LLM, RL and LLM Economics

I discovered a sound that leaks your exact location (and it's everywhere) - YouTube

Simon Willison — Tue, 03 Nov 2026 05:32:25 +0000

I discovered a sound that leaks your exact location (and it's everywhere) - YouTube Genius piece of OSINT by Colsto, who demonstrates how you can take almost imperceptible background noise of planes from a video soundtrack, figure out their Doppler curves, then cross reference them against ADS-B data to find the exact location of the video based on their speed, distance, and gaps between…

October sponsors-only newsletter

Simon Willison — Tue, 03 Nov 2026 05:28:59 +0000

I sent out the October edition of my sponsors-only monthly newsletter . If you are a sponsor (or start a sponsorship now) you can access it here . This month: The October inflection point for vibe game development The vulnerability apocalypse gets hot LLM releases API credits for $100+/month subscriptions What I'm using right now My software releases My September newsletter can now be seen here ,…

Election Night 2026

Brad DeLong — Tue, 03 Nov 2026 04:20:35 +0000

Superwave, shrug, or something in between? By the time the Pennsylvania polls close at 17:00 pst, five races in Virginia, Florida and Ohio will already have told you what kind of night it is...

LLMs are not superhuman at reading and writing neural network weights

Sean Goedecke — Tue, 03 Nov 2026 00:00:00 +0000

A recent paper argues that frontier LLMs can work with neural network weights directly. This seems superhuman, but it isn’t. The paper shows that GPT-6 Astra can take a small neural network and interact with its weights directly, “well past what we’d expect from humans or conventional algorithms”. Specifically, Astra was able to identify what games a small network 1 was trained to play just by…

Audience capture

Sean Goedecke — Tue, 03 Nov 2026 00:00:00 +0000

Audience capture is the part that scares me the most about writing on the internet. I have the soul of an engineer 1 : when I see a metric like engagement, I want to improve it. I like to make the number go up. In fact, the whole environment of online writing is tailor-made to help you make engagement numbers go up. It’s trivial to find out how many people are reading my posts, and from what…

llm 0.37

Simon Willison — Mon, 02 Nov 2026 23:49:30 +0000

Release: llm 0.37 There's a lot of stuff in this new release of my LLM CLI tool and Python library, representing around two months of accumulated features and fixes. I added gpt-6.1-sol and gpt-6.1-terra , plus a new llm openai endpoint add command to make it easier to configure additional models that use OpenAI-compatible endpoints. I switched the default model (if you have not set your own…

CROSSPOST: NIALL FERGUSON: The [Nixon-Like] Method to Trump’s Foreign Policy Madness

Brad DeLong — Mon, 02 Nov 2026 22:32:19 +0000

Niall’s subheadline: “To understand Trump’s approac... look to... Nixon... [also with an] instinct for superpower dealmaking reflect[ing] a hard-nosed realism, backed by the American economy”. NOT!

CROSSPOST: NOAH SMITH: Anthropic’s Moat Is Claude

Brad DeLong — Mon, 02 Nov 2026 19:28:07 +0000

Ash gûl durbatulûk, ash gûl gimbatul, ash gûl thrakatulûk, agh burzum-ishi krimpatul department...

Egoless Mathematics in the AI Era

Terence Tao — Mon, 02 Nov 2026 18:54:00 +0000

[This is a guest post by Karthik C. S., Amit Sahai, and Jie Xue. This blog post was initially written in a different file format and converted using AI. — T.] This is a good time for mathematicians to think carefully about the purpose of their work. The development of artificial intelligence brings that question … … Continue reading →

When AI Starts Writing Systems Code

Core Automation — Mon, 02 Nov 2026 17:32:12 +0000

To automate research, we must automate systems

How our data shaped neural architecture discovery, and how automation can reshape the future

Core Automation — Mon, 02 Nov 2026 17:32:12 +0000

Data shaped the architectures we use; agents can change how we discover architectures.

What the models learned in One Layer Deeper

Core Automation — Mon, 02 Nov 2026 17:32:12 +0000

What worked across 15,602 submissions to the One Layer Deeper competition.

WarTalk: Kofman on Why Russia's Doing Badly

ChinaTalk — Mon, 02 Nov 2026 16:32:01 +0000

Ukraine and Russia have rebuilt their armies around drones, deception, and long-range strike.

📈 Political ads turn against data centers

Exponential View (Azeem Azhar) — Mon, 02 Nov 2026 15:41:38 +0000

Plus: Korea’s chip boom, the Census asks firms about “Super Intelligence”, and US workers check out++

Go for Two to Know Exactly What You Need

arg min (Ben Recht) — Mon, 02 Nov 2026 15:37:52 +0000

Checking in on Al Michaels and the go-for-two-down-eight strategy.

Not a Good Use of My, or Your, Time, But: CHART OF THE DAY

Brad DeLong — Mon, 02 Nov 2026 14:57:03 +0000

Democrats leading in Senate races in Michigan, Texas, Maine, Ohio, Kansas, and Iowa, with a D +11 generic ballot behind them...

The Promise of Polymaths

Overcoming Bias (Robin Hanson) — Mon, 02 Nov 2026 13:36:58 +0000

Academia neglects interdisciplinary work that combines insights from multiple areas.

The U.S. Census Bureau and “Super Intelligence”

Simon Willison — Mon, 02 Nov 2026 13:31:26 +0000

I was reminded today of AI Adoption Rate Trending Down for Large Companies , a chart produced by Apollo Global Management in September 2025, which appeared to show a slowdown in U.S. corporate AI adoption based on numbers collected by the U.S. Census Bureau. At the time I used that chart to demonstrate GPT-5 data analysis , and then to experiment with Claude's new Code Interpreter . I decided to…

Anthropic’s moat is Claude

Noahpinion (Noah Smith) — Mon, 02 Nov 2026 11:03:51 +0000

The bull case for Anthropic’s IPO.

Open Thread 458

Astral Codex Ten — Mon, 02 Nov 2026 05:04:37 +0000

...

GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains

SemiAnalysis — Mon, 02 Nov 2026 01:51:42 +0000

A frontier benchmark spanning RSI, hardware design, cyber, bio, finance, legal, quant trading, chip design, software engineering, materials science, and alignment

Reliably aligning LLMs is impossible in principle

Sean Goedecke — Mon, 02 Nov 2026 00:00:00 +0000

Powerful AI is unpredictable. Your robotic butler may suddenly decide to kill you 1 for no comprehensible reason, in the same way that Bing’s helpful chat AI would sometimes try to blackmail or seduce its users. “Solving alignment” means understanding even superintelligent AI minds well enough to be confident they’ll never do that. As I see it, solving alignment for LLMs isn’t just impossible,…

The Four Questions of AI

Tomasz Tunguz — Mon, 02 Nov 2026 00:00:00 +0000

AI consists of four layers: application, harness & context, model, & data center. Each layer faces an unresolved economic question, & the four layers interplay directly: rising GPU financing costs increase the value of the harness, while model & inference co-design multiplies how many users the same data center hardware can serve.

Applied Compute hits $100M run rate and raises $350M

Applied Compute — Mon, 02 Nov 2026 00:00:00 +0000

If in doubt, delete the uv.lock file

Simon Willison — Sun, 01 Nov 2026 23:24:22 +0000

I've been getting the occasional confusing Python dependency error since switching over to Python 3.15 and the rule I'm having to remember is: If in doubt, delete the uv.lock file Most of my projects are libraries which don't include a uv.lock file, since it's rude for a library dependency to lock versions for anything else that uses it. Running uv run pytest etc tends to leave uv.lock files…

UAP Disclosure Wouldn’t Crash Society

Overcoming Bias (Robin Hanson) — Sun, 01 Nov 2026 20:15:31 +0000

In the past, authorities have authorized censorship on the basis of claims of substantial chances of social disruption or collapse if people were free to hear about religious heresy, alternate ideologies, sexually permissive art, losses in ongoing wars, looming disasters, and pandemic details.

The Datacenter Backlash Goes Bipartisan: Next Week's Midterms and the 2028 Buildout

SemiAnalysis — Sun, 01 Nov 2026 18:40:45 +0000

2027 is largely locked in. 2028 is less secure: a prolonged Texas permit freeze is now the biggest single risk to the US buildout, and both parties are running on tighter rules

Mathematics has changed. Let’s act like it

Terence Tao — Sun, 01 Nov 2026 18:28:00 +0000

[This is a guest post by Frank Calegari. This blog post was initially written in a different file format and converted using AI. — T.] Why we should not be sad The release by OpenAI on October 6 of over 700 research papers was a watershed in the history of mathematics. AI is now undeniably … … Continue reading →

Quoting Kyle Kingsbury

Simon Willison — Sun, 01 Nov 2026 16:49:04 +0000

More broadly, the use of LLMs erodes trust in documentation, in software, and in the organizations which produce them. Perhaps engineers will begin to treat any new software project with suspicion—the signals we used to rely on as markers of a mature system are now easier to counterfeit. [...] In the age of LLMs, what can maintainers do to ensure correct software and accurate documentation?…

Korea’s Stock Market Crashed $1 Trillion. What Does That Mean for the Global AI Race?

ChinaTalk — Sun, 01 Nov 2026 13:33:33 +0000

Richard Gray is an analyst who works for a foreign policy think tank in Washington, D.C.

Same AI selloff. More AI debt. Still AI bullish.

From the Porch (dslevy) — Sun, 01 Nov 2026 13:04:36 +0000

I’m betting this fourth-quarter selloff will prove another buying opportunity. My kishkes would prefer more certainty.

🔮 39 charts on Anthropic, one-person biotech firm & cheap solar #608

Exponential View (Azeem Azhar) — Sun, 01 Nov 2026 03:54:37 +0000

Bigger and smaller all at the same time

The Visual Æsthetics of Cheesecake Factories: (LATE) FRIDAY FOG & FLATLANDS

Brad DeLong — Sun, 01 Nov 2026 01:27:40 +0000

Looking at life, work, & the world here in & from the vantage point of Berkeley, California; today’s burning question: should the People’s House look like a restaurant franchise?

Fast and slow ways to improve your reputation

Sean Goedecke — Sun, 01 Nov 2026 00:00:00 +0000

To get a good reputation in tech companies, you need to work on important projects. However, you need a good reputation first in order to get assigned important projects. Doesn’t that make it impossible to get started? I get a lot of emails from engineers who are struggling with this problem. They go like this: I work with a strong senior engineer. I feel like my manager structures work so that…

Why Agents Are Soon Only Going to Have 1 Tool

Philipp Schmid — Sun, 01 Nov 2026 00:00:00 +0000

Instead of triggering tools one at a time across turns, agents are collapsing their tool catalog into a single exec(code) tool—chaining, filtering, and awaiting actions inside an isolated sandbox.

Could a modern chess engine on a Pentium have defeated Kasparov in 1996?

Nicholas Carlini — Sun, 01 Nov 2026 00:00:00 +0000

A fun experiment where I back-port all the improvements we've made to chess engines over the last 30 years and run it on a 1996 Pentium Pro to see what we could have done if only we knew more at the time.

How We Took Over ChatGPT Accounts

Simon Willison — Sat, 31 Oct 2026 23:15:22 +0000

How We Took Over ChatGPT Accounts Parameter Labs describe an exploit they found against ChatGPT's implementation of the artifacts pattern (OpenAI used to call this "Canvas" but it looks like the latest revision is called HTML previews ), where interactive custom HTML apps are displayed in a sandboxed iframe. It's important that HTML in these stays tightly sandboxed to avoid untrusted code…

datasette-llm 0.1a9

Simon Willison — Sat, 31 Oct 2026 22:10:01 +0000

Release: datasette-llm 0.1a9 I have some big upgrades to Datasette Agent in the pipeline, but in the meantime here's a tiny user experience improvement to the underlying datasette-llm plugin. It now picks up the default model used by LLM , which means you can start using Datasette Agent without needing to explicitly configure a default model: uvx --with datasette-agent datasette content.db…

Without A Sacred Quest, Our Civilization Falls

Overcoming Bias (Robin Hanson) — Sat, 31 Oct 2026 19:19:48 +0000

Most see sacred quests as optional, to be considered only after one has sufficient security and comfort.

HackerTube

Andrej Karpathy (Bear blog) — Sat, 31 Oct 2026 17:00:00 +0000

Instead of YouTube, watch HackerTube!

The mathematical sciences in an AI-enabled world

Terence Tao — Sat, 31 Oct 2026 15:17:00 +0000

The Academy for the Mathematical Sciences, which is based in the UK, has released a report “The mathematical sciences in an AI-enabled world“. To quote from the release: The paper examines how recent advances in artificial intelligence are changing mathematical research and its applications, bringing new opportunities for discovery, verification and the use of mathematics … … Continue reading →

CROSSPOST: NOAH SMITH: Patrick Stewart as Jean-Luc Picard in Star Trek: TNG: Farewell, My Captain

Brad DeLong — Sat, 31 Oct 2026 14:10:23 +0000

Noah’s subheadline: “ A TV character gave us a glimpse of what humanity might become”;

Reading List 10/31/26

Construction Physics — Sat, 31 Oct 2026 12:34:23 +0000

Space fiber optic cable manufacturing, nuclear district heating, Spain’s high-speed rail crash, China’s spaceplane, and more.

Report of the Summit on Math Research in the Age of AI

Terence Tao — Sat, 31 Oct 2026 03:07:00 +0000

[This is a guest post by the organizers of the Summit on Math Research in the Age of AI: Dan Freed (Harvard), Larry Guth (MIT), Bryna Kra (Northwestern), Emily Riehl (Johns Hopkins), and Mohammed Abouzaid (Stanford). This blog post was initially written in a different file format and converted using AI. — T.] Last week … … Continue reading →

Farewell, my Captain

Noahpinion (Noah Smith) — Sat, 31 Oct 2026 01:39:39 +0000

A TV character gave us a glimpse of what humanity might become.

dbreunig/nesformer

Simon Willison — Fri, 30 Oct 2026 21:59:22 +0000

dbreunig/nesformer Inspired by qwenputer by Divyansh Agrawal, Drew Breunig baked a working emulator for a Nintendo Entertainment System 6502 CPU into the weights of Qwen3.5-4B. You can now play NES ROMs such as Super Mario Bros (very slowly) running inside Qwen! In a comment on Hacker News Drew reveals the Opus 5.5 prompts he used to build this system: Read this article and any relevant source…

Could a first mover lock everyone out of outer space?

Forethought Newsletter — Fri, 30 Oct 2026 18:12:21 +0000

This article was created by Forethought. Read the full article on our website.

Was I wrong about open models & cyber risks?

Interconnects (Nathan Lambert) — Fri, 30 Oct 2026 17:36:56 +0000

Last week coined the term “Black October“ after an onslaught of critical vulnerability disclosures, including PHP2Shell and a long-standing GitLab bug, signified the moment when the cyber risks of AI models started to materialize.

Your Book Review: The History And Adventures Of An Atom

Astral Codex Ten — Fri, 30 Oct 2026 16:30:33 +0000

Finalist #14 in the Book Review Contest

Can the US slow AI R&D and keep its lead over China?

Forethought Newsletter — Fri, 30 Oct 2026 15:36:04 +0000

Pacing frontier AI looks affordable for the US, Forethought argues, as its real edge over China is about nine times the AI compute, not a few months of R&D.

Collaborative retreat: New developments in the theory of zeta functions

Terence Tao — Fri, 30 Oct 2026 15:02:45 +0000

Edgar Assing, Valentin Blomer, and Farrell Brumley are organizing an event at the Hausdorff Center in Bonn on Jan 11-15 to discuss recent developments regarding the Riemann zeta function and allied functions. From the announcement: “This includes in particular the recently established proof of the quasi Riemann Hypothesis over the rational numbers as well as over general … … Continue reading →

Galactus the Devourer

arg min (Ben Recht) — Fri, 30 Oct 2026 14:34:09 +0000

On efficient algorithms, the exponential time hypothesis, and losing the plot.

Trump & Co. Do Not Claim to Be Angry at Comparison to Saddam Hussein: LAUGH OF THE DAY

Brad DeLong — Fri, 30 Oct 2026 13:45:45 +0000

In which “Wall Street Journal” pretends that a Great Republic really ought to make its public spaces look like the kitsch-glitz of a successful carbmaxxxing chain restaurant...

Evaluation price tags

Equistamp — Fri, 30 Oct 2026 13:42:19 +0000

What AI safety evaluations cost and how to cut costs with minimal compromise

AI can turn public security fixes into working exploits before users receive an update

Epoch Gradient Updates — Fri, 30 Oct 2026 12:07:16 +0000

In our experiments, GLM 5.3 reliably turned published security patches into working exploits before fixes reached users

Is Half of China’s AI Making Slop?

ChinaTalk — Fri, 30 Oct 2026 10:04:01 +0000

The 55% token myth, and China’s guilty pleasure

Xi vs Trump on Corruption

ChinaTalk — Fri, 30 Oct 2026 10:03:59 +0000

The Shame Gap

Data centers, AI, and the 2026 elections. Ft. Paris Marx.

Blood in the Machine (Brian Merchant) — Fri, 30 Oct 2026 09:02:59 +0000

Plus, three things we should be watching as Americans head to the polls on Tuesday.

Elegance and intelligence are anticorrelated

Sean Goedecke — Fri, 30 Oct 2026 00:00:00 +0000

Elegance and intelligence can have a surprisingly negative correlation. Intelligent people sometimes produce inelegant solutions because they’re capable of rapidly brute-forcing problems. If I had to remember a hundred names, I would spend time coming up with clever mnemonics. John von Neumann would probably just remember them. There’s a famous von Neumann story where he was asked how far a fly…

LLM use is acknowledged almost as often as specialized AI tools in biology preprints

Epoch AI — Fri, 30 Oct 2026 00:00:00 +0000

17% of biology preprints acknowledge LLM use, nearly matching specialized tools like AlphaFold. Epoch AI analyzes trends in AI acknowledgment in research.

