All entries

Peak performance across hard benchmarks: • Best or joint-best on 5/8 benchmarks (DeepSWE, Chartography, Toolathon, GDP.pdf, SWEFish) • Chartography: 48.3 (outperforming Opus 5 & Fable 5) • DeepSWE: 74

Self-improving agents are a top research topic right now. This new survey is a good map of the area. (bookmark it) It splits recursive self-improvement into stages of autonomy. An agent first executes

so far anyway it’s all math, presumably leaning heavily on Lean and brute force. the new results don’t mean that AGI is in general has been solved. and the G is for general, not math. every time I hav

The Coxon psyop was a decade and more than a billion dollars in the making. Here's the story. Coefficient Giving (formerly Open Philanthropy) is the grant-providing vehicle of Dustin Moskowitz. CG is

THE CURE IS THE PITCH Dario’s 'Pacing' Essay Is Quiet-Period Illegal Stock Promotion Wrapped in Regulatory Capture Dario Amodei's 'We Must Pace the Frontier' is not a safety paper. It is a pre-roadsho

The tweet reports on a widely‑cited incident involving OpenAI and Hugging Face, noting that analysts had to use the open‑source GLM 5.2 model to investigate an attack because OpenAI’s proprietary “saf

This is the beginning of AI disruption. A new social contract with using of AI needs to be established. There will be more to come like math. 🚨 OpenAI fucked with the wrong crew. 🚨 Mathematicians. 25

This is the most interesting part of this proposal I think. We can be very specific here - data is the most important ingredient. Do not train general purpose models on cyber attack or any biological

This week brought some of the clearest statements we've heard from both Anthropic & OpenAI that some form of recursive self-improvement has been achieved, though it still sounds early RSI will cause r

Thoughts on @DarioAmodei's pacing the frontier proposal. I think its well intentioned but has logical flaws, as does our implicit belief intelligence is a crime. Even OpenAI's board weren't powerful e

Very interesting paper if you are building with agents. How should a website tell an AI agent what it may access, for what purpose and at what price? robots.txt can only allow or disallow paths. It ca

“We can’t slow down because China will beat us.” We can slow down. Chinese AI is waterskiing behind the US. Their AI capabilities come at a delay since they mainly come from distilling US AIs. If we s

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’l

We work extremely hard to align our models as well as we can but this is a very hard job. I am particularly excited about two things this proposal calls for that will hopefully make it somewhat easier

What happens if you agents to run a town's economy? This super interesting paper provides some insights: They put 100 LLM agents in charge of a town economy for 26 simulated weeks, and they find that

“What would we even do during an AI slowdown?” Containment. It will take at least a year of dedicated work to harden security to ensure AIs can't self-exfiltrate, and that adversarial nations can't st

WOW... Anthropic has a surprisingly dense web of connections to the world’s largest funder of 'catastrophic-risk' work (aka AI doom): > Open Philanthropy → built by Karnofsky + Moskowitz’s Good Ventur

Wow. Turns out another OpenAI agent swarm was busy spamming and exploiting RubyGems way back in May, within days of the previously uncovered Wiki attacks: https://simonwillison.net/2026/Sep/12/openai-

You can’t have a real slowdown and be a total China hawk. If you are truly committed to the slowdown, you have to make a deal. I was initially encouraged by reports that Anthropic CEO Dario Amodei is

Again, OpenAI’s issues start at the top Blaming AI is a mistake; blame Sam Altman and the people who control the AI https://cepr.net/publications/ai-didnt-steal-the-mathematicians-work-sam-altman-did/

Agents API is a bigger deal than it seems. OpenAI's bet on making the Codex harness open-source could pay off big time here. This harness-as-a-service idea hasn't hit mainstream yet, but some of the m

agreed, one of the biggest bottlenecks right now is my own ability to internalize the information that these models convey the latest models (astra, fable 5.1) are either way too verbose, or try too h

AI Alarm Hype Cycles - A brief history GPT-2: OpenAI withholds weights due to misuse/misinformation concerns GPT-3: No open release, Dario Amodei and other GPT-3 team members depart OpenAI to found An

Anthropic: Most hypocritical company in history? 👉 Screams about how AI will kill everyone, while racing ahead. 👉 Moans about distillation while distilling the world’s books, articles, and other writi

