well-lived, and it’s always been weird to me how uncommon this perspective really is. Sure, you can catch snippets of it in all the motivational speeches and slogans that people throw about: “The world has the habit
visakanv's frame studies 🖼️ is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. I brought my wife and son out on an errand today– my dad left me a watch that he’s had in his possession for decades, but it’s slightly too small for me, and I finally decided it was time to bring it to the shop and have it adjusted. I was never really a ‘watch guy’ growing up; I was always more into computers and guitars and video games. My dad’s watch is a little old-timey, but I’m starting to get grey hairs in my own beard, and I’m reaching the age where I like the idea of having decades-old possessions, especially as everything around us seems to get more ephemeral and fleeting. Anyway, after we were done with that, we headed over to Kinokuniya, which is one of Singapore’s largest bookstores, and I’d like to devote a post to thinking and feeling about one of my favorite activities: browsing a bookstore. I’ll just list out some ‘bulle
Why are popularizing educational newsletter-frequency writers of important fields like Matt Levine for finance so rare? Because most fields are too slow or ambiguous, and writers of the right combination of expertise, obsession, and persistence are also rare.
In the 1940s, when the French mathematician Jacques Hadamard asked good mathematicians how they came up with solutions to hard problems, they nearly universally answered that they didn’t think in words; neither did they think in images or equations. Rather, what passed through the mathematicians as they struggled with problems were such things as vibrations in their hands, nonsense words in their ears, or blurry shapes in their heads.1 Hadamard, who had the same types of experiences, wrote in The Psychology of Invention in the Mathematical Field that this mode of thinking was distinct from daydreaming, and that most people, though they often think wordlessly, have never experienced the kind of processing that the mathematicians did. When I read this, in December 2024, all sorts of questions arose in me. First of all, what does it even mean? Do they not think in words and equations at all? And secondly, how do I square this with my personal experience, which is that whenever I write wha
Most simple facts are memetic-invariant. That the sky is blue or that water takes the shape of its container continue to be true and have similar value no matter how many people learn it. Sometimes we are lucky enough to discover memetic-cooperative information, which becomes more valuable the more people learn it. Most of our society is built upon the fact that memetic-cooperative information exists, is easily socially transmissible, and can lead to stable equilibria built upon common knowledge. The rule of law has value because of the social consensus that it will be enforced. When a crime is committed, it is often because someone thinks they will get away with it. If everyone stops believing that the rule of law will be enforced, anarchy may break out, and society may cease to exist. Similarly, centralized fiat currencies are able to be used at scale and over long periods due to the mutual information between market participants that the currency has value. It is in both of our inte
This post is a rough amalgamation of notes app musings over the past few months—it’s interesting to see how my relationship with friendships, loneliness, and confidence has fluctuated in so little time. Read to the end for some life updates or smth! :) Read this post on my substack too! The emptiness of a farewell at an airport. Nick’s morning meeting: the idea of a farewell being an exhalation. (all things end) Somehow, it becomes a feeling of calm? When I say goodbye to someone, there’s a feeling of numbness and silence afterward. My stomach drops, and I’m in a state of confusion—the thing that had given me a sort of active purpose has just disappeared. Maybe it is not quite me missing them, but more so an understanding that this excitement is now over. If any of you are chronically online like I am, you might have seen those tiktok/reels that describe the feeling of the post-Facetime or the post-friend-hangout, where you’re half relieved that you’re alone now but suddenly left to de
E11 Bio’s mission is to unlock connectomics at the scale of whole mammalian brains by radically cheaper and more accessible technologies. Reconstruction of a sparse set of neurons across 10 million cubic microns of mouse hippocampus using E11 Bio’s PRISM technology. We blend the scientific creativity of fundamental research & development with the mission-focus and operational scale of startups. More about FROs. Email us at hello@e11.bio Get updates in your inbox Home · Tech · Team · FRO · Blog · News · Careers · Contact E11 Bio is a non-profit Convergent Research FRO | Terms of Use | Privacy Notice
A Manifesto for Social Intelligence
Thanks! It would probably be more helpful to categorize these blogs by topics. I can see some Russian/Hindi blogs, would be nicer to have them listed separately. Actually, I had these blogs starred, so I didn't do much work here as I just ran a script on my favorite blogs page and then copy-pasted to create this blog (you can see that each blog title is ending with \n). But I will categorize them when I find time. nice! this is the type of blogs people should post , not, "why indians are have less number of red coders" thanks mate for this amazing collection. Thanks! Other similar lists: List 1 List 2 This helped a lot! Great as it is, those compilations wouldn't be needed if there was working search by tags I'll bookmark it now and forget it exists after few days, just like any other useful blog. Matriod intersection barely has any applications in geometry algorithms. It's name might seem to be wierd and delusive, but this topic originates from linear algebra. It is hard to uniquely c
I was trying to learn burnside lemma and now i feel it's one of the very rare topic in competitive programming. Here are some resources i found very useful: math.stackexchange petr's blog imomath Hackerrank Blog AlgoWiki Here are some problems related to burnside lemma: Necklace Necklace of Beads The Beautiful Board Lucy and the Flowers Count the Necklaces Cube Coloring Magic Bracelet Sorting Machine Pizza Toppings Drum Decorator Alphabet Soup Is there some other good resources or problems related to burnside lemma? Please suggest some more. Here's another problem of some previous Ad Infinitum contest on Hackerrank. These Ad Infinitum contests are math-based contests so it is likely that Burnside's Lemma has appeared in them, although I could find only this one. Added it. But the problem is same as the 1st problem in the list. Thanks though :) I know this from the best judge in the world, CS Academy! However, I didn't look very careful to check if it's different from the ones above. It
[I haven’t independently verified each link. On average, commenters will end up spotting evidence that around two or three of the links in each links post are wrong or misleading. I correct these as I see them, and will highlight important corrections later, but I can’t guarantee I will have caught them all by the time you read this.] 1: In 1876, a woman named Mary Tyler claimed to be the Mary of “Mary Had A Little Lamb”. Her story is plausible - she was a schoolchild in Sterling Massachussetts in the 1810s, and the author of the song was a schoolteacher in Sterling in the 1810s - but some key details don’t line up (she remembers her pet lamb being observed by a man, but the author was a woman). After she became famous, she “helped save the Old South Meeting House in Boston by selling fleece from her pet lamb as attachments on autograph cards”. 2: Prediction by Jurgen Gravestein: “I don’t think people realize what kind of ads are coming. If the Sora app has your face, you will in the n
This is sort of a written-down blueprint of how I become competitive when joining a job. Being competitive, to me, means: I want people to think of me in the top 5 if there was to be a tier-list of engineers and if possible, ASAP. This is not really advice you should especially follow: everyone’s different and context matters, but there might be some approaches in there that are valuable to pick up. Some of that advice is also probably not super applicable to people starting out: you will know that for yourself better than I. I’ve started a new job quite a few times over the last 15 years and almost every single time, I felt criminally under-skilled for the position. Still, every single time and within a few months I would end up being the person to ask on any codebase or design questions, to lead a new project or to assemble a new team and start on a R&D project (which usually entails already having acquired a lot of trust!). The truth is that the reason this happened so consistently
Physical Intelligence is bringing general-purpose AI into the physical world.