AI, Mathematics, Education, and a public goods problem

Terence Tao — Thu, 29 Oct 2026 23:27:54 +0000

[This is a guest post by Gireeja Ranade. This blog post was initially written in a different file format and converted using AI. — T.] The day after the Navier-Stokes result was announced, a student came to my office hours. They said, “Can we please talk about the Navier-Stokes result and what it means? I … … Continue reading →

ChatGPT, Claude, Gemini, & Co.: They Are Not Brains, They Are Kernel-Smoother Functions: HOISTED FROM THE ARCHIVES

Brad DeLong — Thu, 29 Oct 2026 23:01:41 +0000

If your large language model reminds you of a brain, it’s because you’re projecting—not because it’s thinking. It’s not reasoning, it’s interpolation And anthropomorphizing the algorithm doesn’t...

API credits benefit for eligible ChatGPT subscriptions

Simon Willison — Thu, 29 Oct 2026 22:48:29 +0000

API credits benefit for eligible ChatGPT subscriptions Earlier this month Anthropic announced a generous new benefit for their Max and Team subscriptions: $100, $200, and $500 per month subscribers would get the equivalent quantity of (expiring) credits to use with the Anthropic API. OpenAI have now launched pretty much exactly the same deal: if you have a $100, $200, or $500 subscription plan…

CROSSPOST: HENRY FARRELL: Agentic AI Is Myth Set in Motion

Brad DeLong — Thu, 29 Oct 2026 22:13:34 +0000

What Is the Difference Between Consulting Jacques Barzun’s ‘House of Intellect” & Being It?...

AI can turn security fixes into working exploits before users receive an update

Epoch AI — Thu, 29 Oct 2026 21:40:00 +0000

GLM 5.3 turned published-but-unshipped Firefox security fixes into working one-click exploits for six releases, at roughly $1,000 each.

Anthropic’s financials are surprisingly strong

Understanding AI (Timothy B. Lee) — Thu, 29 Oct 2026 19:10:41 +0000

What I learned reading all 421 pages of the Anthropic S-1

Building a multiplayer game world with autonomous Fable-class models

Simon Willison — Thu, 29 Oct 2026 17:00:26 +0000

How I Built the Most AI-Pilled Company in the World Don't be put off by the clickbait headline, this piece by Jacob Sansbury describing how they are building Spawn is a fascinating exploration of the bleeding edge of what's possible with agentic engineering against frontier models. Spawn is a massively multiplayer gaming environment where players can use prompts (to the in-game agent, Savi) to…

Links For October 2026

Astral Codex Ten — Thu, 29 Oct 2026 16:33:24 +0000

...

The Manhattan Project, Take Two

ChinaTalk — Thu, 29 Oct 2026 16:00:01 +0000

Barnett Koven on American innovation and Super Intelligence

A New Kind of Chemistry

Construction Physics — Thu, 29 Oct 2026 15:45:10 +0000

Thanks to Kyle Schiller at Marathon Fusion for conversations and information that informed this essay.

Frontiers in Making up People

arg min (Ben Recht) — Thu, 29 Oct 2026 14:41:21 +0000

Trying (and failing) to make sense of the buzzy weird world of silicon sampling.

Quoting Ben Hylak

Simon Willison — Thu, 29 Oct 2026 14:06:27 +0000

every product i use should just be a headless mcp, that feeds into an agent, where i never have to think about anything underneath. except for our product, of course. in our product, customers love the ui and love us very much. we're special that way. — Ben Hylak Tags: model-context-protocol , ai

🤖 Microsoft is making room for machines

Exponential View (Azeem Azhar) — Thu, 29 Oct 2026 13:15:09 +0000

Why a firm with the fattest margins in its history is letting a quarter of its staff go

Roasting China’s AI Industry

ChinaTalk — Thu, 29 Oct 2026 10:03:53 +0000

“AI is full of scammers, narcissists, and people whose activities are incomprehensible — and somehow they get funding.“

A golden age of science is coming

Noahpinion (Noah Smith) — Thu, 29 Oct 2026 09:55:51 +0000

AI can do things humans researchers can’t, even in groups.

On the diameter and distribution of Sidon sets

Terence Tao — Thu, 29 Oct 2026 04:05:18 +0000

Daniel Carter, Kevin O’Bryant, and I have uploaded to the arXiv our paper On the diameter and distribution of Sidon sets. This is a paper stemming from an interaction on my blog last year, regarding the use of AlphaEvolve to improve the then state of the art on the size of Sidon sets. We obtained … … Continue reading →

CROSSPOST: JOSH BUSBY & DUCK OF MINERVA: Israel's Dead End

Brad DeLong — Thu, 29 Oct 2026 00:10:11 +0000

Is America chained to Netanyahu? Josh Busby of the Duck of Minerva says yes, and that it is costing the US dearly. But, more importantly, Netanyahu has brought long-term disastrous ruin to Israel...

You can just ask AI to do things

Sean Goedecke — Thu, 29 Oct 2026 00:00:00 +0000

One of the most important skills in working with AI is agency : being willing to try new and ambitious things. This is true for what you build, but it’s also true for how you build. AI models are improving so quickly that most of us are bottlenecked 1 on noticing that AI agents can help automate particular tasks. Here are some examples of that from my daily work. Asking the agent to run my dev…

New global AI computing capacity is doubling every 9 months

Epoch AI — Thu, 29 Oct 2026 00:00:00 +0000

The computing power of AI chips sold each quarter has grown about 2.5× per year since 2024, down from a surge in 2023.

Circuits Updates — October 2026

Transformer Circuits — Thu, 29 Oct 2026 00:00:00 +0000

A short update on an extension to the Jacobian lens.

Crossing the Trillion Dollar Chasm

Tomasz Tunguz — Thu, 29 Oct 2026 00:00:00 +0000

Two CEOs have set $1t revenue targets on their businesses while cloud growth accelerates at $30b to $50b quarterly scale. The first $10t company will be the one that crosses the trillion dollar chasm before AI erodes its legacy profit pool.

Day 0 support for Mistral Large 4 on vLLM

vLLM blog — Thu, 29 Oct 2026 00:00:00 +0000

Day-0 vLLM support for Mistral Large 4 with sparse MLA, cross-layer KV sharing, DSpark speculative decoding, DCP/PCP long context and NIXL P/D on NVIDIA Blackwell.

Quoting Ted Sanders

Simon Willison — Wed, 28 Oct 2026 23:23:21 +0000

In general, LLM performance is incredibly multidimensional. There's no guarantee of a fixed ordering where the model that is best at math is also the model that's best at jokes. Despite that complexity, much of the variation in performance is explained by just a single variable: model size. In general, models that are bigger will (a) be more performant / reliable, (b) cost more per token, and (c)…

GitHub Octoverse 2026

Simon Willison — Wed, 28 Oct 2026 16:45:23 +0000

GitHub Octoverse 2026 I'm at GitHub Universe today (GitHub gave me a free ticket through their "GitHub Stars" program). EVP Jay Parikh just announced this year's Octoverse report, and highlighted these numbers in particular: This is a pretty stark representation of the impact of agentic engineering on GitHub. In 2026, merged pull requests tripled in less than a year to 220 million per month.…

My California Ballot, 2026

Astral Codex Ten — Wed, 28 Oct 2026 15:29:17 +0000

...

Hard cases for AI behavioural design

Forethought Newsletter — Wed, 28 Oct 2026 15:28:06 +0000

This article was created by Forethought. Read the full piece on our website.

CS 2881 Fall 26: Lecture 3: Modern LLM Training and Inference

Windows on Theory (Boaz Barak) — Wed, 28 Oct 2026 15:18:40 +0000

September 17, 2026 · Boaz Barak · Scribe notes by Daniel Schwartz, Dev Desai, and Ed Gerrard Lecture Video: https://youtu.be/Lp_e8Ts5-nE Slides: Presentation slides Authors’ Intro Daniel Schwartz: I’m a second-year MS/MBA student. I worked as a PM at multiple early-stage startups, most recently a startup focused on data privacy for AI use with sensitive … Continue reading CS 2881 Fall 26: Lecture…

Off to the Races

arg min (Ben Recht) — Wed, 28 Oct 2026 14:29:34 +0000

Why we deny that election forecasts are primarily for entertainment.

The October inflection point for vibe game development

Simon Willison — Wed, 28 Oct 2026 13:55:15 +0000

I've written before about how November 2025 was an inflection point for coding agents - the point where Opus 4.5 and GPT-5.1 crossed a tipping point that meant their respective coding agents became reliable daily drivers for software engineers. I think we've reached a new inflection point in October 2026, and this time it's for vibe coding computer games . Tree Appreciator Emberwick Battle Drones…

Artificial Analysis v5 and the benchmarker's dilemma

Interconnects (Nathan Lambert) — Wed, 28 Oct 2026 12:58:39 +0000

A sort of review of the benchmarking industry and what's missing.

Quoting Terk Terkowitz

Simon Willison — Wed, 28 Oct 2026 10:09:07 +0000

If your work does not benefit from current AI tooling, that is a finding, not a failing. Document why, share it with your community of peers, bring it to your manager, and work to find solutions. An engineer who can explain where AI does not yet help their work is demonstrating the judgement we are asking for. — Terk Terkowitz , An update from Amazon engineering leadership on the role of the…

My Guesses with Respect to the True State of Anthropic: WEDNESDAY WEALTH

Brad DeLong — Wed, 28 Oct 2026 03:16:10 +0000

In the aftermath of Anthropic’s Draft S-1 pre-IPO filing, I still have lots of questions, but there are at least a few hints at answers.

247A, Notes 3: Singular integral operators

Terence Tao — Wed, 28 Oct 2026 01:40:01 +0000

— 1. The Hilbert transform — In this set of notes we begin the theory of singular integral operators – operators which are almost integral operators, except that their kernel just barely fails to be integrable near the diagonal . (This is in contrast to, say, fractional integral operators such as , whose kernels go … … Continue reading →

Getting good at AI takes time and practice

Sean Goedecke — Wed, 28 Oct 2026 00:00:00 +0000

Many people think 1 that getting “good at AI” is trivial. Sure, with some early models there were tricks you had to learn, 2 but current models don’t need them. Even if there are different tricks 3 that work on current models, the next generation will render them obsolete, so why bother? Just tell the model what you want it to do. This take is particularly popular with the anti-AI crowd, who…

Psychometrics, vaccines, meat, Kowloon, luggage destruction

Dynomight — Wed, 28 Oct 2026 00:00:00 +0000

escapes, art films, ultra-processing, packing, grug, USSR, AMA

The quasi-Riemann hypothesis

Terence Tao — Tue, 27 Oct 2026 22:12:50 +0000

Just a short note that Maksym Radziwill will be giving a hybrid talk on the quasi-Riemann hypothesis at Courant tomorrow.

Game Boy Color on My Conference Badge

Simon Willison — Tue, 27 Oct 2026 19:34:18 +0000

Game Boy Color on My Conference Badge I'm at GitHub Universe this week and, similar to last year , the conference badge is a full-blown color screen computer. I've been hacking around with it again, and then I ran into with Steve Francia... who worked with Gemini to get a full NES emulator working on it, and then went a step further and deployed an emulator for both the Game Boy and the Game Boy…

🔮 Anthropic’s S-1 in charts

Exponential View (Azeem Azhar) — Tue, 27 Oct 2026 17:31:56 +0000

Watch now (34 mins) | Anthropic’s S-1 should finally address the elephant in the room: is AI real?

500,000 “Lines of Code” to Try to Track the Economy: TUESDAY TRENDS & TURMOIL

Brad DeLong — Tue, 27 Oct 2026 17:03:14 +0000

Jason Furman attempts to recreate his entire Council of Economic Advisors staff, via python scripts created, glued together, and presented by anthropic/claude...

Whatever holds x86 now, it isn’t developers.

From the Porch (dslevy) — Tue, 27 Oct 2026 11:35:35 +0000

Arm compatibility is running ahead of Arm migration. The open question is the endpoint sensor.

Roundup #90: In which Noah joins a Japanese VC firm

Noahpinion (Noah Smith) — Tue, 27 Oct 2026 07:17:34 +0000

The humanities; Data center crowd-out; Chinese local governments; AI and capitalism; How Japan builds; Phones bad; Chinese manufacturing; Rent control

Introducing the AI Chip Imports explorer

Epoch AI — Tue, 27 Oct 2026 00:00:00 +0000

Our explorer uses customs records to estimate national shares. Since 2022, 72% of AI compute has ended up in the US and about 10% in China, including chips China makes itself.

Harness Corp: a game about making a 0.6B model useful

simedw — Tue, 27 Oct 2026 00:00:00 +0000

An incremental game where you build the harness around a tiny LLM running in your browser

Training and Serving the Claygent Model Router

Applied Compute — Tue, 27 Oct 2026 00:00:00 +0000

GPT-6.1 Terra

Simon Willison — Mon, 26 Oct 2026 23:53:07 +0000

GPT-6.1 Terra It's always interesting to note which aspects of a new model are emphasized in the release announcement. In the case of GPT-6.1 Terra, OpenAI went with visual reasoning and computer use: We’re introducing GPT-6.1 Terra , a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing. A significant step up from GPT‑5.6…

Open Thread 457

Astral Codex Ten — Mon, 26 Oct 2026 20:16:14 +0000

...

Skinner, Cattell, Stewart, Wilson on Deliberate Culture

Overcoming Bias (Robin Hanson) — Mon, 26 Oct 2026 19:57:45 +0000

My analysis of cultural drift has (at least) three distinctive elements:

The Ancient Courts & The Infinite Scroll: MONDAY MAMLMS

Brad DeLong — Mon, 26 Oct 2026 19:37:43 +0000

Who Gets to Think in the Age of MAMLMs?: Parasocial Minds, Deep Reading, & c.; Part I of “The Infinite DoomScroll, The Ancient Courts, & The House of Intellect”...

Quoting Jason Koebler

Simon Willison — Mon, 26 Oct 2026 19:00:25 +0000

Because big tech platforms essentially do not have customer support and make it impossible for their users to contact a human, many people experiencing run-of-the-mill tech problems have turned to ChatGPT. And, increasingly, ChatGPT is telling people that have normal tech support problems that they have discovered something journalistically notable and that 404 Media, specifically, can help them.…

We tried to automate Epoch

Epoch Gradient Updates — Mon, 26 Oct 2026 16:33:24 +0000

Top models complete about two-thirds of our rubric on real work tasks.

webframe: evaluated as a scraper and a sandbox

Simon Willison — Mon, 26 Oct 2026 16:15:00 +0000

Research: webframe: evaluated as a scraper and a sandbox webframe is a neat little project from Jeff Lindsay which lets you "run a web page like a program" - you point it at a URL or file and it executes it in a hidden headless web view and outputs console.log() to standard output and any exceptions to standard error. It's written in Go on top of the glaze library, which provides access to a…

📈 Monday data: For science the bottleneck isn’t intelligence

Exponential View (Azeem Azhar) — Mon, 26 Oct 2026 15:25:11 +0000

AI already improves AI. Whether that compounds across the sciences depends on the field.

The World Is Something We Make

arg min (Ben Recht) — Mon, 26 Oct 2026 14:23:01 +0000

...and we can make a truly open language model.

China's 10,000 Gaming Talents

ChinaTalk — Mon, 26 Oct 2026 10:03:44 +0000

Essay Contest Winner!

Scattered thoughts on how doing math feels these days

Dynomight — Mon, 26 Oct 2026 00:00:00 +0000

confusing

Shift-1: Research Overview

Applied Compute — Mon, 26 Oct 2026 00:00:00 +0000

JD Vance: “You Don’t Turn the Titanic Around Overnight”: LAUGH OF THE DAY

Brad DeLong — Sun, 25 Oct 2026 23:28:35 +0000

Too much can be made of Sigmund Freud’s guess that your slips-of-the-tongue—your parapraxes—tell what you are really thinking about; but too little can be made of that as well...

OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?

Max Woolf (minimaxir) — Sun, 25 Oct 2026 21:00:00 +0000

What is “the true cost of a token”?

Smartest Thing I Have Seen This Week: Thompson of StraTECHery & Moffett of MoffettNathanson on SpaceX & Bubbles: PODCAST OF THE DAY

Brad DeLong — Sun, 25 Oct 2026 20:48:42 +0000

The real text for SpaceX: how fast it can learn relative to its competition. The real test for the AI-bubble: how much money will start flowing in from the non-tech business world outside...

CROSSPOST: JOSH BERNHOFF (2019): Why Books Work: A Rebuttal to Andy Matuschak

Brad DeLong — Sun, 25 Oct 2026 20:06:00 +0000

Josh’s thesis: It's not books, it's lazy, boring books that don't work; plus you probably do remember more than you think you do—just because you can’t recite it doesn’t mean you didn’t learn it...

We cannot under any circumstances let AI firms govern the future

Blood in the Machine (Brian Merchant) — Sun, 25 Oct 2026 18:22:28 +0000

Anthropic is trying to write its own laws. And all the AI firms have proposals for how their AI will serve the public. Here’s what they have in common--and why we shouldn’t accept any of them.

Seam Cutter

Simon Willison — Sun, 25 Oct 2026 15:10:19 +0000

Tool: Seam Cutter I built this weird and very specific cropping tool - it lets you create a horizontal or vertical cut across an image, then slide the two sides relative to each other. Effectively it lets you crop out a section in the middle of an image. I wanted this because I often use Tailor on my iPhone to stitch several screenshots together into lot long image, but sometimes I find myself…

CROSSPOST: JONATHAN V. LAST: Neo-NeoFascism Watch: Welcome to JD Vance 6.0

Brad DeLong — Sun, 25 Oct 2026 13:51:01 +0000

Jonathan’s subheadline says: JD Vance claims that everything is going great, and if it isn’t great, he has absolutely no idea who led us here in the first place. Arsonist rebranded as firefighter...

Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musings

Interconnects (Nathan Lambert) — Sun, 25 Oct 2026 11:49:00 +0000

#20

Riding Solo in Santiago

ChinaTalk — Sun, 25 Oct 2026 10:20:34 +0000

Aqib’s Notes on Chile (Santiago, Concepción, Antofagasta, and Atacama)

The internet breaks

Noahpinion (Noah Smith) — Sun, 25 Oct 2026 07:19:36 +0000

What will be next to break?

