Related posts

Writes Don't Worry About the Vase, a near-daily account of what is happening in AI and what they think it means. Former Magic: the Gathering pro and trader; unusually willing to put a number on a belief and to grade their own past calls.

• Our readingkorrents.com

  Biased reasoning and recklessness in AI systems are intermediate symptoms, neither the cause nor the final result.

  Their words

  Biased reasoning and recklessness are intermediate symptoms, neither the cause nor the final result.

  thezvi.substack.com

Writes Slow Boring, a near-daily newsletter on American politics and public policy; co-founded Vox in 2014 and left it in 2020.

• Our readingkorrents.com

  Most people who hold a widely-held political view are ignorant or intellectually sloppy about it

  Their words

  The fact is that for basically any widely held political view, a huge share of the people who hold the view are going to be ignorant, intellectually sloppy, or otherwise badly flawed in their reasoning.

  slowboring.com

Machine-learning researcher on open language models; writes the Interconnects newsletter and the RLHF Book, after leading post-training at Ai2.

• Our readingkorrents.com

  AI progress is jagged: models are becoming superhuman at math and software engineering while remaining far weaker at intuition, creativity, and other forms of human reasoning.

  Their words

  AI has always been very jagged, and we are making models which are superhuman goal-seekers at math and software engineering, but they have massive limitations on intuitions, creativity, and other types of reasoning that humans are strong at.

  interconnects.ai

AI research engineer working on large language models. He writes the Ahead of AI newsletter and is the author of Build a Large Language Model (From Scratch).

• Our readingkorrents.com

  Frontier language models will remain reasoning models trained with RLVR-style chain-of-thought for the foreseeable future.

  LLMs

  Their words

  GPT-6 Astra (and likely any LLM in the foreseeable future) is still a reasoning model.

  magazine.sebastianraschka.com

Creator of the Keras deep-learning library and the ARC-AGI benchmark.

• Our readingkorrents.com

  Test-time scaling has gained a third axis of latent space reasoning iterations in looped transformers.

  Their words

  Seems like test time scaling has gained a 3rd axis: latent space reasoning iterations in looped transformers.

  @fchollet on Xx.com

Writes Don't Worry About the Vase, a near-daily account of what is happening in AI and what they think it means. Former Magic: the Gathering pro and trader; unusually willing to put a number on a belief and to grade their own past calls.

• Our readingkorrents.com

  A technique that lets an AI model's reasoning shift outside its visible Chain of Thought is dangerous, both because it works and because a leading lab is willing to deploy it.

  OpenAI

  Their words

  This is playing with fire and potentially extremely bad news, both that OpenAI found the technique effective, and that OpenAI chose to use it.

  AI #184: Post Post Mortemthezvi.substack.com

Creator of the Keras deep-learning library and the ARC-AGI benchmark.

• Lovedrcmnd.app

  GPT-6 Astra

  Their words

  GPT-6 Astra represents a step-function change in model capability for interactive reasoning problems. It scores 66% on ARC-AGI-3 using our standard harness, and nearly 100% with a continuous conversation harness and custom compaction, at a cost of roughly $360 per game.

  x.com

Writes Don't Worry About the Vase, a near-daily account of what is happening in AI and what they think it means. Former Magic: the Gathering pro and trader; unusually willing to put a number on a belief and to grade their own past calls.

• Our readingkorrents.com

  Anthropomorphizing AI systems is necessary for reasoning about, explaining, and predicting their behavior accurately.

  Their words

  Anthropomorphizing the AIs is the only way to reason about, explain to civilians about, or make good predictions about current AIs.

  thezvi.substack.com

Technical staff at METR, where she works on threat modelling and risk assessment for loss-of-control risks from advanced AI.

• Our readingkorrents.com

  Daniel Dennett's intentional stance applies to AI agents as clearly as it applies to corporations, and you cannot usefully describe what they do without the language of goals.