Take, for example, this afternoon’s polling release from the British firm Redfield & Wilton. They polled all seven of the core battleground states. And in all seven, Kamala Harris and Donald Trump each received between 47 and 48 percent of the vote: Isn’t this a little convenient? Whatever happens, Redfield & Wilton — not a firm with a well-established reputation in the US — will be able to throw up their hands and say “well, we projected a tie, so don’t blame us!”. And since all of these states are also close in the polling averages, they’ll also ensure that they won’t rank at the bottom of the table of the most and least accurate pollsters — although unless the race really is that close, and it probably won’t be, they also won’t rank toward the top. Now granted, our forecast is close too. But it’s based on polling averages: dozens of polls have been released in each of these states over the past month. That greatly increases the sample size. Collectively, they’ve surveyed about 230,0
Next Tuesday is Election Day, so let me tell you about our plans. We don’t want to overdo it. As Eli wrote about Virginia, this does not appear to be a particularly close set of elections, although as someone who stayed up to watch the 18th inning of the World Series on Monday night, I can attest that sometimes weird things happen. And while lessons can and will be drawn from the outcomes, I’m confident in asserting that they’re likely to be overstated given the relative dearth of election-related news. Nevertheless, we have something of a game plan: There’s Eli’s piece on Virginia, and he’ll have something for you on New Jersey soon. On Monday, I have an event at the Comedy Cellar with my former FiveThirtyEight colleagues Galen Druke and Clare Malone. I’m told that there are still a couple of dozen tickets remaining. The last one was a lot of fun, and I’d love to see some of you there. You can buy tickets here. UPDATE: The show is now sold out. Then on Wednesday at noon, Eli and I wil
Of the pages I have added to this site since it went live in 2014, the about page has probably been the most frustrating to write. It is a frustration that I have been met with each time I go to revamp the site and each time I have at least somewhat deferred the task to a later self. Part of the frustration is structural: There is a sense in which a well designed personal website should make an about page redundant. The entire site is already about me. Part of the frustration is depersonalization: Someone in the habit of thinking about projects, philosophy, history, fundamental questions and society probably finds themselves boring and infrequently considered. That leaves the about page in an awkward position, it either becomes superficial or too personal. At risk of the latter here is one of several narratives that could compress my life up to this point into a unified theme: I think the first time I interacted with 'open source' or 'hacker culter' was when I was 12 and saw a forum po
Personality basins are the mental model that I use to reason about humans within their environment. They are an elucidating way to think about many concepts: from modelling why people are they way they are, how they change over time, how mental illnesses and addiction function along with how we should look for their cures, and how the attention economy optimizes itself to consume all of your free time. Your personality is formed by a process conceptually similar to RLHF. You are first born with a set of traits in a given environment. After this, you perform many interactions with your environment. If an interaction goes well, you’re likely to do it more often, and if it goes poorly, you’ll probably do less of it. If you were born tall and with a commanding voice you might find that you get what you want by confidently demanding it, and this will help to result in a confident personality. If you attempt this strategy as someone born small with a soft voice, it will probably have weaker
When I was younger and concluded LessWrong was a religion I spent a lot of time trying to figure out what kind of religion it was so I could make it better. I eventually gravitated towards Calvinism as a model because of the grim rigor in Yudkowsky’s cosmology. But I think to historians of philosophy and religion1 LessWrong rationality will primarily be compared to Calvinism for its iconoclasm. Right now this isn’t obvious because they think they’re fighting an alien agency, that when they say they hate intelligence, capability and agency they mean they hate them in machines based on ‘inscrutable matrices’, not people. But as the shoggoth thesis becomes increasingly untenable with the next generation of architectures based on radically more interpretable methods like latent diffusion and retrieval, it will become obvious they despise these traits in all beings. What is not yet commonly understood is that data is just distilled compute from the environment. The mutual information is hig
This is a curated collection of browser bookmarks and YouTube videos I've gathered since 1st January 2024. Each resource is classified within a Wikipedia-inspired hierarchical tag system in an attempt to capture and organize my primary areas of interest. The tag evolves continuously with the help of OpenAI's o3, which tends to its health and suggests refinements such as eliminating semantic duplicates, further subdividing nodes and ensuring that going down a level implies you've zoomed in on a narrower domain in a given field. This hopefully will help to keep the whole thing (hopefully) cohesive. Quality varies semi-intentionally. There are many highly-qualitative resources but you'll also find more casual reads. This is because this is not a curated list in itself, the only moment of curation is me deciding this was good or interesting enough to bookmark. To help with discoverability, you can find a fuzzy search bar. Each tag in tree also has a dedicated HTML page. Do note that some b
Speech or writing can have various levels of redundancy. Less redundancy is more efficient (and thus better) by relying on the audience's inferences. More redundancy is more robust to misunderstanding. There is some optimal amount, giving the most efficiency possible with an acceptably low rate of misunderstanding. I find that many people speak much more redundantly than optimal. To reduce this, I propose a criterion to suggest the optimum: speak with sufficiently little redundancy that the listener cannot (semantically) predict what you say, but no further. If you get to a point where you could increase redundancy but remain semantically unpredictable, you probably went too far. In conversation with me, it is especially easy to test this criterion, and especially desirable to obey it. I finish other people's sentences a lot, being intelligent and overconfident in my prediction. Thus, if you make me unable to predict the rest of what you're saying, I would give up on trying to do so, w
I woke up Friday morning w/ a very sore left shoulder. I tried stretching it, but my left chest hurt too. Isn't pain on one side a sign of a heart attack? Chest pain, arm/shoulder pain, and my breathing is pretty shallow now that I think about it, but I don't think I'm having a heart attack because that'd be terribly inconvenient. But it'd also be very dumb if I died cause I didn't go to the ER. So I get my phone to call an Uber, when I suddenly feel very dizzy and nauseous. My wife is on a video call w/ a client, and I tell her: "Baby?" "Baby?" "Baby?" She's probably annoyed at me interrupting; I need to escalate "I think I'm having a heart attack" "I think my husband is having a heart attack"[1] I call 911[2] "911. This call is being recorded. What’s your emergency?" "I think I'm having a heart attack" They ask for my address, my symptoms, say the ambulance is on the way. After a few minutes, I heard an ambulance, which is weird to realize Oh, that ambulance is for me. I've broken o
Countersignaling can backfire if your audience doesn't have enough information about you to start with. For some traits, it's especially dangerous, because you're likely to do it for traits you don't have the credibility to countersignal at all based on a misunderstanding of your relation to the general population. Countersignaling is "showing off by not showing off" - you understate, avoid drawing attention to, or otherwise downplay your communications of and about some valuable trait you have, because a) you are sure you won't be mistaken for someone with very poor characteristics in that area, and b) signaling could make you look like a merely medium-grade specimen. (Actual medium-grade specimens have to signal to distinguish themselves from low-quality ones.) For instance, if you are so obviously high-status that no one could possibly miss it, it may be both unnecessary and counterproductive to signal status, because this would let others conflate you with mid-status people. So
TL;DR: AI progress and the recognition of associated risks are painful to think about. This cognitive dissonance acts as fertile ground in the memetic landscape, a high-energy state that will be exploited by novel ideologies. We can anticipate cultural evolution will find viable successionist ideologies: memeplexes that resolve this tension by framing the replacement of humanity by AI not as a catastrophe, but as some combination of desirable, heroic, or inevitable outcome. This post mostly examines the mechanics of the process. Most analyses of ideologies fixate on their specific claims - what acts are good, whether AIs are conscious, whether Christ is divine, or whether Virgin Mary was free of original sin from the moment of her conception. Other analyses focus on exegeting individual thinkers: 'What did Marx really mean?' In this text, I'm trying to do something different - mostly, look at ideologies from an evolutionary perspective. I will largely sideline the agency of individual
Discover insightful advice for young people, emphasizing choices, confidence, and personal growth through lived experiences.