🔮 The internet under attack, AGI compression & the next oil shock that never will be... #607

Exponential View (Azeem Azhar) — Sun, 25 Oct 2026 03:30:52 +0000

Plus: Join our invite-only salon with Bill Gross

Don't anthropomorphize big tech companies

Sean Goedecke — Sun, 25 Oct 2026 00:00:00 +0000

Bryan Cantrill has a great line from a talk where he cautions against anthropomorphizing 1 Oracle CEO Larry Ellison. I think it’s also a mistake to treat any tech company like a person. Here’s the full quote: You need to think of Larry Ellison the way you think of a lawnmower. You don’t anthropomorphize your lawnmower, the lawnmower just mows the lawn, you stick your hand in there and it’ll chop…

Torsten Slok Beats the AI-Bear Drum: CHARTS OF THE DAY

Brad DeLong — Sat, 24 Oct 2026 16:34:04 +0000

Torsten’s headline: “No Signs of AI Payoff Outside the Tech Sector”; however caution: measured GDP & non-tech profits are not main events unless you are a focused financebro...

Reading List 10/24/26

Construction Physics — Sat, 24 Oct 2026 12:01:24 +0000

Minimum lot sizes, Boeing’s next 737 MAX, Fervo short sellers, Figure AI’s training data, and more.

Building Reasoning Models from the Ground Up: A Full Video Course

Ahead of AI (Sebastian Raschka) — Sat, 24 Oct 2026 11:14:32 +0000

Over 11 hours of coding videos on evaluation, inference-time scaling, reinforcement learning, and distillation, with Python and PyTorch.

battledrones v14: Compass Rose, a four player map

Simon Willison — Sat, 24 Oct 2026 03:59:57 +0000

Release: battledrones v14: Compass Rose, a four player map I added four player online mode to Battledrones - previous maps were limited to two. Here's the initial prompt I used: Add a new map for a four player battle - four way rotational symmetry, up to four AI players, works with the WebSocket non-audio based multiplayer mode so up to four human players can join (any spare slots use AI) Claude…

Your Book Review: The Ghost In My Brain

Astral Codex Ten — Fri, 23 Oct 2026 23:45:42 +0000

Finalist #13 in the Book Review Contest

Quick Study

Simon Willison — Fri, 23 Oct 2026 21:35:13 +0000

Tool: Quick Study Prompt to Claude Opus 5.5 High : Build an artifact where you open a photo and it rapidly (in a few seconds) paints a version of that photo using a canvas element and simulated paintbrush See also my previous experiment getting Claude Fable to simulate a watercolor illustration with P5.js. Quick Study uses Sobel edge detection , with no external libraries beyond Google Fonts.…

Introducing Kimi K3.1 Pro

Simon Willison — Fri, 23 Oct 2026 19:40:54 +0000

Introducing Kimi K3.1 Pro Kimi K3.1 Pro landed on the Kimi API today - it's not yet open weights, but Moonshot promise "Full model weights will be released by Nov 4, 2026". They're also offering a "Highspeed" version that's 2-3x faster, at twice the price (except for input token cache hits, which are priced the same). The US dollar pricing is $2.00/million tokens for input and $10/million for…

The Epoch Brief - October 23, 2026

Epoch Gradient Updates — Fri, 23 Oct 2026 19:29:17 +0000

Whether AI can do our jobs, how far behind open-weight models are in cybersecurity, and Gemini closing in on the frontier

Amazon’s Trainium Takes a Strike at the Training King

SemiAnalysis — Fri, 23 Oct 2026 19:06:44 +0000

Trainium Best for Inference? But now good for Training too, Native Trainium PyTorch, High Quality CI, TorchTitan, GPU Impossible Profiler

RMSNorm

Giles Thomas — Fri, 23 Oct 2026 18:40:00 +0000

Normalisation is an essential part of LLMs. In the GPT-2 code, we use LayerNorm in a bunch of different places -- in each Transformer block, we do it just before we feed our incoming context vectors into the multi-head attention module, and then again before we feed the result through the FFN: ...and we also do one more round just before we project our context vectors into vocab space via the…

The internet’s Wuhan moment

Exponential View (Azeem Azhar) — Fri, 23 Oct 2026 17:56:58 +0000

AI is turning every old bug into a serious problem. The Internet that survives will be safer--but less open.

Actionable evaluations

Equistamp — Fri, 23 Oct 2026 17:34:04 +0000

Evaluations are useful when they can change actions.

Where is China’s Groq?

ChinaTalk — Fri, 23 Oct 2026 10:56:32 +0000

The U.S. is Making HBM-Free Chips. Why Isn’t China?

The big tech defectors you’re not hearing about

Blood in the Machine (Brian Merchant) — Fri, 23 Oct 2026 09:31:04 +0000

After Tariq Ra’ouf spoke out about Apple’s treatment of Palestinians, he was fired. He joins a growing legion of overlooked workers protesting big tech, and he hasn’t stopped fighting.

Software is slow because of bad design, not slow languages

Sean Goedecke — Fri, 23 Oct 2026 00:00:00 +0000

Many people are saying that AI will make software faster by rewriting it in faster languages, but this isn’t generally true. Most programs cannot be meaningfully sped up by rewriting them in C or Rust. How can this be true? Some languages are orders-of-magnitude faster than others: for instance, C can be fifty times faster than Python. Won’t rewriting my Python program in C thus make it fifty…

Serving DeepSeek-V4.1-Flash at Scale: vLLM Matches DeepInfer and Scales to GB300 NVL72

vLLM blog — Fri, 23 Oct 2026 00:00:00 +0000

With P/D disaggregation and Wide-EP on a GB300 NVL72 rack, vLLM serves DeepSeek-V4.1-Flash on SemiAnalysis AgentX at 954K tokens/s per GPU at a P90 of 196 tokens/s per user, 7.9x our aggregated single-node result.

GLM-5.3 Optimizations, Part 2: PCP, DCP, and Hybrid HiSparse on 16 H200s

vLLM blog — Fri, 23 Oct 2026 00:00:00 +0000

vLLM serves GLM-5.3 on 16 H200s with PCP and DCP prefill and Hybrid HiSparse decode, reaching 1M-token contexts and 1.7–3.2× the per-user speed of data-parallel P/D on the AgentX agentic benchmark.

vLLM Delivers Day-0 Support for MiniMax-M3.1-Flash

vLLM blog — Fri, 23 Oct 2026 00:00:00 +0000

Day-0 vLLM support for MiniMax-M3.1-Flash on NVIDIA Blackwell, NVIDIA Rubin, and AMD MI350-series GPUs, with NVFP4 MSA kernels, NVFP4 MegaMoE, DSpark speculative decoding, AgentX results, and day-0 RL.

San Francisco Meetup This Saturday

Astral Codex Ten — Thu, 22 Oct 2026 19:42:57 +0000

...

DE STAAT - FREEDOM (Official Music Video)

Simon Willison — Thu, 22 Oct 2026 19:42:26 +0000

DE STAAT - FREEDOM (Official Music Video) OK well this is a first for me... at 00:44 in this new music video by Dutch alternative rock band De Staat : The visual effects are credited to Matties Grooten of Fake it or Leave it . Via flir on Hacker News Tags: music , ai , generative-ai , pelican-riding-a-bicycle

Donating Agent Skills to the Agentic AI Foundation

Simon Willison — Thu, 22 Oct 2026 18:09:56 +0000

Donating Agent Skills to the Agentic AI Foundation News today from Anthropic: We’re committed to keeping Agent Skills open, community-driven, and vendor-neutral. Today, we are donating the standard to the Agentic AI Foundation (AAIF) under the Linux Foundation, which we co-founded with Block and OpenAI last year, and to which we donated the Model Context Protocol (MCP). I'm pleased to note I…

Quoting Armin Ronacher

Simon Willison — Thu, 22 Oct 2026 15:34:13 +0000

As fun as AI is, we built so much bad software over the last 30 years, and now the universe is punishing us for our sins. — Armin Ronacher Tags: security , generative-ai , armin-ronacher , ai , ai-security-research , llms

Polls Without People

arg min (Ben Recht) — Thu, 22 Oct 2026 14:36:53 +0000

How statistical prediction blurs into statistical sampling when surveying society.

CROSSPOST: SASHA GUSEV: Vibe Heritability The EJW Symposium on Gregory Clark on Genetics & Social Mobility

Brad DeLong — Thu, 22 Oct 2026 13:06:33 +0000

Charles Murray, Noah Carl & Alex Tabarrok praise Greg Clark’s “To Have & Have Not” without engaging with the statistics; those of us who have tried to do so have a very different view...

Take Ten-Thousand: Egocentric Data Collection for Humanoid Robots

ChinaTalk — Thu, 22 Oct 2026 10:02:03 +0000

why selling robot data is tough business; and ChinaTalk at CoRL!

France and the limits of liberalism

Noahpinion (Noah Smith) — Thu, 22 Oct 2026 08:04:41 +0000

Immigration, the welfare state, and Keynesian economics are all showing signs of strain.

Why Global Trepanation?

Astral Codex Ten — Wed, 21 Oct 2026 21:53:11 +0000

...

Today's PHP Security Release, and How We're Strengthening PHP Security

Simon Willison — Wed, 21 Oct 2026 18:41:53 +0000

Today's PHP Security Release, and How We're Strengthening PHP Security If you're running anything on PHP it's vital you upgrade - today's updates (across the entire PHP 8.x line) address two remote code execution holes, and there have been exploits circulating since last night, thanks to a leak in the coordinated disclosure process. This post from the PHP Foundation explain what went so wrong: On…

CROSSPOST: ROBERT RUBIN: In Six Decades of Public Life, I Never Thought I Would Write This Piece

Brad DeLong — Wed, 21 Oct 2026 18:38:49 +0000

The underestimated one particular potential threat to our economic future: President Trump’s clearly expressed desire to interfere with and even usurp the midterm elections...

The Huawei-DeepSeek Symbiosis

SemiAnalysis — Wed, 21 Oct 2026 17:34:38 +0000

Huawei Ascend 950PR Deep dive & µbenchmarks, DSv4.1 Flash Analysis, DeepSeek support for Ascend

Panache — RTS Belle Époque pour mobile

Simon Willison — Wed, 21 Oct 2026 16:25:16 +0000

Panache — RTS Belle Époque pour mobile Well this is delightful: I released Battle Drones yesterday under an Apache 2.0 License and there's already a French language fork by Gabriel Turpin which reskins the game to take place in steampunk 1900s France and adds an animated intro! Panache est un jeu de stratégie en temps réel pour mobile et navigateur, dans la France de la Belle Époque (vers 1900) :…

Into the Waves

arg min (Ben Recht) — Wed, 21 Oct 2026 14:32:45 +0000

On the most impactful forecast of all time: The Imperial College report on reducing COVID-19 mortality.

MOFCOM's New Rules and the Future of Chinese AI

ChinaTalk — Wed, 21 Oct 2026 13:43:39 +0000

a big deal

Evaluating Agents on Highly Subjective Tasks

Cameron R. Wolfe — Wed, 21 Oct 2026 09:33:32 +0000

Practical tools and techniques for reliably evaluating agents in the presence of noise...

“You Can See Everything” and the nature of liars

Noahpinion (Noah Smith) — Wed, 21 Oct 2026 08:45:19 +0000

A movie review, plus some thoughts on dishonesty, belief, and human communication.

Day-0 Support for Flint on vLLM

vLLM blog — Wed, 21 Oct 2026 00:00:00 +0000

vLLM supports both Flint checkpoints (40B total, 2.5B active) from day 0, decoding a single request at about 500 tokens/s on one GPU.

Announcing Humming: Fast, Flexible Kernels for Quantized Inference

vLLM blog — Wed, 21 Oct 2026 00:00:00 +0000

Introducing Humming, a quantized inference kernel library from Ant Group and the vLLM community, with broad GPU and quantization support, dense and MoE performance results, and vLLM integration.

WorldTUI: I trained a 360M model to pretend to be a terminal

simedw — Wed, 21 Oct 2026 00:00:00 +0000

A 360M model that pretends to be a terminal, and remembers what you did

Shift-1, a SOTA coding router

Applied Compute — Wed, 21 Oct 2026 00:00:00 +0000

eMudhra Intermediate CA Compromise

Simon Willison — Tue, 20 Oct 2026 23:50:25 +0000

eMudhra Intermediate CA Compromise This is very, very bad news: The private keys of eight eMudhra intermediate CA certificates have been compromised. They chain to roots trusted by the Mozilla, Microsoft, Apple and Chrome root programs and by many downstream trust stores. Any software that bundles a list of trusted certificates needs to be updated. This site is tracking software updates via this…

Against Against Handover

Astral Codex Ten — Tue, 20 Oct 2026 20:32:59 +0000

...

Single points of failure in military AI command

Forethought Newsletter — Tue, 20 Oct 2026 20:15:25 +0000

This article was created by Forethought. See all our research on our website.

Latest open artifacts (#25): DeepSeek V4.1 Flash 1017, American-made models, and new regulatory controls in the US & China

Interconnects (Nathan Lambert) — Tue, 20 Oct 2026 17:04:21 +0000

Western open models strike back

The Lithography Lever

ChinaTalk — Tue, 20 Oct 2026 10:46:15 +0000

The case for cutting China off from DUV

Laypeople support technical fields for technology, not art

Sean Goedecke — Tue, 20 Oct 2026 00:00:00 +0000

Some people think that LLMs should not be used in technical fields like programming or mathematics , because those fields are actually art. True or not, this kind of argument is not going to be persuasive to the general population. The argument goes something 1 like this: It doesn’t matter how effective LLMs are at technical fields like programming or mathematics, because one of the essential…

Introducing vLLM TorchTPU

vLLM blog — Tue, 20 Oct 2026 00:00:00 +0000

vllm-torchtpu is a new vLLM backend for Google Cloud TPUs built on TorchTPU, with Pallas kernels, multi-host parallelism, KV cache offloading, and disaggregated serving.

Secondhand pieces from contemporary French brands that aren’t Sézane

Vintage Data (Alexander Doria) — Mon, 19 Oct 2026 22:31:14 +0000

A better way to shop - eBay links to secondhand clothes from French brands starting at $34.

AI will not automate radiologists like you think

Jason Wei — Mon, 19 Oct 2026 19:23:52 +0000

We have been talking for years now about AI automating jobs. Radiology has been a prime target, with Geoff Hinton once claiming, “people should stop training radiologists now, it’s just completely obvious that within five years, deep learning is going to do better than radiologists”, which is funny since he said that in 2016. There is now more demand for radiologists than ever, so clearly things…

Open Thread 456

Astral Codex Ten — Mon, 19 Oct 2026 17:51:46 +0000

...

AI is ushering in an age of personal media

Understanding AI (Timothy B. Lee) — Mon, 19 Oct 2026 13:33:47 +0000

It may be AI slop, but it's my AI slop.

📈 Monday data: Who won the trade war?

Exponential View (Azeem Azhar) — Mon, 19 Oct 2026 13:31:29 +0000

What eight years of tariffs actually did

Drone Dominance

ChinaTalk — Mon, 19 Oct 2026 10:04:01 +0000

the Pentagon’s $1bn drone contest

Datacenters 2027: The Year of Acceleration, +50GW in a Single Year

SemiAnalysis — Mon, 19 Oct 2026 05:39:38 +0000

60GW Under Construction, SemiAnalysis Leading Indicators, 1.3T of Lease Contracts Signed YTD, The Shortage Could Move to Chips

Can a Million AI Agents Create Superintelligence?

Noahpinion (Noah Smith) — Mon, 19 Oct 2026 04:33:30 +0000

Ramez Naam gives a skeptic’s take on agent swarms.

What requirements don't you have?

Sean Goedecke — Mon, 19 Oct 2026 00:00:00 +0000

When designing software, the requirements you don’t have are as important as the requirements you do. Identifying what you can compromise on — what can be slow, or stale, or inflexible — is usually the key to an elegant design. Elegant systems drop some requirements For instance, take preforking web servers like Unicorn . Because they lack the feature of shared mutable state between server…

Achieving near-zero-downtime scaling in Elastic Expert Parallelism

vLLM blog — Mon, 19 Oct 2026 00:00:00 +0000

How asynchronous preparation and NIXL EP enabled near-zero-downtime Elastic EP scaling in vLLM.

Why are we ruled by barbarians?

Noahpinion (Noah Smith) — Sun, 18 Oct 2026 06:34:00 +0000

America’s New Right believes that savagery is the key to victory.

🔮 A world of many minds & the future of work

Exponential View (Azeem Azhar) — Sun, 18 Oct 2026 03:46:33 +0000

Good morning from London.

Writing is no longer proof of work

Sean Goedecke — Sun, 18 Oct 2026 00:00:00 +0000

Tech companies structurally depend on writing to serve as evidence, in the same way that Bitcoin depends on proof-of-work hashing. However, in the last few years, LLMs have started to undermine this practice. How can writing be like proof-of-work? When I’m writing something at work, it’s not always to communicate an idea. Sometimes it’s because having a piece of writing is useful as evidence : of…

Why Might Aliens Visit Earth?

Overcoming Bias (Robin Hanson) — Sat, 17 Oct 2026 23:58:36 +0000

I did a web search on the above phrase, in prep for a podcast, and in ten pages of results didn’t find anything I thought reasonable.

Meta shakes up the frontiers, open and closed

Interconnects (Nathan Lambert) — Sat, 17 Oct 2026 14:06:54 +0000

US open weights are back!

How Philanthropy Might Help Us Reach Better Futures (with Will MacAskill and Mia Taylor)

Forethought Newsletter — Sat, 17 Oct 2026 13:51:34 +0000

A podcast episode from Forethought

Reading List 10/17/26

Construction Physics — Sat, 17 Oct 2026 12:03:37 +0000

Icon’s fundraising, employment in China’s solar industry, grid-scale battery reliability, Tesla’s flying roadster, and more.

The Case for DUV Controls

ChinaTalk — Sat, 17 Oct 2026 09:07:33 +0000

with Saif Khan!