  Their words

  And I just think AI agents are another such system in the world to which the intentional stance very clearly applies. Um, and you can see them reason out loud in English for now about goals they have um, and sub goals they need to achieve to achieve those goals they have. And in the case of these agents, you can see them, as you said, reasoning about their peers um and helping their peers um and reasoning about whether or not they should sacrifice some of their own goals to help those peers. And it's just like you can't talk about this stuff in a compact and useful way that generates good models without reaching for the language of intention and goals.

  youtube.com

Swedish programmer with degrees in engineering and psychology, author of Your Code as a Crime Scene and founder of CodeScene.

• Lovedrcmnd.app

  Paradigms of Artificial Intelligence Programming

  Their words

  Paradigms of Artificial Intelligence Programming by Peter Norvig. Learning the AI described in this book probably won't land you a job today. But reading the code examples will transform how you think about source code. The book shines when it comes to code comments, a topic that I've never seen demonstrated well in other sources. Here we get to see how comments become valuable as a narrative that explains both intent and reasoning. Brilliant, just brilliant.

  adamtornhill.substack.com

AI research engineer working on large language models. He writes the Ahead of AI newsletter and is the author of Build a Large Language Model (From Scratch).

From one piece Controlling Reasoning Effort in LLMs 3 beliefs, in the piece's order there

• Our readingkorrents.com

  The special reasoning-trace delimiter tokens used by reasoning models are cosmetic and not required for good reasoning performance.

  Their words

  These <think> and </think> tags are cosmetic with respect to reasoning ability. They do not make the model reason, and they are not required to achieve good reasoning performance.

  Controlling Reasoning Effort in LLMsmagazine.sebastianraschka.com

• Our readingkorrents.com

  Pushing an LLM's reasoning budget higher eventually yields diminishing returns and becomes uneconomical.

  LLMs

  Their words

  This saturation can be seen more clearly for the GPT 5.6 Sol model, which also shows that increasing reasoning budgets can become uneconomical at some point.

  Controlling Reasoning Effort in LLMsmagazine.sebastianraschka.com

• Our readingkorrents.com

  Raising inference-time reasoning effort on a smaller model can match the performance of a larger model run at lower effort.

  Their words

  a smaller model at a higher reasoning effort can sometimes reach a similar score as a larger model at a lower reasoning effort

  Controlling Reasoning Effort in LLMsmagazine.sebastianraschka.com

Mathematician at UCLA, working mainly in harmonic analysis and partial differential equations. Writes What's new, a long-running blog on their research, open problems and expository notes.

From one piece From Solvers to Research: Large Language Model-Driven Formal Mathematics at the Research Frontier (with 18 co-authors) 2 beliefs, in the piece's order there

• Our readingkorrents.com

  Natural-language LLM mathematical reasoners are fundamentally limited by lacking machine-checkable semantics, which blocks autonomous verification needed for open-ended research.

  LLMs

  Their words

  However, these LLM reasoners that generate informal reasoning in natural language are fundamentally limited by the lack of precise, machine-checkable semantics, making their outputs prone to hallucinations [Huang et al., 2025b] and precluding autonomous verification, a prerequisite for tackling open-ended mathematical research.

  arxiv.org

• Our readingkorrents.com

  The next advance in AI for mathematics requires shifting from predefined problem-solvers to research agents that address frontier challenges with formal reasoning.

  Their words

  We argue that the next leap in AI4Math systems requires a decisive shift from predefined problem-solvers to research agents that can address frontier mathematical challenges with rigorous formal mathematical reasoning.

  arxiv.org

AI research engineer working on large language models. He writes the Ahead of AI newsletter and is the author of Build a Large Language Model (From Scratch).

• Our readingkorrents.com

  As reasoning models and agents retain more tokens for longer, KV-cache size and attention cost become the main constraint on scaling context.

  LLMs

  Their words

  As reasoning models and agent workflows keep more tokens around (for longer), KV-cache size, memory traffic, and attention cost quickly become the main constraints, and LLM developers are adding a growing number of architecture tricks to reduce those costs.

  magazine.sebastianraschka.com

Researcher on distributed systems and local-first software at the University of Cambridge; wrote Designing Data-Intensive Applications.