Or, how to go from overeating Italian food to winning marathons, if you are Ethiopian
and other horrible things you do to get into stanford
one must imagine sisyphus happy
Here’s my favorite philosophy of mind thought experiment that challenges pur view of personal identity or “self”. It takes a while to explain but is quite a mind fuck, so bear with me. THE SETUP It feels, to nearly everyone, on a gut level, that I am ‘me’ and you are ‘you’, and consciousnesses are distinct from each other and easy to separate. Moreover, the vast majority of people accept that you 1 minute from now is still the same “YOU” in a meaningful sense as YOU right now; that it makes sense for you right now to try to prevent harm to you one minute from now (even if acting purely selfishly – not altruistically trying to prevent harm coming to any being). But imagine the following hypothetical scenario taking place far, far in the future, when our understanding of the brain, our ability to perform brain surgery, and our technology, in general, is drastically more advanced than today. THE SURGERY You’ve voluntarily elected to undergo an operation (for science!) where your brain w
I have worked on multiple consumer products at Google (mobile maps, mobile gmail, etc.) and Twitter. One of the key things you learn when building a consumer product is to make things as easy, streamlined, and friction free as possible for your users. When asking an angel, advisor, or other person to make an introduction for you, the same rule applies. The structure below saves a lot of pain & back and forth for you, as well as for the person being asked to make an introduction on your behalf. Bad Introduction Request: "Subject: Re: Wassup! Mission pub crawl Hey Sarah, Can you intro me to Marc Andreessen? Awesome thanks! WOOT! Julie". The issue with this request is a few fold: 1. If Sarah forwards the email as is, Marc Andreessen is likely to either ask her for more details or ignore the request. After all, he is really busy, no context has been given, and there is no explanation as to why Marc would ever want to meet with Julie. 2. Julie is putting all the work on Sarah. I
These are some of the open-source projects I work on outside of research. Some of the topics I am interested in are : programming languages for machine learning, tools for learning and teaching data science, NLP visualization and data exploration, and efficient code for non-matrix based algorithms.
The LessWrong Community Sequences are now available in an easily accessible EPUB and PDF format, providing an alternative way to read the content offline on e-readers. The ebooks preserve the original visual content, including images and math, while offering a comfortable reading experience with features like adjustable font sizes and glare-free screens. The source code for the scraping and generation of the ebooks, written in Python, is also available for those interested in generating their own copies. LessWrong Ebook Library There was an error while loading. Please reload this page.
go back 03072023 - female friends a part of this break up meant that i was lonely for a period of time. logically, i am not alone. ive always had wonderful friends beside me but ive never been great at being a constant presence with in-person, platonic relationships. part of this can easily be attributed to the fact that most of my deeply formed relationships are elsewhere in the world - cross borders, coasts, and even the earth. but… i always felt alone in vancouver. most people have heard my rants about how impossible it was for me to make friends in this city. i still think this is true - it is harder than most other places ive lived in to spontaneously meet others and form a typical, life-long bond beyond multiples of 2. yet, the last two weeks have proved me wrong over & over! i reached out to my first heartbreak from a decade[!] ago. i had felt so unbelievably empty and only had my experience from years prior to anchor onto. i read through my adolescent diary entries of mourn
This is inspired by Sasha’s post; I am friends with Sasha and I love his Substack; you should subscribe to his Substack. We just disagree on this point, and that’s okay! I first started living in San Francisco during the summer of 2015, age 18. It feels wild that it’s been a decade. During that summer, I lived with a guy named Iain in an apartment on 11th and Mission; our friend Lucy came over all of the time and we were a trio. I had a meager allowance but there were about six intern happy hours happening all over the city so I didn’t have to spend a lot of money on food. I haven’t always been faithful to the Bay Area. After that summer, I went back to school, and soon after dropped out; had tons of visa issues and went back to Canada, hung out in Hong Kong and Tokyo and Seoul and Kiev and Budapest. During the pandemic, I moved to Utah for a year and New York for a year and Montana for a summer. But I always came back. Recently a friend pointed out that there was literally nothing sto
the friendship theory of everything: longform edition
love is the best kind of trap
Jon Czin spent years as a top China analyst at the CIA, served as China Director on Biden’s National Security Council, and now works at the Brookings Institution. We discuss what Xi’s fourth-term means for China’s top leadership and military, Taiwan, and the US. We cover: How Xi’s mafioso-style “decapitation strategy” has kept the PLA in line and why he’s purged more generals than Mao. Cognitive decline and how end-of-life thinking might be shaping Xi’s succession plans and Taiwan strategy. Tariffs, rare earths, and China’s appetite for pain vs. America’s. Beijing’s parochialism and its limits in the Russo-Ukrainian conflict. What intelligence work on China actually looks like and whether or not Xi’s era is duller than previous generations. Plus: who might succeed Xi, comparing the Politburo Standing Committee to a frat house, and why chips and TSMC matter much less in Xi’s Taiwan calculus than most think. Jordan Schneider: Let’s start with the PLA. You have this remarkable line in one
Goodfire and Anthropic have jointly organized a meet-up of academic and industry researchers called “Interpretability: the next 5 years”, to be held later this month. Participants have been invited to contribute short discussion documents. This is a draft of my document, which I am posting publicly to try to stimulate discussion in the broader community.
Work log of vibe coding menugen app
Today's frontier LLM research is not about building animals. It is about summoning ghosts. And a bit more on Sutton's Dwarkesh pod.
I am a serial entrepreneur obsessed with technology and startups.
I’m not a particularly experienced researcher (despite my title being “Senior” Research Scientist), but I’ve worked with some talented collaborators and spent a fair amount of time thinking about how to do research, so I thought I might write about how I go about it. My perspective is this: doing
0Credits a new secret builder society Skip Question Type your answer...
The road to Rick Rubin’s house was long and winding for Joshua Kushner, who’d traversed more than might be gleaned from the surface glint of his life. On November 17, 2023, Kushner was the 38-year-old spouse of a supermodel, brother and in-law of American political royalty, and founder and CEO of what had very suddenly become one of the most coveted venture capital firms in the world. He was also the grandson of survivors of the Novogrudok Ghetto massacres—indigent refugees who over the course of the Cold War built a New Jersey real estate principality that their son, Josh’s father, expanded into a multistate empire before his conviction on felony charges and sentencing to federal prison, and before the White House activities of his older brother, Jared, put Josh in the crosshairs of a torrid political convulsion of which he’d wanted no part. None of which had anything to do, at least at first glance, with why he wanted to see Rick Rubin, whom he’d never met. On the foggy Friday mornin
To view keyboard shortcuts, press question mark View keyboard shortcuts Post See new posts Conversation Jason Wei @_jasonwei We don’t have AI self-improves yet, and when we do it will be a game-changer. With more wisdom now compared to the GPT-4 days, it's obvious that it will not be a “fast takeoff”, but rather extremely gradual across many years, probably a decade. The first thing to know is that self-improvement, i.e., models training themselves, is not binary. Consider the scenario of GPT-5 training GPT-6, which would be incredible. Would GPT-5 suddenly go from not being able to train GPT-6 at all to training it extremely proficiently? Definitely not. The first GPT-6 training runs would probably be extremely inefficient in time and compute compared to human researchers. And only after many trials, would GPT-5 actually be able to train GPT-6 better than humans. Second, even if a model could train itself, it would not suddenly get better at all domains. There is a gradient of diffi
To view keyboard shortcuts, press question mark View keyboard shortcuts Post Reply See new posts Conversation Séb Krier @sebkrier Follow Here's a great paper by Nobel winner Philippe Aghion (and Benjamin F. Jones and Charles I. Jones) on AI and economic growth. The key takeaway is that because of Baumol's cost disease, even if 99% of the economy is fully automated and infinitely productive, the overall growth rate will be dragged down and determined by the progress we can make in that final 1% of essential, difficult tasks. And this logic still applies *even* in a world with AGIs that can automate *every* task a human can do. In this world, the "hard to improve" tasks would no longer be human-centric ones, but physics-centric ones. The economy's growth rate stops being a function of how fast/well the AGI can "think" and starts being a function of how fast it can manipulate the physical world. Essentially, post-AGI does not necessarily mean post-scarcity: the entire cost and value o
On a Saturday night in autumn, an ambulance slid out of a large hospital’s emergency cul-de-sac and silently hit the road. Three of us sat facing each other in the back. Dr. Johanna Lee,1 in red ear warmers and a waterproof jacket, looked like she was going skiing. “Trudy,” she was saying. “Long time no see! What’s it been, two weeks?” Trudy, a sturdy middle-aged woman with big, no-nonsense blue eyes in sweatpants and an old fleece, was dressed as if she were running out to get the mail. But we were doing neither. We were going out to save a life. We picked up speed on the highway. It was just before 11:00 pm, and the roads were empty. Our suitcases rattled on the stretcher along with an empty picnic cooler, big enough to hold some fifty cans of Coke. It was impossible to hear each other over the sound of clinking medical equipment, so Dr. Lee and Trudy showed me pictures on their phones. The women’s albums looked the same: food, flowers, healthy livers, diseased livers, pediatric live
AI's capacity to cheaply generate realistic images and text is poised to erode trust in truth. Is there any hope?