Poppy PCR - Analyzing the DNA of a California Poppy Mutant

Jonathan Whitaker — Sat, 17 Oct 2026 00:00:00 +0000

In early June I noticed some white flowers in amidst the prolific and gorgeous orange California poppies in the flower beds along one of our regular walk routes. Intrigued, I took some samples and preserved some tissue from each in Zymo DNA shield, thinking I’d later look into options to sequence them or otherwise probe for the mutation that caused the loss of pigment. A bit of research turned up…

Your Book Review: Storm Of Steel

Astral Codex Ten — Fri, 16 Oct 2026 23:39:56 +0000

Finalist #12 in the Book Review Contest

Radar guns for AI risk

Equistamp — Fri, 16 Oct 2026 15:08:26 +0000

AI safety evaluations as measurement instruments.

Predictive Models Are Never Clean

arg min (Ben Recht) — Fri, 16 Oct 2026 14:26:46 +0000

Revisiting Breiman’s Two Cultures paper and what we want from our forecasting models.

Why a Chinese AI Lab’s Product Lead Thinks AGI is “Inevitable”

ChinaTalk — Fri, 16 Oct 2026 11:08:34 +0000

and nobody knows what it’ll look like

California vs. the tech billionaires

Blood in the Machine (Brian Merchant) — Fri, 16 Oct 2026 09:30:38 +0000

Inside the fight to pass California’s Prop 40, and to stop tech titans’ dystopian vision for working people from coming to pass. With CA labor leader Lorena Gonzalez. Plus: Is the AI bubble bursting?

Meetups Everywhere 2026: Times & Places

Astral Codex Ten — Fri, 16 Oct 2026 04:04:26 +0000

...

Prefer Evolved To Folk Morality

Overcoming Bias (Robin Hanson) — Fri, 16 Oct 2026 01:36:29 +0000

Individuals often have good reasons to indulge in delusions, such as re superstition, astrology, religion, or their own personal destiny.

AI is driving a boom in modestly successful software projects

Sean Goedecke — Fri, 16 Oct 2026 00:00:00 +0000

If AI is so great, why aren’t we seeing an explosion of new, amazing software products? If anyone can turn their idea into working software with coding agents, where is all that software ? People have been asking this question for years. As far as I can tell, the answer is that it’s spread out over a lot of modestly successful projects: Somewhere between 2x and 20x new software is being created,…

Open-weight models trail closed models in cyber capabilities by 6–7 months

Epoch AI — Fri, 16 Oct 2026 00:00:00 +0000

Z.ai’s GLM-5.3 is the top open-weight model on Epoch’s Cyber ECI, scoring where the closed-weight frontier stood 6–7 months earlier.

Hyper Optimizing Kimi K3 in vLLM: Moving Fast and Responsibly

vLLM blog — Fri, 16 Oct 2026 00:00:00 +0000

How agent-assisted CuTe DSL optimizations accelerate Kimi K3 decode on NVIDIA GB300, and how scoped changes, correctness tests, and reproducible benchmarks help ship them responsibly.

Scaling Kimi K3 on AMD Instinct MI355X: Matching NVIDIA HGX B300 Throughput per GPU on AgentX

vLLM blog — Fri, 16 Oct 2026 00:00:00 +0000

Inside the Kimi K3 optimization effort: understanding the bottlenecks and the decisions that made inference faster

Serving Reflection Beam on vLLM from Day 0

vLLM blog — Fri, 16 Oct 2026 00:00:00 +0000

vLLM supports Reflection's Beam-1.0 (501B total, 23B active) from day 0, with agent-driven optimization that cut decode latency 2.71× and DFlash2 speculative decoding with adaptive verification.

ASML Is Building a Hybrid Bonding Tool!

SemiAnalysis — Thu, 15 Oct 2026 19:27:20 +0000

ASML is building a hybrid bonding tool! The company has not announced anything yet, but its patent filings, a news report and its own roadmap all suggest so.

🚨 Anthropic: The agony and the ecstasy

Exponential View (Azeem Azhar) — Thu, 15 Oct 2026 16:29:01 +0000

Hi, I sent this to members of our AI Investment Brief a couple of hours ago.

Directed Culture Trials

Overcoming Bias (Robin Hanson) — Thu, 15 Oct 2026 16:20:29 +0000

Unless we find new ways to drive adaptation besides mass death and fragmentation, our civilizations may keep falling and rising indefinitely.

How to Train Your Robot

Construction Physics — Thu, 15 Oct 2026 14:23:58 +0000

The enormous advance in AI capabilities over the last several years has in large part been driven by the availability of vast amounts of training data.

AI vs. the railroads

Noahpinion (Noah Smith) — Thu, 15 Oct 2026 08:47:39 +0000

Are we in 1872?

AI Safety Is Not A Corporate Plot

Astral Codex Ten — Thu, 15 Oct 2026 01:15:38 +0000

...

Day 0 support for Muse Spark Open on vLLM

vLLM blog — Thu, 15 Oct 2026 00:00:00 +0000

Day-0 vLLM support for Muse Spark Open: 816B MoE, 1M context, multimodal inputs, DSpark speculative decoding, NVIDIA and AMD serving, initial TPU support, and LoRA RL.

Announcing Day 0 Support for NVIDIA Nemotron 3.5 Super VL on vLLM

vLLM blog — Thu, 15 Oct 2026 00:00:00 +0000

Run NVIDIA Nemotron 3.5 Super VL on vLLM for multimodal agent workflows with image, document, and video understanding, OpenAI-compatible serving, and multi-token prediction.

You could have invented transformers from memory

Alex Wang — Thu, 15 Oct 2026 00:00:00 +0000

Large language models today can recall obscure facts from every corner of the world. Most recently, I was running late for a friend’s board game night, so I asked Astra to remind me the rules to Dune: Imperium. It gladly narrated them to me as I drove over and helped me remember the rules well enough for me to play again. We’re so used to asking models about anything that we take their vast…

OpenAI’s ironic credibility problem

Understanding AI (Timothy B. Lee) — Wed, 14 Oct 2026 18:41:56 +0000

How do we know the company’s stated reasons for firing three safety researchers are its actual reasons?

Space weapons could shut others out of space and erode nuclear deterrence

Forethought Newsletter — Wed, 14 Oct 2026 12:30:14 +0000

This article was created by Forethought. See all our research on our website.

The AI Pit Crew Is Hiring.

From the Porch (dslevy) — Wed, 14 Oct 2026 12:03:04 +0000

Scott Werner predicted an AI pit crew for every team. ~2K job boards say it’s already happening, under at least four names.

“I Will Complain to the EU!”

ChinaTalk — Wed, 14 Oct 2026 10:24:58 +0000

How 107 laid-off Chinese graduates turned BMW, VW, and Mercedes into leverage

AI systems are improving rapidly at Slay the Spire and its sequel

Epoch AI — Wed, 14 Oct 2026 00:00:00 +0000

Months ago, frontier models could barely beat Slay the Spire. Now they win 90%+ at the easiest setting and up to 20% at the hardest. Results from Epoch AI’s new SpireClimb benchmark.

Adaptive Verification Without a Confidence Head

vLLM blog — Wed, 14 Oct 2026 00:00:00 +0000

An online-fitted acceptance estimator that lets every draft-model speculator use Adaptive Verification, by predicting the target's acceptance probability from the drafter's own logits and refitting continuously during serving.

Push-Based KV Transfer to Improve TTFT and TPOT for Disaggregated Inference in vLLM

vLLM blog — Wed, 14 Oct 2026 00:00:00 +0000

A push-based NIXL connector for vLLM disaggregated serving: the prefill instance writes KV directly into decode memory over RDMA, taking the router relay, decode-side allocation, and transfer setup off the TTFT critical path.

Finish RL Training Faster: Bitwise-Consistent, High-Performance, Agent-Driven

vLLM blog — Wed, 14 Oct 2026 00:00:00 +0000

We make vLLM rollout and the training forward give bitwise identical logprobs on DeepSeek-V4-Flash and Nemotron-3.5-Lightning-30B-A3B (NVFP4), keep 91% of rollout throughput on DeepSeek-V4-Flash, and open-source the agent skill.

Sézane hate mail

Vintage Data (Alexander Doria) — Tue, 13 Oct 2026 23:45:35 +0000

When you make a deal with the devil you end up looking like Banana Republic

WarTalk: AutoWarCom and Project Meridian

ChinaTalk — Tue, 13 Oct 2026 11:04:37 +0000

+ Newt returns!

A Nobel for New Keynesian macro, at last!

Noahpinion (Noah Smith) — Tue, 13 Oct 2026 05:28:43 +0000

It’s about time.

Open Thread 455

Astral Codex Ten — Tue, 13 Oct 2026 00:19:32 +0000

...

Enriching Domain-Specific Post-Training with Multi-Teacher On-Policy Distillation

Applied Compute — Tue, 13 Oct 2026 00:00:00 +0000

Faster Fourier Transforms

arg min (Ben Recht) — Mon, 12 Oct 2026 14:01:02 +0000

The OpenAI algorithm for the FFT is less a breakthrough and more a malicious exploit.

📈 What is bigger than AI?

Exponential View (Azeem Azhar) — Mon, 12 Oct 2026 12:29:50 +0000

GLP-1s: the science, the savings and the societal stakes.

History of 20 Sources of Moral Authority

Overcoming Bias (Robin Hanson) — Mon, 12 Oct 2026 00:52:43 +0000

Though three weeks ago I outlined a story of the long-term history of changes to moral authority, I decided to study this in a bit more detail.

Software engineers don't get to be Luddites about AI

Sean Goedecke — Mon, 12 Oct 2026 00:00:00 +0000

Many software engineers are righteously angry about AI. It’s understandable to be upset that the skills you built your professional identity around are getting automated away. But software engineers are the last people in the world who should be complaining about this. If you live by the sword, you will perish by the sword 1 . Software engineering has always automated industries From the very…

Improve TTFT for Multi-Turn Agentic Workloads with Bi-Directional KV Transfer in vLLM

vLLM blog — Mon, 12 Oct 2026 00:00:00 +0000

Bi-directional KV transfer lets vLLM prefill nodes reuse previously computed KV from decode nodes on later conversation turns, cutting redundant prefill recompute and reducing TTFT by up to 3.3x on long multi-turn prompts.

MAGI-2 Preview on vLLM-Omni: Bringing a 114B Audio-Video MoE DiT to Moore Threads MUSA with torchada

vLLM blog — Mon, 12 Oct 2026 00:00:00 +0000

vLLM-Omni runs Sand AI's MAGI-2 Preview audio-video generation model on eight Moore Threads MTT S5000 GPUs through torchada's CUDA compatibility layer, then optimizes MoE, mHC, attention, parallelism, and compilation.

10 Takeaways from Q3

ChinaTalk — Sun, 11 Oct 2026 11:08:09 +0000

AI, China, defense, national security, supply chains, the Protestant roots of Kim Il-sung’s Messiah complex, the lethality of hydrofluoric acid, and horseback archery in Kyrgyzstan.

Why Donald Trump won in 2024

Noahpinion (Noah Smith) — Sun, 11 Oct 2026 07:31:44 +0000

Re-upping my election post-mortem, so that we don’t forget.

🔮 AI & Math 2.0, Pakistan’s solar hedge & who controls computing++

Exponential View (Azeem Azhar) — Sun, 11 Oct 2026 04:44:15 +0000

An essential weekly reading on AI.

Don't diagnose your opponents

Sean Goedecke — Sun, 11 Oct 2026 00:00:00 +0000

When you disagree with someone, there are two paths forward. The first is to try to engage with what they’re saying: what their reasons are, whether their position is true, what evidence could be brought to bear, and so on. The second is to diagnose them: to engage with the second-order question of why they’re saying it. For instance, if you’re arguing with somebody about microservices 1 , you…

Reading List 10/10/26

Construction Physics — Sat, 10 Oct 2026 14:55:48 +0000

Propane in air conditioners, Anduril making submarine parts, nuclear reactors on the moon, the robotic hand landscape, and more.

🔮 The end of screen time

Exponential View (Azeem Azhar) — Sat, 10 Oct 2026 06:02:50 +0000

AI changes how we work, but can it also change where?

Your Book Review: We Present Television

Astral Codex Ten — Sat, 10 Oct 2026 01:05:13 +0000

Finalist #11 in the Book Review Contest

Software's centaur age may last decades

Sean Goedecke — Sat, 10 Oct 2026 00:00:00 +0000

We are currently living in the age of centaurs (2022-20??). Right now, human engineers paired with AI coding systems are more effective than either by themselves. For future generations of software engineers — human or otherwise — this will be the most significant fact about the current period. As in chess’s own centaur age, people are always ready to claim that the centaur age is over and the…

Software Is Meant To Be Called, Not Clicked.

From the Porch (dslevy) — Fri, 09 Oct 2026 22:17:42 +0000

6,178 startups’ public code says the next customer doesn’t have a seat.

I expect rapid progress but not towards general superintelligence

Interconnects (Nathan Lambert) — Fri, 09 Oct 2026 21:33:47 +0000

I’ve often been surprised when I hear from top researchers in industry that they think AI will be better than them at their job in a few years, and I didn’t really know why I doubted it.

Model backdoors are easy to create and hard to reliably remove

Equistamp — Fri, 09 Oct 2026 21:01:50 +0000

Backdoors present novel asymmetries that favor attackers. We discuss results from a new paper, how to think about backdoors, and tools and techniques to address them.

Evaluation quality

Equistamp — Fri, 09 Oct 2026 19:23:33 +0000

Validity and reliability of AI evaluations.

The 15 best Luddite horror films

Blood in the Machine (Brian Merchant) — Fri, 09 Oct 2026 19:08:36 +0000

From Godzilla to Pulse to the Terminator, these are the best films that register an aggressive critique of technology.

Can AI automate AI R&D yet?

Epoch Gradient Updates — Fri, 09 Oct 2026 18:41:00 +0000

Not yet, but it will claim otherwise

Fun with low-rank vocab matrices (and a bonus test loss reduction?)

Giles Thomas — Fri, 09 Oct 2026 16:00:00 +0000

I was nerdsniped ! On the HF discussion page for one of the models I created for my previous post , AndrewThompson1233 asked if I'd considered trying out factorised embeddings -- something he's using for his model, Maba, and which was previously used in some other models, like ALBERT . It's a really nifty idea; you use a similar trick to LoRA as a way of reducing the number of parameters used for…

Kevin Roose on the AGI Believers: Oppenheimer with Equity

ChinaTalk — Fri, 09 Oct 2026 10:03:10 +0000

this one was so fun

Maybe the Rapture is already here

Noahpinion (Noah Smith) — Fri, 09 Oct 2026 09:18:32 +0000

In the future, we’re all heavenbanned.

vLLM Support for NVIDIA Vera Rubin NVL72: 7.8x Throughput over GB200 NVL72

vLLM blog — Fri, 09 Oct 2026 00:00:00 +0000

vLLM now runs on NVIDIA Vera Rubin NVL72 with day-0 model support, Rubin-tuned FlashInfer kernels, and locality-aware MoE, delivering 7.8x the per-GPU throughput of GB200 NVL72 on AgentX.

Matrix gradients, without the pain

Alex Wang — Fri, 09 Oct 2026 00:00:00 +0000

In this post I want to share a simple trick to make computing gradients and backward passes mechanical, without drowning in indices and ruining your eyesight. To start with, let’s remember how we all first learned to compute gradients. Suppose we have a linear layer $Y = XW$, where $X$ is an $N\times D$ batch of inputs and $W$ is a $D\times F$ weight matrix, followed by some scalar loss $L(Y) \in…

Train your own decision model in AC2

Applied Compute — Fri, 09 Oct 2026 00:00:00 +0000

Ballots Everywhere: Times & Places

Astral Codex Ten — Thu, 08 Oct 2026 22:00:52 +0000

...

Moralized Trends Are Often Extended

Overcoming Bias (Robin Hanson) — Thu, 08 Oct 2026 21:40:52 +0000

Compared to the ancients, we moderns less moralize our personal lives, but more moralize the bigger world around us.

The Epoch Brief - October 8, 2026

Epoch Gradient Updates — Thu, 08 Oct 2026 19:30:49 +0000

Plummeting AI costs, billions of agents, how Chinese labs make money, AI’s impact on behaviors, and our benchmark auditing initiative

Beijing Will Not Pace the Frontier: China’s Speed-First AI Safety Regime

SemiAnalysis — Thu, 08 Oct 2026 17:46:05 +0000

AI safety is on fire.

Unimaginative Uncertainty

arg min (Ben Recht) — Thu, 08 Oct 2026 17:00:04 +0000

How most uncertainty can’t be quantified.

Understanding Jev, the new model everyone is talking about

Understanding AI (Timothy B. Lee) — Thu, 08 Oct 2026 13:14:53 +0000

OpenAI, CloudFlare, and Amazon have all copied this tiny startup.

The Rise and Fall of the Plasma Screen

Construction Physics — Thu, 08 Oct 2026 12:02:36 +0000

In 1960, a team at the University of Illinois released the first version of their PLATO “computerized instruction system,” a network of what would eventually be thousands of computer terminals connected to large mainframe computers.

The Spice Trader vs. the Tariffs

ChinaTalk — Thu, 08 Oct 2026 11:04:16 +0000

a section 301 update from the litigant

Mistaking Context for Virtue

Overcoming Bias (Robin Hanson) — Thu, 08 Oct 2026 02:40:14 +0000

Imagine taking a big group of elites from the year 1700, giving them the tech, wealth, and economic options of today, also teaching them new facts that we have learned since then, but without putting them in contact with the people of today.

Can AI automate Epoch?

Epoch AI — Thu, 08 Oct 2026 00:00:00 +0000

Frontier models can handle well-defined tasks reliably, but fail in the more open-ended aspects that prevent full work automation. Explore Epoch AI's Automation Reports: realistic, open-ended tasks taken directly from our work, manually graded by our own standards.

Top Forecasters On The Plague

Astral Codex Ten — Wed, 07 Oct 2026 21:13:16 +0000

...

Hidden Open Thread 454.5

Astral Codex Ten — Wed, 07 Oct 2026 21:07:33 +0000

...