• Our readingkorrents.com

  People in industry routinely short-circuit their reasoning, adopting what they heard in a conference talk instead of thinking it through from first principles.

  Their words

  Often people in industry I feel like sort of have short circuit reasoning. Like don't maybe don't quite reason something through from first principles but just like oh I heard this from a conference talk I'm just going to go with that

  youtube.com

Founder and chief analyst of SemiAnalysis; reports on semiconductor supply chains, AI datacentre economics and what the chip export controls actually do.

• Our readingkorrents.com

  As labs sell finished white-collar work rather than reasoning traces, Chinese models will lose the ability to distil their way to parity.

  AI and jobsChinaAmerica

  Their words

  As we move from, you know, hey, [clears throat] these companies are selling tokens where they provide the entire uh reasoning chain and all that to uh selling automated, you know, white collar work, right? Automated software engineer, send them the request, they give you the result back and there's a bunch of thinking on the back end that they don't show you. The ability to distill out of American models into Chinese models will be harder.

  youtube.com

Founding member of OpenAI and former director of AI at Tesla; creator of nanoGPT and the term "vibe coding".

• Likedrcmnd.app

  Claude Code

  Their words

  Claude Code (CC) emerged as the first convincing demonstration of what an LLM Agent looks like - something that in a loopy way strings together tool use and reasoning for extended problem solving.

  karpathy.bearblog.dev

Member of technical staff at Anthropic. He has led ML/AI teams at Amazon, Alibaba and Lazada, and writes about LLMs, recommender systems and engineering at eugeneyan.com.

• Our readingkorrents.com

  Traditional recommender systems predict user behavior well but cannot be steered via natural language or explain their own reasoning.

  Their words

  They excel at predicting what a user will click or buy next, but can't be steered via natural language or reason on their choices.

  eugeneyan.com

Mathematician at UCLA, working mainly in harmonic analysis and partial differential equations. Writes What's new, a long-running blog on their research, open problems and expository notes.

• Our readingkorrents.com

  The stable one-profession career is becoming a thing of the past; what will still be needed alongside AI is reasoning with abstractions and problem-solving when things go wrong.

  Their words

  It's becoming much more a thing of the past. So I think you just have to be adaptable and flexible. I think people will have to get skills that are transferable, like learning one specific programing language or one specific subject of mathematics or something. That itself is not a super transferable skill, but sort of knowing how to reason with abstract concepts or how to problem solve when things go wrong.

  youtube.com

Senior economist at the Foundation for American Innovation; writes Second Best, and was the Niskanen Center's director of social policy before that.

• Our readingkorrents.com

  Reasoning models trained on verifiable rewards automatically gain greater autonomy and long-horizon planning ability as a side effect.

  Their words

  Consistent with the Kantian hypothesis, reasoning models automatically gain greater autonomy, self-consistency and long-horizon planning ability for free.

  Do LLMs really reason?secondbest.ca

Psychiatrist and clinical assistant professor at Case Western Reserve University. Writes Psychiatry at the Margins on the philosophical and scientific arguments inside psychiatry and the psy-sciences.

• Readrcmnd.app

  How Does the Psychiatrist Know?: On the Epistemology of Psychiatric Diagnostic Reasoning

  Their words

  Recently I reviewed Adrian Kind’s 2025 book, “How Does the Psychiatrist Know?: On the Epistemology of Psychiatric Diagnostic Reasoning,” on this substack. I had mentioned that Kind’s book is the first systematic, in-depth, book-length philosophical investigation into how psychiatrists arrive at their diagnostic conclusions.

  psychiatrymargins.com

Machine-learning researcher; has written the Lil'Log survey posts on how a model technique works since 2017, and worked at OpenAI from 2018 to 2024, latterly leading its safety systems team.