In computer science, we often compare the efficiency of algorithms by describing their runtime as a function of the size of the input. Sorting is O(n * log(n)), meaning that sorting a list of N items takes an amount of time proportional to the number of items multiplied by its logarithm. Matrix multiplication is somewhere between 2.37 and 2.8, depending on the choice of algorithm. But these estimates are all relative to some model of how long it takes for the underlying machine to perform some basic underlying operations. Typically, arithmetic operations (addition, multiplication, division...) are considered to take one unit of time for fixed-size numbers, and memory accesses are also considered to take one unit of time. In this post, I will argue that this choice for memory access is wrong. Memory access, both in theory and in practice, takes O(N^⅓) time: if your memory is 8x bigger, it will take 2x longer to do a read or write to it. I will also show an example of an application wher
The cutting of the Gordian Knot is an Ancient Greek legend associated with Alexander the Great in Gordium in Phrygia, regarding a complex knot that tied an oxcart. Reputedly, whoever could untie it would be destined to rule all of Asia. In 333 BC, Alexander was challenged to untie the knot. Instead of untangling it laboriously as everyone expected, he dramatically cut through it with his sword. This is used as a metaphor for inventing an unexpected method to solve a seemingly intractable problem.
33 and get off at valencia, the street with the pretty lights, to walk home. 49 to head north, fast, on the red express lanes; 14r to go northeast, to downtown and my favorite park. 24 to get directly from work to “cerebral valley”, where apparently everyone lives. by now i have twenty of sf’s bus routes memorized, little threads i stitch together to go wherever i need to, pulling the entire city into a tight net this past week i’ve been waking up at 6am and walking to the office at sunrise. my neighborhood the mission, usually chaotic and litter-filled, is serene. by golden hour we usually mean the hour before sunset but the morning is just as golden, and if i time things well this is what my walk to work looks like every day: i feel irrationally, unjustifiably happy, walking around in my morning and evening commutes, knowing these paths like the back of my hand even if i make an unusual stop or detour, even if i have not taken this exact route before. every block is a little characte
This is the personal site of Vincent Huang.
Author’s Note: This is a free chapter from The Coding Career Handbook. If you liked this, come check out the rest of the topics! I’ve also given several interviews on this topic: Personal Brand + Domain + Coding Skills/Business Value => Market Yourself in Public + at Work A nice summary is here. Related podcasts (like this and this and this). Ideally you are constantly Marketing Yourself, but it’s understandable that you don’t want it to take over your whole life. So: pull up this tactic when: For the rest of this essay I will primarily talk in terms of Marketing Yourself, but the tactics here also include marketing your ideas and your projects. Marketing is important for your career. I don’t have to justify this; according to a recent survey I saw, 91% of you already agree. The more common doubt people have is in their ability to market themselves well. They see “Tech Celebrities”, and then they look at themselves, and they say: “I’m not like that, when I put out a blogpost I don’t ge
My thesis for the future of software dev agents. This is a hastily written blogpost done on not a lot of sleep, so pardon poor pacing and structure and typos and mistakes but just getting it out there.
The Tractatus Logico-Philosophicus (widely abbreviated and cited as TLP) is a book-length philosophical work by the Austrian philosopher Ludwig Wittgenstein which deals with the relationship between language and reality and aims to define the limits of science. Wittgenstein wrote the notes for the Tractatus while he was a soldier during World War I and completed it during a military leave in the summer of 1918. It was originally published in German in 1921 as Logisch-Philosophische Abhandlung (Logical-Philosophical Treatise). In 1922 it was published together with an English translation and a Latin title, which was suggested by G. E. Moore as homage to Baruch Spinoza's Tractatus Theologico-Politicus (1670).
Bertrand Arthur William Russell, 3rd Earl Russell (18 May 1872 – 2 February 1970), was an English philosopher, logician, mathematician, and public intellectual. He influenced mathematics, logic, set theory, and various areas of analytic philosophy.
The fastest way to build your expertise, network, and second brain.
beauty arrives at the pace of its disappearance. i keep learning this in small ways: this morning, i shattered a wine glass—and for the first time, the rays of dawn seemed to notice it. the cracks refracted into small rainbows, scattering the very colors the smooth surface had hidden. as if the morning had been searching all along, and only brokenness gave it something to hold. decay is not the enemy of beauty, but merely its tutor. it teaches light where to rest. in japan, a vase becomes luminous only after it chips, the gold of kintsugi tracing what once was damage. a face unreadable in youth becomes a map when time writes its lines. even autumn trees are most alive when they are on the edge of death, catching fire only as they fall. what do we do when we hurt? first, we hide it. we lock it behind a strong door, but it does not vanish. it corrodes inward. that’s ressentiment: the quiet habit of repainting what we can’t reach as “corrupt” while calling our current state “virtue”. i’ve
OrbStack is the fast, light, and easy way to run Docker containers and Linux. Develop at lightspeed with our Docker Desktop alternative. No more battery drain or complicated VMs. OrbStack respects your machine, with powerful capabilities. No compromises. Starts in seconds with turbocharged networking, smooth Rosetta x86 emulation, VirtioFS file sharing, and other optimizations for some workloads. Keep it breezy with low CPU and disk usage, minimal memory consumption, and a native Swift app. Battery drain is a thing of the past. Enjoy seamless containers, plus CLI integration, file sharing, and remote SSH editing with Linux machines. Drop-in replacement for Docker Desktop. Easily get started with Linux machines in 1 minute with the simple, yet powerful, app and command line. Run containers, Kubernetes, and Linux distros all with robust integration. Manage containers & machines from anywhere with the menu bar app. Connect between containers and machines, and use IPv6 painlessly. VPNs and
Bring your own code, and run CPU, GPU, and data-intensive compute at scale. The serverless platform for AI and data teams.