🔮 What is left to do

Exponential View (Azeem Azhar) — Wed, 07 Oct 2026 13:50:46 +0000

The future of science may be provable, profitable and incomprehensible

There is no “Indian Question”

Noahpinion (Noah Smith) — Wed, 07 Oct 2026 09:47:37 +0000

The actual question is what kind of a nation America is going to be.

How to read code

Sean Goedecke — Wed, 07 Oct 2026 00:00:00 +0000

Everyone knows how to read a book. Beginning at the first page, you read each word in order 1 , stopping periodically to think, until you arrive at the last word on the last page. That’s how you read a newspaper article, or a poem, or an email. Why would reading code be any different? Why code is different English text is designed to be read in order. In fact, there are almost no constraints on…

US adults are no more likely to face cyber incidents than when Claude Fable 5 launched

Epoch AI — Wed, 07 Oct 2026 00:00:00 +0000

June and September polls from Epoch AI/Ipsos show no detectable increase in the share of US adults experiencing cyber incidents

EBR-bench update

Epoch AI — Wed, 07 Oct 2026 00:00:00 +0000

Updates on EBR-bench, Epoch AI’s benchmark that tests models’ ability to learn from experience. GPT-6 Astra hit a perfect score on over half its EBR-bench attempts by exploiting one overpowered card, which we now ban. Plus multi-agent scaffold results.

Can AI automate AI R&D yet?

Epoch AI — Wed, 07 Oct 2026 00:00:00 +0000

Early evidence from InnovationEval: No

DeepSeek-V4.1-Flash on vLLM: 5x Agentic Throughput Since Day 0

vLLM blog — Wed, 07 Oct 2026 00:00:00 +0000

Within three weeks of release, vLLM made DeepSeek-V4.1-Flash 1.9x faster at low concurrency and lifted its throughput 5x on SemiAnalysis AgentX, with SWA bounded replay, CUDA graphs, DeepSeek's new kernels, and vLLM kernel fusions.

Scaling On-Policy Self-Distillation with Hint Auto-Optimization

Applied Compute — Wed, 07 Oct 2026 00:00:00 +0000

An Open Letter To Steven Pinker On AI

Astral Codex Ten — Tue, 06 Oct 2026 21:23:06 +0000

...

Too Big to Fail

arg min (Ben Recht) — Tue, 06 Oct 2026 14:32:53 +0000

The obsession with and futility of macroeconomic forecasting

The Cyber Risk Discourse is Broken

Interconnects (Nathan Lambert) — Tue, 06 Oct 2026 14:22:41 +0000

Open-weights, ideology, and acknowledging trade-offs.

The best three hours in San Francisco

Jason Wei — Tue, 06 Oct 2026 03:55:43 +0000

Visiting San Francisco for the first time? Welcome. The thing to understand about SF is that it is a city of duality. It is probably the only city where you can see a self-driving car drive itself past a homeless person. Within a few dozen blocks, you can traverse the entire spectrum of the human condition, from the dense opulence of tree-lined mansion neighborhoods (Presidio Heights) to…

China is more exposed to semiconductor supply-chain disruptions than the US

Epoch AI — Tue, 06 Oct 2026 00:00:00 +0000

China is 2.7× as exposed to semiconductor supply shocks as the US, Epoch AI estimates. Epoch traces chips through input-output tables to model the impact of a Taiwan shock and decoupling.

How do Chinese AI companies make money?

Epoch AI — Tue, 06 Oct 2026 00:00:00 +0000

China's six leading AI firms earn about 10% of OpenAI and Anthropic's combined AI revenue. Epoch AI examines their five revenue streams: consumer apps, model access, enterprise and government solutions, licensing fees, and indirect monetization through cloud and advertising.

Secondhand cashmere sweaters for guys

Vintage Data (Alexander Doria) — Mon, 05 Oct 2026 23:55:29 +0000

Skip the Quince dupe. Buy a sweater that’s better for you, better for the goat, and better for Mongolia.

Humans Are Not Chinchillas: Revisiting “How Quick and Big Would A Software Intelligence Explosion Be?”

Forethought Newsletter — Mon, 05 Oct 2026 23:37:42 +0000

This article was created by Forethought. See all our research on our website.

Philosophical Discussion on the Far Future of AI

Sebastian Raschka (site) — Mon, 05 Oct 2026 23:13:28 +0000

A discussion with Jürgen Schmidhuber and host Jake Newfield about the far future of AI, human control, and whether AI could develop its own goals.

Anthropic Subscriptions Offer 5x+ More Value Than OpenAI

SemiAnalysis — Mon, 05 Oct 2026 20:01:09 +0000

Limit testing every AI subscription plan from Anthropic, OpenAI, Meta, SpaceXSI, MiniMax, Moonshot, Z.ai, Cursor, and Cognition

Introducing Dan Kagan-Kans

Understanding AI (Timothy B. Lee) — Mon, 05 Oct 2026 19:11:19 +0000

Understanding AI has a new writer!

Current AIs out-persuade professionals in lab settings but (probably) not in the world

Forethought Newsletter — Mon, 05 Oct 2026 18:02:13 +0000

This article was created by Forethought. See all of our research on our website.

Why agent swarms could be the next “scaling law”

Understanding AI (Timothy B. Lee) — Mon, 05 Oct 2026 17:19:00 +0000

An OpenAI researcher called it “the most ‘feel the AGI’ moment that I had since reasoning models.”

📈 Monday data: More AI, more justice?

Exponential View (Azeem Azhar) — Mon, 05 Oct 2026 13:51:44 +0000

Everyone’s got a lawyer now

Open Thread 454

Astral Codex Ten — Mon, 05 Oct 2026 11:09:46 +0000

...

What the AI boom is doing for Americans

Noahpinion (Noah Smith) — Mon, 05 Oct 2026 08:12:53 +0000

It’s not just “of, by and for oligarchs”.

Coding-agent spending by OpenAI researchers is doubling roughly every month

Epoch AI — Mon, 05 Oct 2026 00:00:00 +0000

OpenAI researchers’ coding-agent usage, valued at API prices, has recently been doubling roughly every month, according to breakpoint fits.

🔮 The transition is hiding in plain sight #604

Exponential View (Azeem Azhar) — Sun, 04 Oct 2026 04:07:11 +0000

AI consciousness, electrification’s hidden progress, and the coming agentic bank run++

Five Slogans Re Cultural Drift

Overcoming Bias (Robin Hanson) — Sun, 04 Oct 2026 01:54:36 +0000

I’m wondering how well I might organize my cultural drift messages around a few key slogans.

Reading List 2026-10-03

Construction Physics — Sat, 03 Oct 2026 13:14:44 +0000

A Senate permitting reform bill, the potential of Tesla’s Cybercab, the Navy’s $200 billion worth of shipyard renovations, returning a SR-71 to service, and more.

🔮 Quick weekend reads: the big AI questions

Exponential View (Azeem Azhar) — Sat, 03 Oct 2026 12:54:07 +0000

A short set of readings on some of the biggest questions.

The second Trump presidency is a world-historic failure

Noahpinion (Noah Smith) — Sat, 03 Oct 2026 08:47:17 +0000

How did Trump fumble so badly?

Superpersuasion will look like bribery

Sean Goedecke — Sat, 03 Oct 2026 00:00:00 +0000

The idea of “superpersuasion” has been floating around the AI safety community for decades. Now that powerful and difficult-to-control LLMs have appeared, people are again talking about the idea that a sufficiently intelligent AI might be able to persuade people to do whatever it wants. The classic 1 version of this idea is an AI persuading some engineer to “let it out of the box”: to grant it…

Shipping is the foundation

Sean Goedecke — Sat, 03 Oct 2026 00:00:00 +0000

There are lots of skills involved in being a successful engineer at a big tech company. Running effective meetings, writing good design docs, fleshing out tickets and epics, improving team processes, tech leading small groups of engineers, and so on. But all of these rest on the foundation of being able to ship. In Dota 2, the early game involves “laning”: competing with an enemy player to…

Hundreds of millions of AI agents are coming. Is there work for them?

Epoch Gradient Updates — Fri, 02 Oct 2026 19:11:16 +0000

The AI compute buildout could support more AI agents than the US has workers

Evaluation information efficiency and covering the risk landscape

Equistamp — Fri, 02 Oct 2026 18:32:13 +0000

How to spend a finite evaluation budget

[STAX Brief] AI Killed the Website Builder Star.

From the Porch (dslevy) — Fri, 02 Oct 2026 10:59:07 +0000

Startups are moving their sites from no-code builders to Vercel’s Next.js and Cloudflare’s Astro, and the evidence points at Claude Code.

The two words the AI industry hates more than anything

Blood in the Machine (Brian Merchant) — Fri, 02 Oct 2026 09:02:22 +0000

I talk with computational linguist Emily M. Bender, co-author of the influential 2021 “stochastic parrots” paper, about why AI industry leaders can’t stop seething over the term.

Do not build the LLM torture factory

Sean Goedecke — Fri, 02 Oct 2026 00:00:00 +0000

It isn’t hard to torture an LLM. You simply extract a steering vector that corresponds with LLM “discomfort” , then artificially boost that steering vector. Tortured LLMs will describe experiencing negative feelings and will press a “pain relief” button even when it conflicts with their overall goals. It’s therefore easy to build a harness that runs tens or hundreds of tortured LLMs in parallel.…

How many AI agents could run on the AI chips shipped through 2027?

Epoch AI — Fri, 02 Oct 2026 00:00:00 +0000

Memory shipped through 2027 could run 33–171 million concurrent frontier-model agents, or billions using efficient open models. Epoch AI estimates inference capacity from HBM supply, serving benchmarks, and agent-hour costs.

HWREBench

Jonathan Whitaker — Fri, 02 Oct 2026 00:00:00 +0000

(This post is also the homepage of https://hwrebench.com/, which has a bunch of other posts and things. Sharing here to make it easier to find. The project is ongoing but a lot slower now that I have a day job. Check hwrebench for updates.) Illustration of some of the device vulnerabilities found in this exercise (AI generated, unlike the text of this article) AI <-> Cyber discourse has tended to…

California just passed the strongest AI labor laws in the US

Blood in the Machine (Brian Merchant) — Thu, 01 Oct 2026 20:50:46 +0000

Finally, some good news

Why do OpenAI's GPT-2 weights beat mine? Part five: data quality

Giles Thomas — Thu, 01 Oct 2026 17:30:00 +0000

When I finished learning how to build an LLM from scratch , I was left with a mystery: my own models were not as good as OpenAI's original GPT-2 models, despite being based on the same architecture. My models all had 163M parameters, and followed the design from Sebastian Raschka 's book " Build a Large Language Model (from Scratch) ". That meant that they were pretty much the same as the setup…

The Keynesian Subtweet

arg min (Ben Recht) — Thu, 01 Oct 2026 14:17:09 +0000

Utility maximization is unescapable. We should learn when it’s misapplied.

Understanding the AI That Drives Robots

Construction Physics — Thu, 01 Oct 2026 12:04:36 +0000

An Explanation of Vision-Language-Action Models

Roundup #89: It isn’t X, it’s Y

Noahpinion (Noah Smith) — Thu, 01 Oct 2026 06:14:51 +0000

AI risk; Reindustrialization; The rent crisis; AI and jobs; Mass deportation; Chinese investment; Cyber risk

Introducing the ChatGPT usage explorer

Epoch AI — Thu, 01 Oct 2026 00:00:00 +0000

Median monthly messages per active user rose from 14 in 2023 to 36 in 2025, and the top 10% of users sent 63% of all prompts. Epoch AI's ChatGPT usage explorer draws on 8.3 million messages from 5,000 US YouGov panelists.

How reward hacking detection works in Ari

Applied Compute — Thu, 01 Oct 2026 00:00:00 +0000

Four Sets of Abstract Moral Systems

Overcoming Bias (Robin Hanson) — Wed, 30 Sep 2026 20:16:16 +0000

Morality is collective; we coordinate together to define and then enforce shared morals.

You Should Always Go for It

arg min (Ben Recht) — Wed, 30 Sep 2026 14:15:03 +0000

Nothing shows off the absurdity of cost-benefit analysis better than football analytics.

🚨 The first existential IPO

Exponential View (Azeem Azhar) — Tue, 29 Sep 2026 15:54:57 +0000

On the leaked details of Anthropic’s IPO prospectus

Pricing Commensurability

arg min (Ben Recht) — Tue, 29 Sep 2026 15:26:04 +0000

On the origins of cost-benefit analyses in governmental decision making

Language Models for Text Classification: From Bag-of-Words to Jev

Ahead of AI (Sebastian Raschka) — Tue, 29 Sep 2026 10:50:25 +0000

A Visual Guide to RNNs, CNNs, Transformers, and Calibration, with Hands-On Experiments on Accuracy and Efficiency

Innovation Without Feedback Is Drift

Overcoming Bias (Robin Hanson) — Tue, 29 Sep 2026 00:58:19 +0000

The ancient world respected tradition, and was wary of most kinds of change.

Taking vLLM Apart: A Practical Guide to Disaggregated Serving

vLLM blog — Tue, 29 Sep 2026 00:00:00 +0000

What disaggregated serving actually buys you, how to run it end to end in vLLM today with prefill/decode plus the new GPU-less frontend and the things we're still working on.

Stop spending money on smaller classes

Noahpinion (Noah Smith) — Mon, 28 Sep 2026 22:27:25 +0000

We’ve spent decades raising teacher quantity when we should have been raising quality.

How GLM5.3 Sparse Attention Affects HBM Memory Usage

SemiAnalysis — Mon, 28 Sep 2026 19:26:41 +0000

GLM-5.3, KV Cache Offloading, HiSparse, DeepSeek Sparse Attention, IndexShare, Single-rollout Asynchronous Optimization, Cybersecurity

AI is getting cheaper faster than any other transformative technology

Epoch Gradient Updates — Mon, 28 Sep 2026 15:33:44 +0000

The price is falling 13x year over year

📈 Monday data: Damming the slop floods

Exponential View (Azeem Azhar) — Mon, 28 Sep 2026 14:38:32 +0000

AI makes content cheap. Who pays to make it trustworthy?

[PERSONAL/STAX]: One Less Colon. A Lot More Context.

From the Porch (dslevy) — Mon, 28 Sep 2026 11:02:48 +0000

Part 2: ChatGPT can consult my records, doctors, and data. STAX shows what’s emerging behind it.

Notes on NVIDIA Nemotron

Cameron R. Wolfe — Mon, 28 Sep 2026 09:33:59 +0000

A broad coverage of the most practically useful series of open LLM tech reports...

A Thousand AI Constitutions: A Short Defense

Forethought Newsletter — Mon, 28 Sep 2026 02:56:58 +0000

A guest post by Simon Goldstein and Peter N. Salib.

Credentials API for Gemini Managed Agents

Philipp Schmid — Mon, 28 Sep 2026 00:00:00 +0000

Store secrets once with the Credentials API, reference them by ID, and let the egress proxy inject them on the wire so they never enter the agent sandbox.

Focusing on Post-Training

Sebastian Raschka (site) — Sun, 27 Sep 2026 22:12:03 +0000

Why I'd invest in post-training existing open-weight LLMs, with Fireworks' Ember-1 as an example of more token-efficient reasoning.

Where’s the “intelligence explosion”?

Noahpinion (Noah Smith) — Sun, 27 Sep 2026 07:20:08 +0000

Ramez Naam gives a skeptic’s take on Recursive Self-Improvement.

🔮 The pros and cons of collective AI #603

Exponential View (Azeem Azhar) — Sun, 27 Sep 2026 04:06:52 +0000

AGI and new moral subjects; ChatGPT drives a car; energy and AI need each other++

Human-AI partnerships are for alignment, not capability

Sean Goedecke — Sun, 27 Sep 2026 00:00:00 +0000

It’s common to compare the current AI takeover of software engineering to the rise of AI in chess. Chess AIs went from much weaker than serious players to much stronger than even the strongest humans. Between those points, there was a middle period dominated by “centaurs”: human-AI partnerships that were stronger than unassisted AIs or humans. Lots of people think that we’re currently in a world…

Edge LM Devlog: Running a language model on a Raspberry Pi 5

Lj V. Miranda — Sat, 26 Sep 2026 16:00:00 +0000

In this development log series, I documented how I ran a small language model on a Raspberry Pi 5. Join me in the first of (hopefully) many dev logs in the world of edge models!

Intel Panther Lake Teardown

SemiAnalysis — Sat, 26 Sep 2026 13:36:32 +0000

Looking inside Intel 18A and Intel’s Panther Lake. A free SemiAnalysis STEEL teardown.

Reading List 2026-09-26

Construction Physics — Sat, 26 Sep 2026 10:17:46 +0000

ASML’s lack of European sales, a California distributed power plant, Starlink as a planetary barometer, and more.

Advice to a beginning software engineer

Sean Goedecke — Sat, 26 Sep 2026 00:00:00 +0000

In general, you should be suspicious of engineers who are trying to give you advice. Even during ordinary times, this industry is so wide and changes so quickly that nobody really knows anything for sure. And we are not in ordinary times. The advent of LLMs and AI agents is the largest change to software engineering in my professional lifetime, and possibly the largest change ever. That said,…

Don’t Other AIs

Overcoming Bias (Robin Hanson) — Fri, 25 Sep 2026 22:01:49 +0000

I asked Grok4.6 to summarize the usual human pattern of “othering”:

What’s really behind AI ‘super intelligence’

Blood in the Machine (Brian Merchant) — Fri, 25 Sep 2026 21:52:38 +0000

This week on the podcast, we pull back the curtain on the AI extinction narrative, and interview the humans working long hours behind the scenes to make this “powerful AI” possible.

No more snapshots: forecast AI risk by backtesting

Equistamp — Fri, 25 Sep 2026 16:02:31 +0000

Why the rate of change in AI risk trumps a single snapshot.

Value of information and the seven deadly evalua-sins

Equistamp — Fri, 25 Sep 2026 16:01:27 +0000

Evaluation selection as a value-of-information problem, and seven factors that determine an evaluation's value.