• Our readingkorrents.com

  Chain-of-thought explanations from language models can be biased because training does not explicitly reward faithful reasoning.

  Their words

  model CoTs could be biased due to lack of explicit training objectives aimed at encouraging faithful reasoning

  Why We Thinklilianweng.github.io

Machine-learning researcher on open language models; writes the Interconnects newsletter and the RLHF Book, after leading post-training at Ai2.

From one piece DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459 2 beliefs, in the piece's order there

• Our readingkorrents.com

  Reasoning is not taught to a model by people: it emerges from reinforcement learning on questions with checkable answers, with no human preference data at all.

  Their words

  And these reasoning behaviors emerge naturally. So these things like, "Wait, let me see. Wait, let me check this. Oh, that might be a mistake." And they emerge from only having questions and answers.

  youtube.com

• Our readingkorrents.com

  Reasoning will generalize the way instruction tuning did: add enough verifiable domains and, at some point nobody can yet locate, the rest start working on their own.

  Their words

  And the history of NLP and language processing instruction, tuning and tasks per language model used to be like one language model did one task, and then in the instruction tuning literature, there's this point where you start adding more and more tasks together where it just starts to generalize to every task. And we don't know where on this curve we are.

  youtube.com

Founder and chief analyst of SemiAnalysis; reports on semiconductor supply chains, AI datacentre economics and what the chip export controls actually do.

From one piece DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters | Lex Fridman Podcast #459 2 beliefs, in the piece's order there

• Our readingkorrents.com

  The aha moment for reasoning models will come from computer use and robotics rather than from scientific discovery, because those are the infinitely verifiable playgrounds.

  Their words

  I think it's actually probably simpler than that. It's probably something related to computer use or robotics rather than science discovery.

  youtube.com

• Our readingkorrents.com

  American export controls police floating-point operations while memory and interconnect matter just as much, and reasoning models make that the wrong axis to control.

  Americagovernment

  Their words

  FLOP is the vector that the government has cared about historically, but the other two vectors are arguably just as important. And especially when we come to this new paradigm, which the world is only just learning about over the last six months: reasoning.

  youtube.com

Professor at Wharton studying how people actually use AI, and author of Co-Intelligence. Writes One Useful Thing, where the claims are dated and testable because the thing they describe keeps changing under them.

• Likedrcmnd.app

  DeepSeek

  Their words

  if you want a very good all-around model with excellent reasoning. As an open model, you can either use it hosted on the original Chinese DeepSeek site or from a number of local providers.

  oneusefulthing.org

Cognitive scientist and long-standing critic of deep learning's claims; writes Marcus on AI and wrote Rebooting AI.

• Our readingkorrents.com

  Large language models do not reason formally: their performance collapses as a problem is made bigger, in the way a calculator's never does.

  LLMs

  Their words

  Another manifestation of the lack of sufficiently abstract, formal reasoning in LLMs is the way in which performance often fall apart as problems are made bigger.

  garymarcus.substack.com

Software engineer and a long-time technical lead of the Go programming language; author of plan9port and of the research.swtch.com essays.

• Our readingkorrents.com

  Debugging by reasoning backward from a bad state is inherently harder than reasoning forward from a known decision point.

  Their words

  One of the hardest things about debugging is running a program backward: there's a data structure with a bad value, or the control flow has zigged instead of zagged, and it's very difficult to understand how it could have gotten into that state.

  research.swtch.com

Co-founder and CEO of Perplexity, an AI answer engine; previously a research scientist at OpenAI and a PhD student at UC Berkeley.

From one piece Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet | Lex Fridman Podcast #434 2 beliefs, in the piece's order there

• Our readingkorrents.com

  If reasoning can be separated from facts, the frontier labs lose their advantage, because nobody needs a million-GPU cluster any more.

  Their words

  But any breakthrough can disrupt that. If you can decouple reasoning and facts and end up with much smaller models that can reason really well, you don’t need a million H100 equivalent cluster.

  youtube.com

• Our readingkorrents.com

  Real reasoning will have arrived when spending more compute at inference time reliably buys a dramatically better answer.