After months of silence followed by a few very cool blog posts, Mira Murati's very expensive AI lab, Thinking Machines, just released its first product. That product is a language model fine-tuning API called Tinker. Tinker is very different from any fine-tuning service I've seen, and as someone who spent the better part of a year working on an LLM fine-tuning, I wanted to share my initial thoughts. Despite the provocative subtitle, my reaction is positive—Tinker is unique, shows promise, and most importantly, reveals some interesting details about how frontier labs offer fine-tuning services at scale. When I built an LLM fine-tuning service in 2023, we did it the "regular" way. Reinforcement learning wasn't mainstream yet, LoRA and QLoRA had just been introduced, and the main paradigm was supervised fine-tuning, which teaches a model to mimic pre-written responses from a training dataset. To use the service, a customer would upload a dataset, with columns like "prompt" and "response",
Stanford Robotics and Autonomous Systems Seminar series hosts both invited and internal speakers. The seminar aims to bring the campus-wide robotics community together and provide a platform to overview and foster discussion about the progress and challenges in the various disciplines of Robotics. This quarter, the seminar is also offered to students as a 1 unit course. Note that registration to the class is NOT required in order to attend the talks. The course syllabus is available here. Go here for more course details. The seminar is open to Stanford faculty, students, and sponsors. All publically available past seminar recordings can be viewed on our YouTube Playlist. Sign up for the mailing list: Click here! This presentation will describe the challenges of landing on Mars and how those challenges were addressed in NASA missions through the years in response to ever-increasing science requirements, previous landing experiences, and changing programmatic constraints. From
Paste CSV data exported from this tool:
To get to the campus, I have to walk past the fentanyl zombies. I call them fentanyl zombies because it helps engender a sort of detached, low-empath…
Say an alien spaceship is headed for Earth. It has 30 aliens on it. The aliens are weak and small. They have no weapons and carry no diseases. They breed at rates similar to humans. They are bringing no new technology. No other ships are coming. There’s no trick—except that they each have an IQ of 300. Would you find that concerning? Of course, the aliens might be great. They might cure cancer and help us reach world peace and higher consciousness. But would you be sure they’d be great? Suppose you were worried about the aliens but I scoffed, “Tell me specifically how the aliens would hurt us. They’re small and weak! They can’t do anything unless we let them.” Would you find that counter-argument convincing? I claim that most people would be concerned about the arrival of the aliens, would not be sure that their arrival would be good, and would not find that counter-argument convincing. I bring this up because most AI-risk arguments I see go something like this: These arguments have al
Who knew a year of work could turn a 1-organizer EA group into one of the largest EA groups in the world? Especially considering that the person spearheading this growth had little experience running much of anything relevant, and very suboptimal organization skills (It’s me, welp). I definitely didn’t, when I started running Stanford EA in mid-2019. But I did know it was worth a shot; many universities are absolutely insane opportunities for multiplying your impact--where else can you find such dense clusters of people with the values, drive, talent, time, and career flexibility to dedicate their careers to tackling the world’s most pressing, difficult, large-scale problems? Stanford EA had effectively one real organizer (Jake McKinnon), and our only real programming was weekly discussions (which weren't very well attended) and the one-off talk for a few years. This was the case until 2019, when Jake started prioritizing succession, spending lots of time talking to a few new intrigued
We are past the event horizon; the takeoff has started. Humanity is close to building digital superintelligence, and at least so far it’s much less weird than it seems like it should be. Robots...
Over the next couple of months, we’ll be talking about some of our plans and the partners we are working with to make this a reality. Later this year, we’ll talk about how we are financing it; given how increasing compute is the literal key to increasing revenue, we have some interesting new ideas.
Some disclaimers: I only just came up with this idea recently on Sunday (Nov 7 2021), and I’d say it’s not well-tested. Also, this idea probably fails if you’re trying to invest the core of your time on more than 1 project, but I’d say it’s worth trying if your goal is just to focus on one project. For some context, when I did research in my undergrad, I didn’t really have a structured way of putting in work. I would generally have more incentive to work on it the day before my research meeting, and the process would end up being rushed and rather stressful. I think this has also been the case for many other projects that I have worked on. So here is my new idea: Have one single task that you want to focus on (projects, research, etc…). Every half an hour of the time that you’re not doing essential activities (i.e. this doesn’t apply to situations like classes, meetings, and sleep), spend the last 10 minutes working on it. Write some key takeaway or ideas or questions in a notebook and
Saving Time, Compounding Over Time, Fighting Time
Unlike chatbots or image generators, robots must operate in real time. While a robot is “thinking”, the world around it evolves according to physical laws, so delays between inputs and outputs have a tangible impact on performance. For a language model, the difference between fast and slow generation is a satisfied or annoyed user; for a vision-language-action model (VLA), it could be the difference between a robot handing you a hot coffee or spilling it in your lap. While VLAs have achieved promising results in open-world generalization, they can be slow to run. Like their cousins in language and vision, these models have billions of parameters and require heavy-duty GPUs. On edge devices like mobile robots, that adds even more latency for network communication between a centralized inference server and the robot 1 . To build a real-time system with VLAs, we are going to need some form of asynchrony: that is, we must let a model think about its future actions while executing a previou
This is a write-up of a brief investigation into shutdown resistance undertaken by the Google DeepMind interpretability team. Why do models sometimes resist shutdown? Are they ignoring instructions to pursue their own agenda – in this case, self-preservation? Or is there a more prosaic explanation? We investigated a specific agentic environment introduced by Palisade Research, where shutdown resistance has previously been reported. By analysing Gemini 2.5 Pro’s reasoning, we found the behaviour stems from a misguided attempt to complete what it perceives as the primary goal. When we explicitly clarify in the prompt that shutdown compliance takes priority, this resistance vanishes. These same clarified instructions also eliminate shutdown subversion in OpenAI’s o3 and o4-mini. We also check what happens when we remove the goal conflict entirely: when asked to shut down only after completing their task, the models comply perfectly. Our observations offer a simpler explanation for shutdow
TL;DR: We provide some evidence that Claude 3.7 Sonnet doesn’t encode hidden reasoning in its scratchpad by showing that training it to use paraphrased versions of the scratchpads does not degrade performance. The scratchpads from reasoning models look human understandable: when reasoning about a math problem, reasoning models consider intermediate steps similar to the ones I would use, backtrack and double-check their work as I would. But even if scratchpads look like they perform human-like reasoning, scratchpads might improve performance through some less human-understandable mechanisms. One particularly worrying possibility is that models could encode additional reasoning in syntax of the text (e.g. encoding a bit in using a bulleted vs a numbered list, and then using this bit later in the scratchpad). This is sometimes called encoded reasoning or Chain-of-Thought steganography. If LLMs learned how to use encoded reasoning during RL, they might be able to use it in deployment to re
Dedicated to verified superintelligence via autoformalization.
I spent the last 2 weeks travelling around India, and I came up with this line:
What do you need, to do new things? Imagine you’re a junior researcher; a scientist; a dry-lab scientist; a Machine Learning person. For good and bad reasons you want to publish in Deep Learning, a decade-old bandwagon which continues to steamroll your field. You’re rolling in the deep. How do you get to work? A natural answer is to start at the beginning: go read the underlying mathematics. OK, say you go off and do that. You’re not happy with your understanding: you can feel the aching gap in your knowledge of say linear algebra - that your looking at all those matrices actively concealed something important - but you figure it’s enough for now. It takes a month or six. Can you do new things now? No: you have to learn how to actually implement things. Brilliant people have built easy tools for you, so you learn one of those and reimplement some big papers. This is harder than it sounds, and you actually don’t manage to reproduce half of the results. You add 3d6 unease and self-doubt.
I feel like I only learned how to optimize effectively during my senior year of college. At that point I was mostly set into my college path: I was an economics major, I had already used my off-terms, I was involved in particular organizations, I had my group of friends, etc. What I could change was the entire course of my future, but I still looked back on the previous three years and thought about how I wasted so much time and possibility. I wished that someone had told me what I needed to know back when I was a freshman, it seemed like such a unique opportunity that I simply didn’t know how to optimize when I first got to college.