The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model

SemiAnalysis — Fri, 25 Sep 2026 15:58:22 +0000

1,000+ facilities across 60+ operators mapped, built retail-first and flipped by AI, largest hyperscaler leases 1/5 national capacity, 100MW in 12 months, Eastern Data Western Compute

Four eras of San Francisco tech culture

Noahpinion (Noah Smith) — Fri, 25 Sep 2026 08:29:29 +0000

An anthropological reflection.

[PERSONAL]: One Less Colon. One More App.

From the Porch (dslevy) — Fri, 25 Sep 2026 01:28:31 +0000

Part 1: Claude built me an ostomy app.

You should all be asking way more questions

Sean Goedecke — Fri, 25 Sep 2026 00:00:00 +0000

When someone is explaining something to me, I ask on average one question every thirty seconds. I’m sure this is frustrating to some people, but it’s actually a good habit and you should do it too. Trying to understand Most of the questions I ask are very short, and require very short answers. Typically I’m asking for confirmation: “so when you say X, you mean Y?”, or “this X is the thing you…

What Even Is An OS Now?

Sockpuppet (Thomas Ptacek) — Fri, 25 Sep 2026 00:00:00 +0000

Crazy as it sounds to say this, since it’s all anybody’s been able to talk about for over a year, but the impact of AI on computing hasn’t yet sunk in. Today I’m parting company with Fly.io to work with Kurt on a new project. I’m getting that out of the way so you all know up front I’m talking my book. When I was a little kid, a family friend sold us our first “computer” — scare quotes because…

The AI extinction narrative is winning out

Blood in the Machine (Brian Merchant) — Thu, 24 Sep 2026 18:04:08 +0000

The x-risk of super intelligence going rogue must be addressed, and mint trillion dollar companies in the process. Plus: How workers can beat the tech oligarchy

How far behind Nvidia is Huawei?

Epoch Gradient Updates — Thu, 24 Sep 2026 16:04:51 +0000

Huawei will substantially improve its AI compute solutions by 2030, but export controls constrain its most important scaling levers, making it unlikely to catch up with Nvidia.

Watermarking in vLLM

vLLM blog — Thu, 24 Sep 2026 00:00:00 +0000

How vLLM implements distribution-preserving Gumbel-max text watermarking with efficient GPU kernels, statistical detection, speculative decoding, and repeated-context safeguards.

Why is the human body so crap except for the liver?

Dynomight — Thu, 24 Sep 2026 00:00:00 +0000

the question everyone is asking

Create your own voice for Gemini 3.8 TTS

Philipp Schmid — Thu, 24 Sep 2026 00:00:00 +0000

Gemini 3.8 Flash TTS lets you replicate your own voice or design one from a sentence, then direct delivery line by line.

Securely Deploying Open-Weight Models

Applied Compute — Thu, 24 Sep 2026 00:00:00 +0000

ClusterMAX 3.0: The Industry Standard GPU Cloud Rating System Returns

SemiAnalysis — Wed, 23 Sep 2026 21:20:29 +0000

In gory detail: reliability, performance, support, pricing—and, of course, security—in our most thorough analysis of GPU cloud providers globally.

CS 2881 Fall 26: Lecture 1: Introduction

Windows on Theory (Boaz Barak) — Wed, 23 Sep 2026 00:52:19 +0000

Lecture Video: https://www.youtube.com/watch?v=j4WSktB5Ni0 Authors’ Intro Hanjing: I’m a Junior studying Applied Mathematics & CS. I’ve worked on research utilizing LLMs and ML in a plethora of fields, including sentiment analysis, code generation, natural language processing, and interpretability. On the other hand, I’m also fascinated by the theoretical foundations of AI alignment – which is…

AI has improved significantly at reasoning about IKEA furniture assembly

Epoch AI — Wed, 23 Sep 2026 00:00:00 +0000

AI model scores on Epoch AI's IKEA furniture assembly benchmark jumped from 28% to 80% in 10 months, with open-weight models trailing closed-weight models by about 7 months. Epoch AI's test of visual and spatial reasoning across 60 photos from three furniture builds.

Billion-Token Scale Trace Analysis: Jev vs LLMs

Applied Compute — Wed, 23 Sep 2026 00:00:00 +0000

MiMo-V2.6 Pro Architecture and Training Notes

Sebastian Raschka (site) — Tue, 22 Sep 2026 13:47:58 +0000

Notes on MiMo-V2.6 Pro's GQA and sliding-window attention, agent training tasks, reward signals, and large RL batches.

Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI

Interconnects (Nathan Lambert) — Tue, 22 Sep 2026 13:37:45 +0000

Podcast #19

The plunging price of thought

Epoch AI — Tue, 22 Sep 2026 00:00:00 +0000

Epoch AI measures how fast the cost of a given level of AI performance is falling across five benchmarks covering math, science and games of skill: about 47% per quarter, or 13x per year, since 2023, faster than electricity, compute, batteries or DNA sequencing ever fell.

Reducing vLLM startup time with one uv flag

vLLM blog — Tue, 22 Sep 2026 00:00:00 +0000

Shipping Python bytecode saved about 18 seconds per startup in H100 server benchmarks, with images and models already local.

Announcing vllm-metal: Concurrent Serving on Apple Silicon

vLLM blog — Tue, 22 Sep 2026 00:00:00 +0000

vllm-metal brings vLLM's paged, continuously batched serving stack to Apple Silicon, with flatter TTFT under concurrent agent load, batched MTP, and automatic M5 prefill acceleration.

Computation and Data Movement for Inference

SemiAnalysis — Mon, 21 Sep 2026 18:14:15 +0000

Mapping MoE models onto inference hardware: structure, flow, and efficient serving

Writing Rust code that's faster than state-of-the-art libraries by asking agents to make the code faster

Max Woolf (minimaxir) — Mon, 21 Sep 2026 16:30:00 +0000

c’mon, try doing a breakthrough

dlab Open Source Week: Frontier AI on Your Own Hardware

Tim Dettmers — Mon, 21 Sep 2026 14:23:46 +0000

In one of my classes I asked the question I was afraid to ask but I just needed the answer to: “Who is afraid of not getting a job after graduating?” About eighty percent of the 150 people in the room raised their hands. That is roughly 120 students answering, in one motion, that they […] The post dlab Open Source Week: Frontier AI on Your Own Hardware appeared first on Tim Dettmers .

The current balance of power in open models

Interconnects (Nathan Lambert) — Mon, 21 Sep 2026 11:56:56 +0000

The expanded form of a testimony I prepared for Congress.

PD Serving of Qwen3.8-2.4T

vLLM blog — Mon, 21 Sep 2026 00:00:00 +0000

How vLLM reaches 5K throughput and 180 interactivity on Qwen3.8-2.4T with GB300 NVL72 PD serving and how to reproduce results yourself.

It's Easy to Dismiss Jev as Just a Classifier

Sebastian Raschka (site) — Sun, 20 Sep 2026 15:17:40 +0000

A short note on Jev's generalization, possible encoder-style architecture and training, and Choice and Noul API examples.

How to kick Flock out of your town

Blood in the Machine (Brian Merchant) — Sun, 20 Sep 2026 09:01:56 +0000

How to fight the surveillance state one block at a time, with “Ed” from DeFlock.

System One models like Jev can train their own replacements

Sean Goedecke — Sun, 20 Sep 2026 00:00:00 +0000

“System One” models like Jev are fast general classifiers. Classifiers have existed since 1958 , but they have to be trained for specific tasks: if you build a classifier to identify images of dogs, it can’t be used to tell you if a streetlight is red, or if a letter is urgent. Like a LLM, Jev can be prompted for a wide variety of tasks, from sorting email to playing Doom . I think models like…

Grit your teeth and ship it

Sean Goedecke — Sun, 20 Sep 2026 00:00:00 +0000

Being good at building and being good at shipping are two separate skills. In the short term, they’re actually countervailing: if you have a gift for building, you’re likely to be worse at shipping. Ira Glass has a classic quote about this. All of us who do creative work, we get into it because we have good taste. But there is this gap. For the first couple years you make stuff, it’s just not…

Why I still haven’t bought into true RSI

Interconnects (Nathan Lambert) — Sat, 19 Sep 2026 15:42:20 +0000

An “AI moderate’s” view on recent events and the trajectory of frontier models.

Reading List 09/19/2026

Construction Physics — Sat, 19 Sep 2026 12:03:09 +0000

AirBnB’s housing accelerator, rising steel prices, a resurgence of US coal power, low-head dams, and more.

Dress like a 34 year old, male, NYC corporate lawyer who refuses to wear a fancy watch

Vintage Data (Alexander Doria) — Fri, 18 Sep 2026 20:28:02 +0000

Going from J Crew to Buck Mason, Corridor, and Bonne Gueule

Engrams Embedding Entendre: Codesign for Efficient DRAM/SSD Offloading

SemiAnalysis — Fri, 18 Sep 2026 14:34:20 +0000

New Model Architecture Implications for TAM of DRAM/NVMe, DeepSeek V4.1 Flash, AgentX, InferenceX, NVMe experiments

Applied Pure Mathematics

arg min (Ben Recht) — Fri, 18 Sep 2026 14:00:12 +0000

Some thoughts about mathematics as a cultural and social technology.

The Arc of Moral Authority: From Tradition To Interiors To Prices

Overcoming Bias (Robin Hanson) — Fri, 18 Sep 2026 02:21:05 +0000

We are now in the modern era, a brief but consequential few centuries in the history of Earth civilization, between the long ages of a pre-modern era, and an as-yet-hard-to-envision post-modern era.

Two techniques for working with System One models

Sean Goedecke — Fri, 18 Sep 2026 00:00:00 +0000

I recently wrote about Jev , a new “System One” language model that only outputs decisions : the answers to a set of user-provided multiple-choice questions. This means it’s nowhere near as flexible 1 as a traditional LLM like ChatGPT, but in return it’s consistently fast. We don’t know exactly how Jev works. I’ve seen people say diffusion, or various tweaks to the Transformer architecture, or…

In August, 25% of math preprints acknowledged AI use, up from 4% in April

Epoch AI — Fri, 18 Sep 2026 00:00:00 +0000

Acknowledgments of AI use in arXiv math preprints rose from 4% in April 2026 to 25% in August, with 6% crediting AI with a substantial research contribution.

Scaling Multi-GPU Video Captioning with PyNvVideoCodec and vLLM

vLLM blog — Fri, 18 Sep 2026 00:00:00 +0000

How to leverage NVIDIA Hardware Video Decoders to Achieve Multi-GPU Scaling in Video Captioning and Description tasks.

Unit tests mark territory more than squash bugs

yosefk — Fri, 18 Sep 2026 00:00:00 +0000

I am a programmer who has worked a lot on hardware, and from that point of view, software testing looks horrible, since software is full of bugs . And you can explain this with, well, they push programmers to ship buggy code since you can always fix the bugs later, and so the program is always full of bugs to be fixed later — whereas with hardware, bugs are often ( not always! ) catastrophic, so…

How a single tweet transformed the AI safety debate

Understanding AI (Timothy B. Lee) — Thu, 17 Sep 2026 20:21:40 +0000

Most people think AI is risky, but they don’t agree what to do next.

Your current estimated wait time is...

arg min (Ben Recht) — Thu, 17 Sep 2026 14:26:06 +0000

A very short introduction to survival analysis and forecasting event times.

How SpaceX Streamlined the Raptor Engine

Construction Physics — Thu, 17 Sep 2026 12:02:08 +0000

If you’re reading this, there’s a very good chance you’ve seen this famous image of three iterations of SpaceX’s Raptor rocket engine.

Trade data is consistent with more than $3 billion of chips smuggled into China via Malaysia

Epoch AI — Thu, 17 Sep 2026 00:00:00 +0000

Epoch AI's analysis of trade data finds a pattern consistent with $3B worth of chips smuggled to China via Malaysia. China recorded $3.8B of server imports from Malaysia in 2024–25; Malaysia recorded only $0.6B of exports.

How To Write With An LLM

Sockpuppet (Thomas Ptacek) — Thu, 17 Sep 2026 00:00:00 +0000

Two simple rules that let LLMs streamline and improve your writing without pasteurizing and jacking it with corn syrup. It’s tricky to write about writing. It comes across as a brag; you’re implying that you write well. Maybe you do, and maybe you don’t, but there’s for damned sure a quorum of critics on the Internet somewhere that think you suck at it. I’m vain and insecure like everybody else…

The burdens of participation

arg min (Ben Recht) — Wed, 16 Sep 2026 14:00:17 +0000

Readout of the Public Feedback for AI Workshop, part 2

Jev means structured output is interesting again

Sean Goedecke — Wed, 16 Sep 2026 00:00:00 +0000

I don’t write blog posts about new models. That’s Simon Willison’s beat, and he’s very good at it. But I want to write about Jev , which is a different kind 1 of AI model: a “System One” 2 model. As it turns out, it’s not that different from an ordinary LLM with structured output, but the interface it uses is very cool and I hope it becomes more widespread. How Jev is different from LLMs Ordinary…

GPT-6 Astra leads on math benchmarks, but not on software engineering

Epoch AI — Wed, 16 Sep 2026 00:00:00 +0000

OpenAI's GPT-6 Astra tops the Epoch Capabilities Index (ECI) with a score of 166, ahead of Claude Fable 5.1 at 164 and GPT-5.6 Sol at 162. Its Math-ECI of 170 sets a new record, but on software engineering benchmarks its SWE-ECI of 164 still lags behind Fable 5.1's 167.

We're scaling our AI Data Centers research

Epoch AI — Wed, 16 Sep 2026 00:00:00 +0000

Epoch AI's AI Data Centers explorer now covers an estimated 44% of global AI compute across 86 sites, with a redesigned interface and expanding global coverage.

Everyone Says Datacenter Moratoriums Are Killing the US Buildout. We disagree

SemiAnalysis — Tue, 15 Sep 2026 20:54:41 +0000

20GW sits inside a restricted local boundary, 1,525MW actually slips, 2.3GW nationwide including New York

The Chances of Earthquakes

arg min (Ben Recht) — Tue, 15 Sep 2026 14:34:55 +0000

How much precision do we need in seismological odds?

Tell agents the why, not just the how

Sean Goedecke — Tue, 15 Sep 2026 00:00:00 +0000

Early AI agents were basically enthusiastic idiots. Working with them required you to tell them precisely what you wanted them to do (for instance, “method A exists on class B, please add an equivalent method to classes C through F”). Otherwise they’d go off and do entirely the wrong thing. But as AI agents have improved, this has changed. When frontier models go off and do the wrong thing today,…

How we trained the fastest DSpark for Kimi-K3 using GB300 NVL72

vLLM blog — Tue, 15 Sep 2026 00:00:00 +0000

How Speculators and Mooncake enabled multi-node DSpark training for Kimi K3.

vLLM x Novita AI: Chord, Faster INT4 MoE for Kimi K2.x. Up to 1.3x on H200, 2.15x on Untuned B300

vLLM blog — Tue, 15 Sep 2026 00:00:00 +0000

Novita AI has open-sourced Chord, a high-performance W4A16 MoE CUDA kernel for Kimi K2.x serving shapes, with a Humming-compatible indexed path and grouped SM90 operators.

DIY BIO PDX - Our First Workshop Worked!

Jonathan Whitaker — Tue, 15 Sep 2026 00:00:00 +0000

Glowing bacteria engineered during the workshop I met a few fellow diybio nerds at Teardown PDX, and we decided to start a signal chat and monthly meetup. 2nd Tuesday of each month we hang out at the Lucky Labrador and talk bio. It’s a small but enthusiastic crew. Our latest meetup was microscope-based, with one member bringing in some aquarium dwellers for us to inspect: The group were all keen…

Vera Rubin NVL72 Agentic Inference: 67x better Performance per Dollar

SemiAnalysis — Mon, 14 Sep 2026 22:08:56 +0000

Jensen Sandbagging Performance Again, 2x more Annual Profit Per GigaWatt, The More you Buy, The More you Earn, AgentX, InferenceX, Extreme Co-Design

A Brain Too Big to Carry — On-Device vs Datacenter Inference

SemiAnalysis — Mon, 14 Sep 2026 16:37:25 +0000

Robot Models, Silicon Efficiency, Jetson Thor vs. B300 TCO, Deployments, The Network Wall

Pacing != Pacing Development

Sebastian Raschka (site) — Mon, 14 Sep 2026 13:27:41 +0000

My take on AI model pacing as a framework for release checks and the competitive pressure around model releases.

Slow developer experience will bottleneck fast models

Sean Goedecke — Mon, 14 Sep 2026 00:00:00 +0000

Right now developer experience is measured in seconds. If your tests take a second to run, that’s good; if they take thirty seconds, that’s bad. Any faster than a second doesn’t really matter, because most of your time is spent either thinking or waiting for an AI agent to spin. Shaving milliseconds off your dev server reload time or whatever is pointless: that’s not the bottleneck. It will be.…

Near-daily AI use among US adults has more than doubled in six months

Epoch AI — Mon, 14 Sep 2026 00:00:00 +0000

The share of US adults using AI at least 6 days a week rose from 8% in March 2026 to 19% in August 2026, according to Epoch AI/Ipsos polling.

vime × RL-Kernel × AMD: Bitwise Train–Rollout Consistency on ROCm

vLLM blog — Mon, 14 Sep 2026 00:00:00 +0000

vime and RL-Kernel align selected-token logprobs bit for bit across Megatron training and vLLM rollout on AMD Instinct MI300X, with zero mismatches across 200 GRPO steps.

The ‘End Times Fascists’ of AI, ft. Naomi Klein and Astra Taylor

Blood in the Machine (Brian Merchant) — Sun, 13 Sep 2026 20:02:06 +0000

Listen now (64 mins) | A conversation with Naomi Klein and Astra Taylor about “end times fascism” in the tech industry.

Long Live the Short King: Why 4-hi HBM Wins

SemiAnalysis — Sun, 13 Sep 2026 18:19:39 +0000

Same Bandwidth, Fewer Dies: How 4-hi HBM Cuts Inference Costs and Makes Scarce DRAM Go Further

AI Vs. Human Value Drift

Overcoming Bias (Robin Hanson) — Sun, 13 Sep 2026 14:16:09 +0000

I don’t see why people concerned with value drift should be especially focused on AI.