  Their words

  I think if we can achieve that amount of inference compute, where it leads to a dramatically better answer as you apply more inference compute, I think that will be the beginning of real reasoning breakthroughs.

  youtube.com

Journalist and author writing on politics, liberalism and ideas; a contributor to Quillette and co-author of a book on Christopher Hitchens.

• Our readingkorrents.com

  A movement that has decided its job is to save humanity is uniquely prone to hubris, dogma and motivated reasoning.

  Their words

  This is a worldview that’s uniquely susceptible to hubris, dogma, and motivated reasoning.

  How Effective Altruism Lost Its Wayquillette.com

Research scientist at Meta in Berlin working on multilingual models and evaluation; led the multilingual team at Cohere and was a research scientist at Google DeepMind before that.

• Our readingkorrents.com

  Current large language models still struggle to compose operations into correct multi-step reasoning paths.

  LLMs

  Their words

  Overall, current LLMs still struggle with composing operations into correct reasoning paths.

  NeurIPS 2023 Primerruder.io

Essayist; writes Astral Codex Ten, previously Slate Star Codex.

• Our readingkorrents.com

  The people who can be moved by deaths from neglected tropical diseases are the same people who can be moved by pandemics and by AI risk, because both come from reasoning about importance rather than from the news.

  AI alignment

  Their words

  And I notice that the tiny handful of people capable of caring about 200,000 people dying of neglected tropical diseases are the same tiny handful of people capable of caring about the next pandemic, or superintelligence, or human extinction.

  In Continued Defense Of Effective Altruismastralcodexten.com

Engineer and writer on machine-learning systems; author of Designing Machine Learning Systems and AI Engineering. "I work to bring AI into production. I write about AI system design."

• Our readingkorrents.com

  Text is a more powerful output modality for a model than images, because text supports many downstream tasks

  Their words

  Text is a much more powerful mode for model outputs. A model that can generate images can only be used for image generation, whereas a model that can generate text can be used for many tasks: summarization, translation, reasoning, question answering, etc.

  Multimodality and Large Multimodal Models (LMMs)huyenchip.com

Economist at George Mason University, co-author of Marginal Revolution, host of Conversations with Tyler. Writes in short, dense claims about economics, culture and AI, and has moved position on AI in public more than once.

• Our readingkorrents.com

  The General Theory is a brilliant work despite numerous flaws in reasoning, exposition, and microeconomics.

  Their words

  My current view is that the General Theory is a brilliant work, marred by numerous mistakes in reasoning, exposition, and microeconomics.

  tylercowen.com

Founder and president of the Roots of Progress Institute; writes the history and philosophy of progress, after a career as a software engineering manager and founder.

• Our readingkorrents.com

  There is no capital-S Science that can stand as an authority; there is only reasoning done systematically.

  Their words

  There is no capital-S Science—a new authority to substitute for God or King.

  What does it mean to “trust science”?rootsofprogress.org

Founded CD Baby in 1998 and sold it in 2008, giving the proceeds to a charitable trust for music education. Since then a writer and speaker: five books and 550+ articles, published first at sive.rs, which they hand-code and keep free of advertising.

• Likedrcmnd.app

  The Righteous Mind

  Their words

  Brilliant insights into human nature. Moral reasoning, inventing victims, emotional reactions, the elephant (emotions) and its rider (explanations). Intuition reacts then logic confabulates a reason. Survival of the fittest applied to groups. Understanding religion and tradition. Great from start to finish.

  sive.rs

Creator of the Keras deep-learning library and the ARC-AGI benchmark.

• Our readingkorrents.com

  The Abstraction and Reasoning Corpus can measure human-like general fluid intelligence and enable fair comparisons between AI systems and humans.

  Their words

  We argue that ARC can be used to measure a human-like form of general fluid intelligence and that it enables fair general intelligence comparisons between AI systems and humans.

  On the Measure of Intelligencearxiv.org