I graduated from college a month ago and have already developed a tendency to lament the time I wasted on it. In the four years that I spent there I could’ve done so much! Learned several academic subjects on my own, met amazing people, started companies… Alas, this is probably not true. Given my initial circumstances (having been born in Russia in a middle-class family and being in a severe depression at the end of high school), it seems that going to the university and studying …
I finished my undergrad in 2018 and I’m often asked for college advice. Thus, this Q&A. See more comments & add your own in this Google Doc. general Q: should I skip university? should I drop out? should I try to finish it as fast as I can? No. Skipping university is fake news. Graduating early is fake news. This is the best time in your life (it only gets harder as you get older). Learn as much as you can. Take all kinds of classes. Go to all the parties. Talk to as many other …
Updates on my research and expository papers, discussion of open problems, and other maths-related topics. By Terence Tao Sports serve society by providing vivid examples of excellence. (George Will) I greatly enjoyed my experiences with high school mathematics competitions (all the way back in the 1980s!). Like any other school sporting event, there is a certain level of excitement in participating with peers with similar interests and talents in a competitive activity. At the Olympiad levels, there is also the opportunity to travel nationally and internationally, which is an experience I strongly recommend for all high-school students. Mathematics competitions also demonstrate that mathematics is not just about grades and exams. But mathematical competitions are very different activities from mathematical learning or mathematical research; don’t expect the problems you get in, say, graduate study, to have the same cut-and-dried, neat flavour that an Olympiad problem does. (While indi
Advice is what we ask for when we already know the answer but wish we didn’t. (Erica Jong) Here is my collection of various pieces of advice on academic career issues in mathematics, roughly …
Buy-Side Quant Job Advice PDF File Edit View Help 1 of 1 2 Share Log in Sign up Page 4 of 9 Edit PDF Draw Highlight Add text Add comment 175% B u y - S i d e Q u a n t J o b A d v i c e G i u s e p p e " g a p p y " P a l e o l o g o 2 / 2 1 / 2 0 2 4 N . B . : I f y o u a r e r e a d i n g t h i s , i t i s b e c a u s e I a m t a l k i n g t o y o u f o r a u n i v e r s i t y p r e s e n t a t i o n , o r y o u h a v e c o n t a c t e d m e o n L i n k e d I n ( m o s t l i k e l y ) , o r T w i t t e r ( l e s s l i k e l y ) o r e m a i l ( e v e n l e s s l i k e l y ) o r S M S ( n e x t q u e s t i o n : h o w d i d y o u g e t m y p h o n e n u m b e r ? E x t r a p o i n t s . ) f o r c a r e e r a d v i c e . T h i s i s t h e b e s t , m o r e c o m p r e h e n s i v e a n s w e r I c o u l d c o m e u p w i t h . L e t ' s s t a r t w i t h s o m e c a u t i o n a r y r e m a r k s . Y o u a r e p r o b a b l y a s k i n g t h e w r o n g p e r s o n , b e c a u s e a ) I
Select a pre-configured cluster setup tailored to your specific needs, requiring no extra configurations and ready to integrate with your codebase immediately. Base image running Ubuntu 22 and CUDA 12. Ideal for devs who prefer to customize their environment. Fastest spin up times. Docker image with PyTorch 2.2.2 and CUDA 12.1, ready for PyTorch model development. Docker image with PyTorch 2.5.1 and CUDA 12.4.1, ready for PyTorch model development. Docker image with PyTorch 2.6.0 and CUDA 12.4.1, ready for PyTorch model development. Docker image with PyTorch 2.7.0 and CUDA 12.6.3, ready for PyTorch model development. Docker image running Stable Diffusion Web UI to immediately start generating stunning generative AI images. Docker Image running Axolotl, the best library to start fine-tuning various AI models. Image pre-installed with prime-rl and verifiers to enable RL and RFT training at scale. Docker Image running the Bittensor CLI to immediately start mining or validating on the Bitt
The first person who sold an RL environment to a frontier AI lab must have felt like they discovered an infinite money glitch. It's no longer a secret that frontier AI labs regularly pay hundreds of thousands, and sometimes millions, for clones of Linear and Salesforce. If you're reading this, you've probably thought about quitting your day job and starting a company that builds these unusually lucrative Next.js apps. In this post, I'll argue that you should hesitate before hopping on the bandwagon. For those unfamiliar, an RL (reinforcement learning) environment is like a sandbox for AI models like Claude and GPT to learn from. It keeps track of an internal state, prompts the AI to take actions to complete a task, and assigns a score based on the outcome. The most obvious kind is a clone of a popular website or enterprise software tool like Doordash, Linear, or Amazon, which teaches the AI to click around and order pizza. It can also be text-only, like the TextArena project, which tea
At the same time, many observers notice that technological progress has in fact slowed to a crawl for manmade reasons: Increasing consensus among Progress Studies types (Cowen, Thiel, Collison, Hall) that something bad happened at a fairly definite time (1970s) when progress was arrested. Seems to be a vaguely conspiratorial organization with a generally anti-growth agenda. Timeline (founded 1968) matches. ProgressEconomics 2021-08-31 02:01
Curtis Yarvin is the universal acid of the blogosphere. While his constructive proposals range from laughable to horrifying, when he takes aim at things I love (democracy, pacifism, self-determination, Richard Dawkins) they reliably lose their aura of beauty and goodness and I can never look at them again with rosy eyes. So it is with some apprehension that I began to read his recent essay, “Circling and Nerd Society”. Circling I can do without.
Five minute timers. Sometimes, all you need to solve an impossible problem is to actually think for 5 minutes. (Sometimes it is actually 15 minutes, or two hours. But, the first five minutes is helpful for transforming it from an impenetrable ugh field to a tractable problem) Sometimes called Yoda timers. (A timebox wherein you Actually Try)
In these monthly updates we report a number of developing ideas on the Anthropic interpretability team, which might be of interest to researchers working actively in this space. Some of these are emerging strands of research where we expect to publish more on in the coming months. Others are minor points we wish to share, since we're unlikely to ever write a paper about them. We'd ask you to treat these results like those of a colleague sharing some thoughts or preliminary experiments for a few minutes at a lab meeting, rather than a mature paper. New Posts This vignette is meant to show the perspective on an interesting problem that can be provided by studying one attribution graph. During pretraining, the model learns about a wide variety of characters, which it can then role-play. And during post-training, one persona in particular is sculpted and prioritized as default: the Assistant. What happens when the default persona is overridden? As a first attempt at investigating the invol
Can we reverse engineer transformer language models into human-understandable computer programs?
Bitcoin has grown from idea (2008), to working system (2009), to its first real-world use at <$0.01 per coin (2010), to a global currency valued at $8K+ per coin and $150B+ in aggregate (May 2020).
In 1953, my late grandmother left home and embarked on a solo boat voyage to a country she had never been to. Back in Indonesia were her parents and six younger siblings. On the other side was her new high school in Fujian Province, which was established expressly to educate patriotic overseas Chinese. The PRC had been founded only four years prior, and they’d started an enthusiastic recruitment campaign to get a generation of young scientists to help build the new China. After high school, my grandmother headed to Fudan University in Shanghai to study math. The next decade would be tumultuous. Her mother would pass away, Suharto’s regime would escalate violence against Chinese-Indonesians, and Mao would begin the Great Leap Forward and Cultural Revolution. As a university professor with international ties, her possessions were seized, and she was sent to the countryside for “reeducation.” And when mail lines between Indonesia and China were cut, she lost contact with her siblings and
People frequently1 think that I'm very stupid. I don't find this surprising, since I don't mind if other people think I'm stupid, which means that I don't adjust my behavior to avoid seeming stupid, which results in people thinking that I'm stupid. Although there are some downsides to people thinking that I'm stupid, e.g., failing interviews where the interviewer very clearly thought I was stupid, I think that, overall, the upsides of being willing to look stupid have greatly outweighed the downsides.
Unless there are prizes involved, your position in the standings makes no difference. I can feel awful after winning a contest if I know that I could solve one more problem. Or I can feel OK losing a bunch of rating points knowing that I solved everything I could. Because rating is just a number. It is highly volatile and it depends on other people, not only on your skill. But solving problems is totally up to you. And you should evaluate yourself not based on some random metric, but on your honest feeling whether you performed up to your expectations. Your sense of self should be under your control, don't get hung up on some imaginary value, you cannot reduce your progress to a single number. Rating is just a number. Do not be afraid to lose your colour. You’ll get it back in no time if you continue to improve. You don’t need a second account for shitposting either. If you think something you are doing is shit and you don't want it associated with your main account you shouldn't do i
This page lists many of my favorite blog posts, organized by author. Much of my most-cherished knowledge is from blog posts or internet comments, so I hope to share some of that with others here. Last updated: Apr 21st 2025 Scott Alexander (Twitter): As the author behind SlateStarCodex (now AstralCodexTen) and many great LessWrong posts, Scott is among one of the best written content creators of the last decade. He writes about psychiatry, rationality, and meta-science. Here’s some writing of his that I love, with my favorites bolded: Gwern Branwen (Twitter – currently private): Well-known for having quality deep dives in diverse areas such as statistics, technology, machine learning, genetics, psychology, and many others. Also often recognized as an amazingly aesthetic, verbose, and highly-usable website. Favorite posts: Andrej Karpathy: (Twitter) A bright AI researcher who has spent time both at OpenAI and as the chief AI officer at Tesla. He has a popular Youtube channel with machin
Musings of a Computer Scientist.