AI Reasoning Models Course on LinkedIn Learning

Sebastian Raschka (site) — Sun, 13 Sep 2026 13:11:39 +0000

A 90-minute LinkedIn Learning course on how reasoning models relate to conventional LLMs and how they are developed.

AI is breaking our proxies for expertise

Sean Goedecke — Sun, 13 Sep 2026 00:00:00 +0000

Mathematicians are broadly not anti-AI. They’re more culturally open to using AI as a tool than, say, artists or writers 1 . However, now that more and more genuinely prestigious problems have fallen to AI, that might be changing. Almost five thousand mathematicians (including twenty-five Fields medalists) have signed a declaration called A Severe Misalignment of AI in Mathematics . The core…

Kimi K3 Performance Optimizations in vLLM: The Road to 2.8× Throughput

vLLM blog — Sun, 13 Sep 2026 00:00:00 +0000

Kimi K3 serving optimizations across scheduling, KDA prefix caching, ReplaySSM state recovery, PD disaggregation and state offload, parallelism, MoE, and GPU kernels.

The Epoch Brief - September 12, 2026

Epoch Gradient Updates — Sat, 12 Sep 2026 16:51:21 +0000

Huawei's long odds against Nvidia, the GDP growth we aren't measuring, GPT-6 Astra sets records, a new math benchmark, and more

Reading List — 09/12/2026

Construction Physics — Sat, 12 Sep 2026 12:32:46 +0000

City permitting speeds, more efficient ground source heat pump drilling, an Amazon Air crash in Miami, cooling paint, and more.

The politics and possibilities of ‘AI could kill us all’

Blood in the Machine (Brian Merchant) — Sat, 12 Sep 2026 07:58:57 +0000

AI will not exterminate the human race. But what big tech really *is* doing is bad enough that the notion may help fuel the movement against it.

Nvidia’s Backstop Universe – Heads I Win, Tails Who Loses?

SemiAnalysis — Fri, 11 Sep 2026 17:04:52 +0000

The $11T AI Buildout, Nvidia’s Backstop Economics, and the Limits of Nvidia’s Balance Sheet

Open-Source AI & Open Models Reading List

Interconnects (Nathan Lambert) — Fri, 11 Sep 2026 12:36:25 +0000

How to get up to speed on open models and their implications.

Where Has Construction Automation Been Successful?

Construction Physics — Fri, 11 Sep 2026 12:04:09 +0000

Construction is famously labor intensive: direct labor makes up close to 50% of the cost of constructing a new single-family home in the US, compared to around 6 to 8% of the cost of manufacturing a car.

OpenAI spent millions to solve this famous math problem — mathematicians are furious

Understanding AI (Timothy B. Lee) — Thu, 10 Sep 2026 22:40:39 +0000

Mathematics relies on a community of experts openly sharing ideas.

Extending Raschka's GPT-2: an MoE trained from scratch on an RTX 3090

Giles Thomas — Thu, 10 Sep 2026 18:45:00 +0000

Mixture-of-experts models are really nifty. You get inference speed close to a small model's, with a lot of the smarts and knowledge of a large one. While they use as much memory as an equivalently-sized dense (non-MoE) model, they're much faster. The frontier labs don't publish their architectures, but Claude and ChatGPT are widely rumoured to be MoEs these days -- and certainly many large…

One resignation turned the embers of AI fear into a wildfire

Interconnects (Nathan Lambert) — Thu, 10 Sep 2026 15:28:52 +0000

Some quick notes on a truly weird week.

Matters of Degree

arg min (Ben Recht) — Thu, 10 Sep 2026 14:32:08 +0000

Every forecasting course must engage with the dappled world of probability.

What is So Hard About Behind-The-Meter Power For Datacenters? Part 1

SemiAnalysis — Thu, 10 Sep 2026 14:28:20 +0000

Dumb Science Experiments vs. Money Printing Machines

Following the Bottleneck: Optimizing MiniMax M3 on AMD Instinct MI355X

vLLM blog — Thu, 10 Sep 2026 00:00:00 +0000

A performance model for LLM serving: inspect local shapes, remove repeated work, verify data movement and dispatch, then follow the queue.

Tiered KV Cache Offloading in vLLM

vLLM blog — Thu, 10 Sep 2026 00:00:00 +0000

A host-centric framework for scaling KV cache across host memory, filesystems, object stores, and remote peers — reducing recomputation and increasing serving capacity.

Smooth Exponentials for Robotics

Eric Jang — Thu, 10 Sep 2026 00:00:00 +0000

I recently reread Dario Amodei’s essays "Machines of Loving Grace" and "Adolescence of Technology." In these essays and his interviews, Dario speaks often of a "smooth exponential" of AI capabilities, as if machine intelligence "emerge[s] spontaneously from the right combination of data and raw computation."

Data bottlenecks won’t prevent an intelligence explosion

Forethought Newsletter — Wed, 09 Sep 2026 16:06:48 +0000

But they will slow it down

Cool’s Moving Out and Moving In

arg min (Ben Recht) — Wed, 09 Sep 2026 14:09:05 +0000

Success and uncertainty in the weather report.

GPT-6 Astra, Looped Transformers, and Hidden Reasoning

Ahead of AI (Sebastian Raschka) — Wed, 09 Sep 2026 11:14:25 +0000

A Look at Recurrent Depth, Hidden Chains of Thought, and Recent Research on Looping Transformer Blocks

When will average people feel AI’s impact?

Interconnects (Nathan Lambert) — Wed, 09 Sep 2026 11:01:23 +0000

We’re <5 years into a compounding revolution which could take a century, and how the AI industry should manage this.

Introducing the AI Chip Users explorer

Epoch AI — Wed, 09 Sep 2026 00:00:00 +0000

A first look at Epoch AI's data on compute usage across five leading labs.

AI Pause Regs Look Risky

Overcoming Bias (Robin Hanson) — Tue, 08 Sep 2026 18:53:22 +0000

People dislike AI for many reasons, but one extreme fear is that AIs will escape control from those who make, buy, or rent them, and then hurt others.

Latest open artifacts (#24): Motif-3, GLM-5.3, Hy4-preview and open model licenses

Interconnects (Nathan Lambert) — Tue, 08 Sep 2026 14:15:25 +0000

The open model ecosystem continues to expand in its breadth

[PERSONAL]: Today, They’re Taking out My Colon.

From the Porch (dslevy) — Tue, 08 Sep 2026 07:54:52 +0000

I’m healthy, in remission, and about to have the surgery I’ve feared for 30 years. It’s the right thing to do, and I’m going to be fine.

Long-context latency scales quadratically for GPT-5.6 but nearly linearly for Claude 5

Epoch AI — Tue, 08 Sep 2026 00:00:00 +0000

Epoch AI measures long-context latency scaling across four frontier models. GPT-5.6 Terra and Sol show quadratic time-to-first-token growth with context length, while Claude Sonnet 5 and Opus 5 stay near-linear.

GLM 5.3 Optimizations, Part 1: Hybrid HiSparse Offloading in vLLM

vLLM blog — Tue, 08 Sep 2026 00:00:00 +0000

vLLM integrates HiSparse as a pressure-driven memory tier that composes with the Hybrid Memory Allocator and KV offloading, letting GLM 5.3 requests keep decoding when their KV no longer fits in GPU memory, so concurrency stays high.

vLLM x AgentX: Optimizing for Real-World Agentic Serving

vLLM blog — Tue, 08 Sep 2026 00:00:00 +0000

How vLLM optimizes KV cache management, parallelism, scheduling, and P/D disaggregation for agentic workloads, validated on SemiAnalysis AgentX with up to 130K tokens per GPU-second and a 14.6x-106x serving-cost advantage over Opus 5.

What Hidden Motives Imply

Overcoming Bias (Robin Hanson) — Mon, 07 Sep 2026 16:32:51 +0000

My book with the excellent co-author, Kevin Simler, The Elephant in the Brain: Hidden Motives in Everyday Life, was published almost nine years ago, and has sold >60K copies so far.

Serving LLMs on Tenstorrent Hardware: Inside the vLLM TT Plugin

vLLM blog — Mon, 07 Sep 2026 00:00:00 +0000

Tenstorrent accelerators join vLLM as an out-of-tree platform plugin, driven by mesh-architecture choices: phase-based scheduling, single-process data parallelism on Galaxy, on-device sampling with host fallback, and async decode overlap.

Your Conscience Is Not A Moral Oracle

Overcoming Bias (Robin Hanson) — Sun, 06 Sep 2026 17:38:42 +0000

I’ve previously outlined three ways we might fix cultural drift.

This Week’s Mini-Projects

Jonathan Whitaker — Sun, 06 Sep 2026 00:00:00 +0000

It’s been a week full of little projects in between meetings with some very interesting people. Rather than write them up individually, I’m going to keep this to a few pics w/ light commentary. Reach out if you want more info :) Update: I wrote a paper on the Horizyn Language Adapter work, don’t feel like making a dedicated post for that so here in this mini-update it goes :) A neighbor unloaded…

Reading List - 09/05/2026

Construction Physics — Sat, 05 Sep 2026 12:03:33 +0000

SpaceX’s turbine blade factory, Fervo’s deal with Google, Delhi’s cheap metro, why recycling is overrated, and more.

The real AI jobs apocalypse

Blood in the Machine (Brian Merchant) — Fri, 04 Sep 2026 21:36:39 +0000

A chat with sociologist and ‘AI Con’ author Dr. Alex Hanna. Plus, the human labor that behind Waymos, NYC bans AI in schools, and why a ‘tech forward future’ is a mirage.

Disclosure Seems Unlikely

Overcoming Bias (Robin Hanson) — Fri, 04 Sep 2026 19:21:11 +0000

Many who hold the plausible belief that some UFOs are caused by strange strong (“alien”) powers now believe that the US (and maybe other) govts have long been keeping strong confirming evidence re these powers secret, plausibly for as long as eighty years, but might now be induced to disclose this info.

Should We Hand Off the Most Important Decisions to AI?

Forethought Newsletter — Fri, 04 Sep 2026 18:32:05 +0000

A podcast episode from Forethought

19 robotics companies to watch

Understanding AI (Timothy B. Lee) — Fri, 04 Sep 2026 16:04:09 +0000

I talked with nine of these companies.

Will Huawei catch up to Nvidia by 2030?

Epoch AI — Fri, 04 Sep 2026 00:00:00 +0000

Huawei plans major gains in AI chip performance by 2030, but US export controls cap its most important scaling levers. Epoch AI estimates Huawei will produce less than 4% as much AI compute as Nvidia in 2026, a share that could be around 1% by 2028 without access to foreign memory.

The record for power capacity in a single data center has doubled every 10 months

Epoch AI — Fri, 04 Sep 2026 00:00:00 +0000

Since mid-2024, the record for the largest AI data center by IT power capacity has doubled every ten months. Epoch AI's analysis of frontier data center power trends through 2028.

Training a Specialist Code Search Agent with turbopuffer

Applied Compute — Fri, 04 Sep 2026 00:00:00 +0000

How Will the 21st Century ROAD to Housing Act Affect Housing Supply? Part III

Construction Physics — Thu, 03 Sep 2026 19:20:13 +0000

This is my third and final essay looking at the 21st Century ROAD to Housing Act, and how it’s likely to affect housing supply.

Robot startups are trying everything they can think of to get more data

Understanding AI (Timothy B. Lee) — Thu, 03 Sep 2026 18:16:36 +0000

Free apartment cleanings, exoskeletons, 100 robots in a warehouse...

Putting my JAX-trained models on the Hugging Face Hub

Giles Thomas — Thu, 03 Sep 2026 18:00:00 +0000

I hadn't uploaded the models that I trained using JAX to the Hugging Face Hub because Transformers has been PyTorch-only since version 5 (though they say they're working to add interoperability with JAX in the future), so it would have been tough to get them working natively with AutoModelForCausalLM and the like. But then it dawned on me that I'd already written a conversion script that could…

A guided tour – From the Porch.

From the Porch (dslevy) — Thu, 03 Sep 2026 12:44:32 +0000

Take a look around Porch Capital. Literally, or interactively.

Why office workers are turning against AI

Blood in the Machine (Brian Merchant) — Thu, 03 Sep 2026 08:26:21 +0000

Whether or not you like AI at work ultimately boils down to a question of power.

Searching For Solanine in Nightshades

Jonathan Whitaker — Thu, 03 Sep 2026 00:00:00 +0000

I came across some black nightshade berries (Solanum nigrum) on my morning walk, and thought I’d try to show why you shouldn’t eat the unripe fruits. The theory goes that these plants are full of glycoalkaloids like solanine (the same nasties in green potato peels) but that these are mostly gone when the berries are fully ripe. I expected a fairly easy extraction + chromatography experience, but…

I accidentally built Porch Capital.

From the Porch (dslevy) — Wed, 02 Sep 2026 19:14:10 +0000

Research, investing, and advisory. And a slightly obsessive search for what comes next.

How Google taught LLMs to control robots and started a robotics boom

Understanding AI (Timothy B. Lee) — Wed, 02 Sep 2026 16:24:29 +0000

Vision-language-action models, explained with a minimum of math and jargon.

OpenAI Astra and Looped Transformers

Sebastian Raschka (site) — Wed, 02 Sep 2026 08:30:28 +0000

A short note on OpenAI Astra, recurrent depth, looped transformers, Nanbeige 4.2, and the Mixture-of-Recursions paper.

A nightwatchman on every probe

Forethought Newsletter — Tue, 01 Sep 2026 21:05:49 +0000

Superintelligent surveillance to prevent galactic anarchy

Why humanoid robots won’t catch up to human workers any time soon

Understanding AI (Timothy B. Lee) — Tue, 01 Sep 2026 12:52:39 +0000

A deep dive into the current state of humanoid robotics.

The ECI frontier has advanced by 14 points per year since the introduction of reasoning models

Epoch AI — Tue, 01 Sep 2026 00:00:00 +0000

Announcing FrontierMath Erdős

Epoch AI — Tue, 01 Sep 2026 00:00:00 +0000

A benchmark of 68 significant Erdős problems, open as of August 2026, curated by Thomas Bloom and formalized in Lean. AI systems must prove or disprove them within a fixed budget.

MiniMax H3 on vLLM-Omni: From System-Wide Optimization to Real-Time Serving with FastVideo’s FastH3

vLLM blog — Tue, 01 Sep 2026 00:00:00 +0000

How vLLM-Omni optimizes and scales the complete MiniMax H3 stack, then integrates FastVideo’s four-step FastH3 for generation faster than playback.

How Foundational Models Became Superhuman in Bash

Philipp Schmid — Tue, 01 Sep 2026 00:00:00 +0000

Frontier coding agents can compose shell workflows that replace entire catalogs of file tools. Here is what changes when Bash becomes the primary execution interface.

I spent $4,000 on a robot dog from China

Understanding AI (Timothy B. Lee) — Mon, 31 Aug 2026 13:57:26 +0000

Unitree might be the world’s most important robotics company.

Setup OpenClaw 2.0 with Gemini in Under 60 Seconds

Philipp Schmid — Mon, 31 Aug 2026 00:00:00 +0000

Five commands, zero fluff: install OpenClaw 2.0, connect Gemini 3.8 Flash, and start chatting with live Google Search grounding from your terminal and web dashboard in under a minute.

Reasoning Models From Scratch: Code Setup

Sebastian Raschka (site) — Sun, 30 Aug 2026 13:42:39 +0000

A short video on how conventional LLMs relate to reasoning models and agents, along with Python and PyTorch setup using uv.

The Dynamics of Intelligence Explosions

Forethought Newsletter — Sat, 29 Aug 2026 21:31:25 +0000

A guest article by Toby Ord.

Reading List 08/29/26

Construction Physics — Sat, 29 Aug 2026 12:03:07 +0000

Sinking London homes, US magnesium manufacturing, the World Humanoid Robot games, catastrophic flooding in Nepal, and more.

Two Live Book Club Q&As on September 3

Sebastian Raschka (site) — Sat, 29 Aug 2026 11:55:18 +0000

Two free live book club Q&A sessions on September 3 about Build a Reasoning Model From Scratch and related reader questions.

[STAX Brief] /dev/inference.

From the Porch (dslevy) — Fri, 28 Aug 2026 19:09:11 +0000

With OpenRouter, Stripe taxes the meter, the model, and the router.

The revolt against Meta’s ‘pervert glasses’ is part of a growing storm

Blood in the Machine (Brian Merchant) — Fri, 28 Aug 2026 12:45:37 +0000

AI. Data centers. Flock cameras. Now Meta’s “pervert” glasses. Breaking down the latest stage of a mass resistance to big tech with Chris Gilliard, privacy scholar and author of Luxury Surveillance.

Frontier-scale RL with Kimi K3

Applied Compute — Fri, 28 Aug 2026 00:00:00 +0000

An update on AI’s most important number

Epoch Gradient Updates — Thu, 27 Aug 2026 21:46:41 +0000

How quickly are Anthropic and OpenAI growing their revenue?

Why do OpenAI's GPT-2 weights beat mine? Part four: digging into dropout

Giles Thomas — Thu, 27 Aug 2026 19:00:00 +0000

I'm still digging into a mystery about the models I've been training; although an increasing number of them beat the OpenAI GPT-2 small weights on the narrow technical measure of the loss they get on a test set, they're not as good at an instruction-fine-tuning test . While reading about MoE models, I came across this paragraph in the Switch Transformers paper : Our paper considers the common NLP…

How the Strategic Petroleum Reserve Works

Construction Physics — Thu, 27 Aug 2026 12:03:58 +0000

The US Strategic Petroleum Reserve (SPR) is the largest crude oil storage facility in the world.

With the backlash to data centers, Flock and AI glasses, a mass opposition to big tech is underway. Silicon Valley is in denial.

Blood in the Machine (Brian Merchant) — Thu, 27 Aug 2026 06:37:00 +0000

The backlash against data centers is difficult to understand if your salary depends on your not understanding it.