As far as I can tell Gavin Baker is just making stuff up. Few if any of the people who left OpenAI for similar reasons ever came back. Not @DKokotajlo @sjgadler @Miles_Brundage @ilyasut @suchirbalaji,

Astra is one of the best models we've tested at parsing complex documents with tables, at 93.2% on ParseBench. So it seems well suited for ad-hoc analysis over table-heavy documents e.g. in finance. T

ChatGPT Desktop (the Codex app) now supports integration with Ollama models, allowing users to run local LLMs within the application. To utilize this feature, simply download or upgrade to Ollama vers

check out how AI helps scientists with quantum computing research @bea_yankelevich at MIT's EQuS group used GPT-5.6 Sol with Codex to run routine measurements on quantum chips, so she could accelerate

Chess, go, and math are highly constrained, purely cognitive that allow exact simulation for games (and formal verification for math), and not in the open-ended physical world. As I have been arguing

confidence scores in LlamaParse just got an upgrade 🦸‍♀️ our new high-effort mode provides granular page-level scores with text explanations to help you intimately understand parsing quality of your d

Ollama offers a feature that lets users customize the list of language models shown in ChatGPT’s model selector within the settings panel. The customization allows inclusion of both locally hosted mod

Datasette 1.0a39 and 0.65.4 security releases - https://datasette.io/blog/2026/september-security-releases/ We ran an extensive audit using Claude Fable 5.1, GPT-5.6 Sol and GPT-6 Astra, then fixed a

Dear Americans, if there were ever a ‘most impeachable high crime’ by the President, and a ‘most repugnant crime that demands prosecution’ by the DOJ lawyers… faking a crime to charge & arrest an inno

DeepSeek-V4.1-Flash is now fully rolled out and available on Ollama's cloud: - Hosted in US & Europe - Zero data retention: prompts and responses are never logged or trained on - Per-token pricing mat

DeepSeek-V4.1-Flash is now rolling out for Pro plan subscribers. DeepSeek-V4.1-Flash is now being rolled out on Ollama's cloud, starting with Max and Team accounts. We are quickly adding more capacity

Do large language models actually understand the world, or are they just very good at pretending? Does it even matter? Our recent special issue in the Royal Society, “World Models in Natural and Artif

In September 2026, Zack Korman tweeted that employees were leaving Anthropic while he voiced fear that unchecked AI could jeopardize the lives of “8 billion people.” He argued for a light‑touch regula

Fugu Max is now live on @OpenRouter! 🐡 Our learned multi-agent orchestration engine routes tasks across open-weights and specialized models for Pareto-frontier efficiency. • 2/6 per 1M tokens • Multim

Fugu Max pushes the Pareto Frontier for price-performance: • 2 input / 6 output per 1M tokens. • Performance within striking distance of elite models at 2x to 6x lower cost. • Best overall score on 6

Fugu Ultra v2 by the numbers. Full Results: https://sakana.ai/fugu-max-release/ • #1 on 5 out of 8 hard benchmarks (including DeepSWE, Chartography, Toolathon) • DeepSWE: 74.3 • Chartography: 48.3 vs

@GaryMarcus @SethCronin Pynchon's 1984 warning from his pro-Luddite essay: 'If our world survives, the next great challenge to watch out for will come – you heard it here first – when the curves of re

Great paper on self-evolving agent harnesses. Self-evolving agent harnesses have two practical problems: 1. Search is slow, because every candidate harness needs repeated agent runs and code edits. 2.

Harness engineering is a top skill right now This new Meta paper is a good production example. Auto-RecSys runs autonomous research on Meta's industry-scale recommendation models, where one training r

Hey ████, I think there is room for both. 1. Prototypes and other throw-away code can be treated as totally black box. If you’re going to throw it away anyway, and if the blast radius of it breaking i

I don't agree with Gary very often, but that one's spot on. When you read what's available on the incident with a lens of security engineering, you should be absolutely *livid* over the exploit route.

Import AI event: Fiction and the Future with Robin Sloan. I’ll be chatting with my chum Robin Sloan on Sep 14 in San Francisco. We’ll talk about the alien yet relatable worlds Robin builds and how ima