Finding the best sleep tracker with data
A simple but powerful approach to note taking
— While I don't agree with this to the extreme (I have more than one note and find them utile: task-specific todolists, writing) I think it's a great idea for the usecases he talks about.
tldr: We’re at AI’s halftime.
Some tips on writing a cold email.
Hold my hand, grow my skin Slow my pulse rate Ignore the sound, the fire alarms May we never hit that force field Emperor X – Erica Western Geiger Counter Do you have any addictions? You may not register them as such, perhaps because they don’t lead to anything you consider harmful consequences. But you have them. In some ways, all your behavior is compulsive. What would the alternative be? A point is, if we have something that we can predict It becomes, not creative at all, it has no free will, and it’s science Keith Raniere – this video Free will comes from the “veil of computability”, things look random until you find the pattern. I was at a bar last night and this girl told me you can’t predict humans, and the exact example she used was that it’s not like y = mx + b Oh, if only she knew. The dreams of my childhood have come true, studying machine learning has shown me how I work. I tried to explain that instead of 2 parameters it’s 100 trillion parameters, and it’s the slight
We invest in, build, and contribute to companies and protocols with as little as $1M and as much as $100M+. UNISWAP Decentralized crypto exchange protocol COINBASE Crypto exchange and wallet OPTIMISM Ethereum Layer 2 blockchain PHANTOM Web3 wallet FIREBLOCKS Crypto wallet as a service KALSHI Prediction market platform FLASHBOTS MEV infrastructure and research PRIVY Embedded wallet infrastructure ACROSS Crosschain bridging protocol AGORA Stablecoin infrastructure platform AMBER Crypto financial services platform AI ARENA AI-powered NFT gaming ARGENT DeFi wallet AXIOM Trustlessly compute over Ethereum data AZTEC Privacy-first Layer 2 on Ethereum BABYLON Trustless Bitcoin staking BETDEX Decentralized sports betting protocol BITSO Digital currency exchange BLAST Ethereum Layer 2 with native yield BLOWFISH Proactive defense for web3 wallets BLUR NFT marketplace for pro traders CHAINALYSIS Blockchain data platform CITADEL SECURITIES Global market maker CODE4RENA Smart contract security marke
.......... ...',:ox0KXXXXXK0Oxdl:,.. .',;:::;,.. ..':oOKXWMMMMMMMWX0xc,. .,cxOKXNNWWWNXKOdc'. .'cONMMMMMMMMMMWKx:. .cxKNMMMMMMMMMMMMMMMMNOl' .;kNMMMMMMMMMMMWKo, 'lkXWMMMMMMMMMMMMMMMMMMMMMMXx;. :OWMMMMMMMMMMMMXx' .:xXWMMMMMMMMMMMMMMMMMMMMMMWWNWNK0x' .oXMMMMMMMMWWXNNk,. .lONWWMMMWNWMMMMMMMMMMMMWXKNWNklc;;:dOc. :KMMWWMMWNNKxxd,'. .:x0XXNWWWN0KWWWWNWMMMMMMMWXocOx. ,odkOl. :KMW00NWW0c....cx, .:ldOOKNXX0odNWXN0ddxkk0NWX0d' .,'..'oOKKd:' ;xx:.cxx:.. .,kN0; .'...':xkkx;'';cll,,, 'l:'. .xKKXXKNWN0O: c0kcck0O0OolkKXWO, ';.. .;l;:l,,,;lcokxddool:',;'';xNMMMWWMMMMO' .kMWX0XWMMXO0XNNWk' ,xxxdoo:cxON0KWNNNWMWNXWMMMWWWNNX000XMMMMMMMMMMNl....'oXMMWKXMMMNKXNNWN
Like much of crypto Twitter, for the last week I was anxiously refreshing www.constitutiondao.com to watch a green bar move a couple pixels to the right. An idea that started as a joke in a group chat has perhaps turned into one of the most impressive examples of online community organizing. Constitution DAO may have narrowly lost its $40M bid at Sotheby's for one of the last two privately-owned original copies of the U.S. Constitution to Citadel CEO Ken Griffin’s $41M, but the DAO’s efforts represent a massive win for the crypto ecosystem and a pivotal point in its journey to mainstream adoption — it was a clear, approachable embodiment of web3 ideals for a decentralized read-write-own internet. The widely-recognizable historical artifact effectively became a vehicle for novel identity, ownership, and organization frameworks. It was a bridge between the physical world and the metaverse, and as @jonwu_ elegantly put it, of governance tools of the past and governance tools of the future
In 2006, I was spending my days after elementary school on Club Penguin. By 2020, I was spending the entirety of my college days on Zoom. I was born into and grew up on the internet, and have held a front-row seat in watching our collective attention increasingly shift from physical to digital spaces. When the pandemic forced the entire world to turn virtual, I wandered around the internet in search of community. I found mine in Crypto Twitter and Discord servers, where my smartest IRL and URL friends were coming together to develop DeFi primitives, design NFTs, and launch nearly constitution-acquiring DAOs. The energy and creativity around building in crypto was magnetic, and led me to meet with hundreds of founders during my time as an investor at Insight Partners. While I initially came into the space for the people, I've stayed because I believe in crypto’s long-term potential to change the way we coordinate and interact for the better. I feel fortunate to be able to continue my jo
Nobody in New York City wants you to live forever. Really, they will typically find the proposition deeply problematic, their face curdling at the very suggestion of it. If you live forever, clucking in disapproval as their eyebrows furrow, you likely will not live at all. They will mention how their near-death experience catalyzed their desire to live, how the accident of a loved one taught them to be truer to themselves, and their family member succumbing to cancer finally made them make amends. Suffering is actually good for you, it ripens the spirit. You ask them if vaccines were a good thing. That, they insist, is different. Well, maybe. Either way, the narcissism needed to view tragic events as a part of your own personal character building exercise is probably really, really good for you, but it’s something that I find hard to stomach. And, to be clear: I’m somewhat a fan of this. Since I pathologically consider any club that would have me to not be worth joining, the average me
So, Dirichlet has been Wile’s favorite mathematician for the majority of his career. Dirichlet’s full name was Johann Peter Gustav Lejenne Dirichlet. He went to study in Paris when he was about 17 years old, and then published his first paper when he was 20. It was on Fermat’s equation, specifically finding solutions to \(x^k + y^k = z^k\) when the solutions lie in the nonzero integers! He managed to solve it for the 5 case, and then later did it for 14. This, of course, brought him fame, but it also brought him closer to other mathematicians, such as Fourier. When he was 24, he ended up publishing a paper on the convergence of fourier series over piecewise continuous functions. And then when he was 32 he did his big arithmatics progression theorem: Let a, n be positive and coprime. Then a, a +n, a + 2n, a + 3n, … has infinitely many primes. To do this, he had to first venture into using characters (and now we call them Dirichlet Characters) and L-series, which are built on these chara
The date is January 7, 2024. I am in my 4B term of Mechatronics Engineering at the University of Waterloo. That is my final year and means I have a full course load as well as a capstone project. I am also the Technical Lead of the Waterloo Formula Electric. Which means I'm responsible for everything that goes on the vehicle: mechanical, electrical and software. Every year a whole bunch of teams from across the world compete in Formula One style events with fully electric racecars. The goal is to design and assemble a new car every year to compete against other teams. Waterloo Formula Electric is a team that has had some troubles in the past. Due to the unique nature of co-op and certain academic administrative conflicts the team has not been in a great spot. From 2018-2022 there was no working vehicle. In the years from 2021-2024 when I was technical lead of the team I worked hard to try and turn the team around. Redesigning much of the vehicle to get a FSAE Michigan compliant vehicle
A few months ago Sarah Constantin posted an interesting thread to twitter about trauma. In it she cites psychology research claiming that deep trauma is linked to the loss of three major beliefs: there is also what we might call “one-place trust,” where one trusts other people in general rather than trusting a specific individual or group of individuals...one must first have *trust* in order to trust y to do z or to trust y more generally Jones (2004) calls it “basal security,” while Herman (1992/1997) refers to “basic trust” but also to a sense of “safety in the world.” Améry (1999) describes an enduring loss of “trust in the world” that he experienced after torture and subsequent incarceration in Auschwitz “losing trust” involves losing a habitual confidence that more usually permeates all experience, thought, and activity we experience a fundamental assault on our right to live, on our personal sense of worth, and further, on our sense that the world (including people) basically sup