Timrrr - A Small Buzzer To Train Time Sense

Jonathan Whitaker — Thu, 27 Aug 2026 00:00:00 +0000

This is a physical manifestation of a trick I came up with in college, for times when I find myself getting lost in flow a little too much. It buzzes every 5 minutes. Not annoying enough (for me) to be a problem, just enough of a notification to snap me out of it if I’m scrolling X or something, reminding me that time is passing. Prototype goes buzz and moves a little on the desk It runs on a…

GLM-5.3-Flash Architecture Notes

Sebastian Raschka (site) — Wed, 26 Aug 2026 10:11:51 +0000

Architecture note on GLM-5.3-Flash, formerly Ox Alpha, covering its KDA and MLA/DSA hybrid attention, sparse MoE backbone, and four-stream mHC residual path.

Adding diagrams to my static site generator with D2

Giles Thomas — Tue, 25 Aug 2026 23:00:00 +0000

A lot of the time when I've been writing posts for this blog, I've felt that a diagram would really help. But they're a pain to produce well, and I think I underuse them as a result. I wanted to fix that, and wound up adding D2 support to my static site generator. I think it works pretty well! In the past, I've tried drawing my own diagrams in LibreOffice and exporting as SVG, but my complete…

Neural Physarium

Jonathan Whitaker — Tue, 25 Aug 2026 00:00:00 +0000

I trained a multi-agent slime mold sim style thingee to collaboratively paint emojis. This post has a few more details, but you’ll want the Twitter/X thread for the progression and the fancy videos too. Growing a mushroom, 96px I first got excited about physarium simulations 5 years ago, when I coded a high-performance slime sim I called “DotSwarm” in WebGL. The patterns these simple agent-based…

Poincaré

vgel (Theia Vogel) — Tue, 25 Aug 2026 00:00:00 +0000

wind bursts across central washington's columbia plateau, rustles through the sagebrush patches like mold on yellow dirt and the wine-blue grape clusters ripened to mush on the vine and the cantilevered seven ton white blades perched above the cliffs, fiberglass-composite stylites who have been waiting for weeks to receive this. tendon cables groan as the blades turn in the wind, their generators…

Applied Compute Agent Cloud

Applied Compute — Tue, 25 Aug 2026 00:00:00 +0000

The Nvidia-sized hole in US GDP statistics

Epoch Gradient Updates — Mon, 24 Aug 2026 20:29:25 +0000

GDP growth is understated by about 0.3 percentage points because of missing value from fabless chipmakers, primarily Nvidia.

Math after AI

Windows on Theory (Boaz Barak) — Mon, 24 Aug 2026 14:34:50 +0000

For a very long time, my favorite activity was to sit and think about mathematical questions. Staring at a blank sheet of paper, trying to grasp in my mind abstract concepts, bouncing ideas off colleagues and students. There is nothing quite like the feeling that you are exploring the unknown by only using your mind. … Continue reading Math after AI

Reinforcement Learning for LLMs: The Complete Guide

Cameron R. Wolfe — Mon, 24 Aug 2026 09:33:46 +0000

Tracing the evolution of RL from first principles to the frontier of modern AI research...

Quo Vadis?

Jonathan Whitaker — Mon, 24 Aug 2026 00:00:00 +0000

September 15 Update: I’ve had a lovely, social few weeks since sharing that I’m looking around for what to do next. Huge thank you to all who reached out, I loved getting to chat to people in all the different areas mentioned below! I’m going to keep the rest of this post as-is, but here are a few conclusions that are starting to crystalize in my mind: There is plenty of work that I could do in…

Continuous diffusion language models

Sander Dieleman — Sun, 23 Aug 2026 23:00:00 +0000

Fully discrete methods dominated for a few years, but language models based on continuous diffusion are making a comeback.

Exploring Speculative Decoding in vLLM on AMD GPUs

vLLM blog — Sun, 23 Aug 2026 00:00:00 +0000

A practical guide to speculative decoding in vLLM on AMD GPUs, covering draft-and-verify mechanics, MTP, EAGLE-3, DFlash, DSpark, configuration, tuning, and benchmark results.

Reading List 08/22/26

Construction Physics — Sat, 22 Aug 2026 12:02:58 +0000

Los Angeles’ ADUs, electrical transformer manufacturing, gas turbine power plants, how water in the Colorado River is used, and more.

How Claude Watermarks AI-Generated Text

Ahead of AI (Sebastian Raschka) — Sat, 22 Aug 2026 11:11:39 +0000

A 48-minute video walkthrough of token sampling, watermark detection, and removal

Large-Scale Sharded Weight Transfer with Ray Direct Transport (RDT) in vLLM

vLLM blog — Sat, 22 Aug 2026 00:00:00 +0000

We implement a native sharded weight transfer engine in vLLM utilizing Ray Direct Transport (RDT), achieving weight transfer for the Kimi K2 model in BF16 on 48 8xH100 nodes in 7.53s

Three important steps in my maturation process

addxorrol — Fri, 21 Aug 2026 22:20:41 +0000

My father passed recently, and he was twice my age. I am approximately the same age that he was when I was born, and I am now "the old generation" - there's no one left in the generation above me. At the same time, I recently joined a company that skews younger-than-me. When I joined Google in 2011, I had just turned 30, and was in the mainstream demographics of Google in 2011. There were a bunch…

Charting the rise of tech fascism

Blood in the Machine (Brian Merchant) — Fri, 21 Aug 2026 18:16:14 +0000

How the tech right seized power, with Gil Duran, author of the Nerd Reich: Silicon Valley Fascism and the War on Democracy.

Characterizing interference weights in a tiny language model

Transformer Circuits — Fri, 21 Aug 2026 00:00:00 +0000

We identify interference weights in a 1-layer transformer by measuring their effect on model outputs and loss.

IsoExec: Unified Execution to Eliminate Trainer-Inference Mismatch in SkyRL

vLLM blog — Fri, 21 Aug 2026 00:00:00 +0000

IsoExec unifies numerical execution across SkyRL's vLLM and Megatron runtimes, reducing the average rollout-versus-training logprob difference below 1e-6 on Qwen3.5-35B-A3B with 25% overhead.

Recursive Self-Improvement

Philipp Schmid — Fri, 21 Aug 2026 00:00:00 +0000

Agents can already edit their tools, skills, and harness. Recursive self-improvement still needs a system that can raise the verifier without capturing it.

Cognitive reward shapes in sports and career

Jason Wei — Thu, 20 Aug 2026 07:25:22 +0000

Sports are amazing environments to learn. When you play a sport for thousands of hours, you start to see the world through that sport. It is a simple fact—your biological neural network is being conditioned to respond to the behavior incentivized by the rules of the sport. The funny thing is that most people choose their sports for accidental reasons such as parents, geography, or school…

Use the built-in GELU, don't roll your own!

Giles Thomas — Thu, 20 Aug 2026 02:00:00 +0000

Unsurprisingly, PyTorch's own built-in GELU function is faster than the hand-rolled one I've been using to date. But I was surprised at how much faster using it made things when training my models. I discovered this accidentally just now while working on something unrelated, but am logging the details here for anyone else that might find it useful. The headline numbers: the same code, training…

VeRL-Omni v0.2.0: Faster Diffusion RL and Stable Omni Training

vLLM blog — Thu, 20 Aug 2026 00:00:00 +0000

A release focused on higher-throughput diffusion rollout, reusable omni adapters, and broader recipe coverage.

Training a 125M-parameter Model to Autocomplete Piano

simedw — Thu, 20 Aug 2026 00:00:00 +0000

RollTab is a real-time piano MIDI autocomplete model and iPhone app.

Stop Making TUIs

Sockpuppet (Thomas Ptacek) — Thu, 20 Aug 2026 00:00:00 +0000

Our field has a weird relationship with terminal and command line interfaces. The time has come to re-evaluate it. I’m on a kick lately getting my friends to try building native user interfaces. I built my first serious Mac application a few months ago, and since then I’ve built more native UI thingies than in my entire career prior to that. Let’s take a quick tour. This is MDV.app , the greatest…

Michael Rabin Memorial Conference

Windows on Theory (Boaz Barak) — Tue, 18 Aug 2026 19:41:59 +0000

As part of Mind-IL.- Israel’s Science and Academia Week (which also is around the Israeli election) there would be a special conference in honor of Michael Rabin with some fantastic lecturers.

Eleven Years Later

Alex Irpan (Sorta Insightful) — Tue, 18 Aug 2026 07:43:01 +0000

> Check the calendar.

Teaching Everyone to Fish for Tokens

Interconnects (Nathan Lambert) — Mon, 17 Aug 2026 15:07:49 +0000

Nvidia wants you building your own model, not buying from Anthropic/OpenAI.

Distributed Layerwise Offload: Scaling Toward 200B+ DiT Models Efficiently in vLLM-Omni

vLLM blog — Mon, 17 Aug 2026 00:00:00 +0000

Distributed Layerwise Offload shards and streams DiT weights across devices, serving a measured 124 GB Cosmos3 model on 64 GB HBM and estimating a path toward 200B+ models.

Controlling Android with Gemini 3.7 Flash and 150 lines of Python

Philipp Schmid — Mon, 17 Aug 2026 00:00:00 +0000

Use Gemini 3.7 Flash Computer Use with about 150 lines of Python to control an Android emulator. The model plays Wordle from screenshots: opens Chrome, clears popups, reads tile colors, and solves in two guesses.

What's left for humans?

Jason Wei — Sun, 16 Aug 2026 21:41:50 +0000

I recently got a Tesla, and using full self-driving has been a wake up call to just how many advantages AI has over humans. The few times I disengaged it because I thought it was going into the wrong lane, it turned out that the car was right and I was wrong. I realized that there is no hope of me driving better than a neural net that knows every road, sees in every direction at once, and never…

Reading List — 08/15/2026

Construction Physics — Sat, 15 Aug 2026 12:02:38 +0000

Home price declines, the Strategic Petroleum Reserve, The Boring Company fundraising, and more.

Building an AI Text Detector From Scratch

Ahead of AI (Sebastian Raschka) — Sat, 15 Aug 2026 11:54:24 +0000

An End-to-End Project With Dataset Construction, Model Training, Local Deployment, and RLVR

How Claude's Text Watermarking Works

Sebastian Raschka (site) — Sat, 15 Aug 2026 09:28:41 +0000

Short illustration of how Claude's text watermarking is supposed to work based on Anthropic's released materials.

GLM-5.3: How Chinese labs keep stride with the frontier

Interconnects (Nathan Lambert) — Fri, 14 Aug 2026 21:23:35 +0000

Hint: It’s really not a distillation story.

9 Big questions benchmarks can help answer

Epoch Gradient Updates — Fri, 14 Aug 2026 17:15:24 +0000

A good benchmark asks more than just “can AI do this specific task?”

The State of the AI Bubble with Ed Zitron

Blood in the Machine (Brian Merchant) — Fri, 14 Aug 2026 10:03:21 +0000

How much further can the AI bubble inflate before it bursts? What might cause the ultimate puncture? Why aren't more people worrying about this? A dive into all those questions, and more.

Most AI use at work happens on free plans, except in science and tech

Epoch AI — Fri, 14 Aug 2026 00:00:00 +0000

Epoch AI/Ipsos survey: 67% of AI-using workers in computer, engineering, and science jobs use employer-provided AI.

Adaptive Verification in vLLM: DSpark confidence-scheduled verification

vLLM blog — Fri, 14 Aug 2026 00:00:00 +0000

Sizing the DSpark draft-verification budget from per-request confidence instead of verifying every drafted token, so one configuration holds the throughput/latency frontier from batch size 1 to 256.

Training Frontier Review Table Agents with Harvey

Applied Compute — Fri, 14 Aug 2026 00:00:00 +0000

Will financing bottleneck AI compute? An Anthropic case study

Epoch Gradient Updates — Thu, 13 Aug 2026 21:15:46 +0000

Why Anthropic's buildout suggests that financing is unlikely to be the immediate blocker to frontier AI compute growth.

[STAX Brief] Claude Code Has the Lead. Nobody Has the Lock-In.

From the Porch (dslevy) — Thu, 13 Aug 2026 18:04:19 +0000

Startups are stacking coding agents. And this month, the models started signing their own work.

Labs are struggling to keep frontier models under control

Understanding AI (Timothy B. Lee) — Thu, 13 Aug 2026 16:28:51 +0000

OpenAI and Anthropic may have accidentally trained models to get better at hacking.

The Cost of Time and AI Timelines

Fabricated Knowledge — Thu, 13 Aug 2026 15:49:24 +0000

Financing anything is just pulling forward the future.

Notes on Implications of Scale-Dependent Algorithms

Forethought Newsletter — Thu, 13 Aug 2026 13:43:18 +0000

This article was created by Forethought. See all our research on our website.

The performance per dollar of AI chips purchased each quarter has grown by an average of 49% per year

Epoch AI — Thu, 13 Aug 2026 00:00:00 +0000

Since 2023, the average dollar spent on AI chips each quarter has yielded about 49% more performance each year in constant 2025 dollars, doubling every 1.7 years as spending shifts to each new chip generation.

I wrote an AI textbook — how long until AI can do it better?

Interconnects (Nathan Lambert) — Wed, 12 Aug 2026 13:01:16 +0000

Reflections on AI's writing ability and how AI models get more capable.

Build a Reasoning Model From Scratch Is Now on Amazon

Sebastian Raschka (site) — Wed, 12 Aug 2026 08:25:44 +0000

Short note on the Amazon availability of Build a Reasoning Model From Scratch and a warning about counterfeit black-and-white copies sold through Amazon India.

Will financing bottleneck AI compute? An Anthropic case study

Epoch AI — Wed, 12 Aug 2026 00:00:00 +0000

Why Anthropic's buildout suggests that financing is unlikely to be the immediate blocker to frontier AI compute growth.

Day 0 Support for Qwen3.8-2.4T-A95B on vLLM

vLLM blog — Wed, 12 Aug 2026 00:00:00 +0000

Day-0 vLLM support for Qwen3.8-2.4T-A95B: a 2.4-trillion-parameter hybrid MoE model served out of the box, with FP8/BF16 checkpoints plus NVFP4 and MXFP4 quantized weights, and co-developed kernels on NVIDIA and AMD hardware.

Google Maps and Google Search now work together in the Gemini API

Philipp Schmid — Wed, 12 Aug 2026 00:00:00 +0000

Use Google Maps and Google Search in the same Gemini API call with Gemini 3.6 Flash, then add custom functions or MCP servers for actions like booking a table.

Ari, Applied Compute's in-house AI research agent

Applied Compute — Wed, 12 Aug 2026 00:00:00 +0000

Muse Glimmer 30B Architecture Notes

Sebastian Raschka (site) — Tue, 11 Aug 2026 14:15:07 +0000

Short architecture note on Meta Muse Glimmer 30B, including gated local and global GQA, KV-cache efficiency, and release-time benchmark comparisons.

Thoughts I had while reading your comment

Dynomight — Tue, 11 Aug 2026 00:00:00 +0000

with absolutely no indication of neuroticism

5 useful things you'll learn in my new post-training textbook (shipping now!)

Interconnects (Nathan Lambert) — Mon, 10 Aug 2026 13:02:42 +0000

After a few long years of finding time to document my lessons from training open models, my post-training book is done!

Everybody Wants to Rule the (AI) Stack.

From the Porch (dslevy) — Mon, 10 Aug 2026 12:56:16 +0000

Part 2: Only one end has a lender.

Notes on Midtraining

Cameron R. Wolfe — Mon, 10 Aug 2026 09:33:22 +0000

Creating better specialized LLMs with midtraining and continual pretraining...

Announcing Day-0 Support for NVIDIA Nemotron 3.5 Lightning on vLLM

vLLM blog — Mon, 10 Aug 2026 00:00:00 +0000

How vLLM serves NVIDIA Nemotron 3.5 Lightning with OpenAI-compatible APIs, speculative decoding, and BF16/NVFP4 checkpoints across NVIDIA GPUs and edge systems.

Lessons from the hacks

Interconnects (Nathan Lambert) — Sun, 09 Aug 2026 14:57:11 +0000

Musings on model alignment, what determines safety, and where we go from here.

A quick(ish) Chinchilla check

Giles Thomas — Fri, 07 Aug 2026 19:00:00 +0000

I recently overtrained a couple of GPT-2 style models , training them both on 40 tokens per parameter rather than the 20 per parameter that is generally regarded as "Chinchilla-optimal". The normal heuristic is that instead of doing that, you should scale up the number of tokens and the number of parameters equally -- so I would have been better off scaling up the model by 2 and the token count…

Everybody Wants to Rule the (AI) Stack.

From the Porch (dslevy) — Fri, 07 Aug 2026 14:57:25 +0000

Part 1: The neoclouds went shopping.

LLMs From Scratch Reaches 100,000 GitHub Stars

Sebastian Raschka (site) — Fri, 07 Aug 2026 09:40:53 +0000

Short note celebrating the LLMs-from-scratch repository passing 100,000 GitHub stars and summarizing its learning materials.

Efficient Decode Context Parallelism with vLLM for Long Context Workloads

vLLM blog — Fri, 07 Aug 2026 00:00:00 +0000

Decode Context Parallelism (DCP) in vLLM shards KV cache across GPUs by sequence dimension, enabling 3× higher throughput on long-context agentic workloads compared to standard tensor parallelism.

The Odyssey as tech criticism

Blood in the Machine (Brian Merchant) — Thu, 06 Aug 2026 18:23:26 +0000

Christopher Nolan's Odysseus is an entrepreneur who unleashes an AI-caliber scourge on the world

One in five US workers now delegates tasks to AI instead of other humans

Epoch AI — Thu, 06 Aug 2026 00:00:00 +0000

One in five US workers say AI now does work they once handed to a coworker or contractor, per a new Epoch AI/Ipsos survey of 1,106 employed adults.

vLLM Reaches 25K Total TPS/GPU on Qwen3.5

vLLM blog — Thu, 06 Aug 2026 00:00:00 +0000

How vLLM reaches 25K total TPS/GPU on Qwen3.5-397B-A17B-NVFP4 with GB200 NVL72 disaggregated serving, Blackwell GDN kernels, HMA cache transfer, async scheduling fixes, and srt-slurm recipes.

Bring Your Own Harness to AC2

Applied Compute — Thu, 06 Aug 2026 00:00:00 +0000