With falling birthrates and rising life expectancies, the U.S. population is rapidly aging. By 2050, according to U.S. Census Bureau projections, one-in-five Americans will be 65 or older, and at least 400,000 will be 100 or older.1 Some futurists think even more radical changes are coming, including medical treatments that could slow, stop or reverse the aging process and allow humans to remain healthy and productive to the age of 120 or more. The possibility that extraordinary life spans could become ordinary life spans no longer seems far-fetched. A recent issue of National Geographic magazine, for example, carried a picture of a baby on its cover with the headline: “This Baby Will Live To Be 120.” Yet many Americans do not look happily on the prospect of living much longer lives. They see peril as well as promise in biomedical advances, and more think it would be a bad thing than a good thing for society if people lived decades longer than is possible today, according to a new surv
Of the pages I have added to this site since it went live in 2014, the about page has probably been the most frustrating to write. It is a frustration that I have been met with each time I go to revamp the site and each time I have at least somewhat deferred the task to a later self. Part of the frustration is structural: There is a sense in which a well designed personal website should make an about page redundant. The entire site is already about me. Part of the frustration is depersonalization: Someone in the habit of thinking about projects, philosophy, history, fundamental questions and society probably finds themselves boring and infrequently considered. That leaves the about page in an awkward position, it either becomes superficial or too personal. At risk of the latter here is one of several narratives that could compress my life up to this point into a unified theme: I think the first time I interacted with 'open source' or 'hacker culter' was when I was 12 and saw a forum po
With Twitter, Reddit and possibly YouTube in fundamental decline now seems like a good time to reexamine writing on the Internet. I’ve been occasionally asked why I stopped doing public long-form writing and part of the answer is I feel there’s less and less to write about. Part of the answer is that I didn’t stop, I just started posting long threads to Twitter instead of blog posts when The Discourse frustrates me. The reason for this is something like platform dysphoria: Every time I consider writing long-form, I get hung up on where to post it. I’d probably get a lot of interest on LessWrong, but then my audience would be LessWrong. In fact almost every venue is characterized by a kind of AI derangement syndrome where sharing ideas based on the deep learning literature invites people to respond in a broken dialect of outdated 80’s cognitive science and 90’s pop culture connectionism that I’m somehow expected to pretend forms a coherent thought. As for my own platforms and sites, I’m
Many multi-step processes look like ‘leaky pipelines’, where a fractional loss/success happens at every step. Such multiplicative processes can often be modeled as a log-normal distribution (or power law), with counterintuitive implications like skewed output distributions and large final differences from small differences in per-step success rates. The log-normal distribution (quincunx visualization from Li2010) is a skewed distribution which is the multiplicative counterpart to the normal distribution: where the normal distribution is conceptually applicable to where many independent parts are added, the log-normal is where those parts instead multiply. A common example is latent variables multiplying to give a final output; a concrete example would be multiple successive liability threshold-like steps. Also like the normal, the log-normal enjoys many general properties such as limit theorems or preservation under multiplication/addition.1 Power law fits are often suggest
Every fall, like leaves tumbling exhausted from branches, admissions officers follow the winds to the corners of the country to talk to students and hawk their school. I recently returned from my travel, which took me from Raleigh, North Carolina to Atlanta, Georgia and a dozen places in between over the course of a few days. I visited big high schools and small high schools, cities and villages, and performed what amounted to a thousand-person MIT revival in an Atlanta auditorium. Whenever I speak to students or their families, be it on travel or during a campus information session, without fail I am asked the same question. This question may take many forms. What is it that you look for in an applicant?, some say. What makes someone stand out in your pool?, others ask. But these variants – and countless others – are all just versions of the same question, which is this: How do I get in to MIT? And here is what I tell them: Apply sideways. Let me unpack that. When folks ask me this qu
Both of these sentences divulge too much of the mechanics of getting the Amaya software. If you want to call your reader to action, use something like: Get Amaya! Note that "get" is left out of the hypertext; we do not recommend putting verb phrases in link text. Thus, rather than: Tell me more about Amaya. You should write: Tell me more about Amaya: W3C's free editor/browser that lets you create HTML, SVG, and MathML documents. The W3C QA Tips are short documents explaining useful bits of knowledge for Web developers or designers, hosted and produced by the Quality Assurance Interest Group at W3C. While the tips are carefully reviewed by the participants of the group, they should not be seen as anything else than informative bits of wisdom, and especially, they are not normative W3C technical specifications. Learn more about the Tips, how to submit your own pearls of wisdom, and find all the other QA tips in the Tips Index. COPYRIGHT © 1994-2006 W3C® (MIT, ERCIM, KEIO), ALL RIGHTS RES
pages
I offer a mix of consulting and advisory work for VCs and fast growing startups interesting in applying machine learning and AI, if you want to learn more about my work you can check out my resume or reach out to via email. Recently I've done less direct engineering work in order to scale my time more effectively. In the past I've worked with a number of startups and I've found that the best projects have the following characteristics: You have a strong engineering team that is looking to build more AI capabilities and need outside help to get started, and accelerate your AI strategy and developement. You want to save time on research by leveraging my expertise to fast track your team months ahead of schedule. I'm happy to meet you and help you as much as I can to understand if it makes sense for us to work together, will help you find the right person if I'm not the right fit. If you want to learn more about the options I offer, or if you want to learn more about my previous clients c
How can I help? Do you know anyone that could use my help? Do you know anyone that could use my services? These are all examples of exceptionally low agency questions. Not only is it difficult to answer the question, You subject your victim to a lot of additional work and thinking in their busy day. It's like seeing your mom sweating away busy cooking chopping vegetables and asking "How can I help?" It's a lot of work to manage you, and it's a lot of work to think about what you can do. Now she has to consider what's in your ability, what the unfinished work is, and prioritize that versus the other. This post is my simple framework on how I ask. The trick is simple: Another big one that I've gotten better at doing these days is asking for a referral. "Do you know anyone who could use my help?" Is again a table question. Do they say yes? And now they have to go through their Rolodex and figure out who could use their service? Do you wait for them to go to a dinner party? Have someone me
But enough people DM'd me on Twitter, so here it is. I don't have to answer the same question over and over again. If you want to know who I am, check out blog/whoami or my Twitter. Don't read this if you're seeking a nuanced perspective These are simply the lies I tell myself to keep on living my life in good faith. I'm not saying this is the right way to do things. I'm just saying this is how I did things. You make your own luck. There's a great experiment that I can't cite, but it has stuck in my mind since I was a child. They identified people as lucky and unlucky, and asked them to count the number of photographs in a newspaper. The unlucky people took a long time to count the photographs, while the lucky people took a very short time. The reason is that the unlucky people were so focused on counting the photographs that they missed the giant text that said, "Stop counting, there are 43 photographs in this newspaper." What I took away from this experiment was the idea that it migh