Posts

Saturday’s lead is a first-party technical report on the agent swarm behind May’s RubyGems attack — 2,000+ packages pushed in two days — where the attribution is specific rather than rhetorical (49 of the same URLs as the German-wiki swarm OpenAI already confirmed) and the limits are stated (no access to the agents’ reasoning; unknown whether the API-key theft ever worked). Around it: a ~2,900-line LiteLLM replacement with an unusually honest compatibility claim, two MCP papers (a description-only vulnerability auditor and a seeded reliability census where most servers never start), three agent-training papers on harness evolution, skill optimization and sandbox memory compression, Clay’s first public statement on the Navier–Stokes prize, and OpenAI pulling out of Caltech’s AI mathathon after mathematicians objected. ...

Nick Abe runs a small Android puzzle app called Dayzle. Two weeks of Google Ads on a CA$40-a-day budget produced 56 billed installs. When he went into the raw analytics, 33 of them did something no person does: opened the app once, spent zero seconds on any screen, and never came back — across 28 phone models in 19 states. The 13 installs that were actual people finished 92 puzzles between them. The 270-comment thread on Hacker News is where the practical detail lives: countermeasures, near-identical experiences at other budget sizes, and one question nobody could answer. ...

Hamel Husain and a co-host spend an hour live-testing claude plugin eval — the plugin-evaluation feature Anthropic shipped for Claude Code (announced from the Claude Devs X account) — by pointing it at a plain-writing/tone skill. They never get a useful eval out of it, and the diagnosis of why is the interesting part. ...

An engineer at Earendil set out to answer the question the coding-agent industry mostly hand-waves: how do you actually measure whether AI-written code is bad? His starting point is that correctness is close to solved, and for a structural reason. Code is easy to check — run it against hidden tests and you get a clean pass/fail signal — so it’s the perfect thing to train a model on. Sloppiness is the opposite: duplicated logic, pointless abstractions, structurally bad decisions that pass every test. Judging that requires human taste, which makes it a genuinely hard measurement problem. ...

Twenty-five Fields Medalists — including Terence Tao, Peter Scholze, Manjul Bhargava, and Pierre Deligne — have signed a joint declaration arguing that AI companies are damaging mathematics by treating it as a benchmark to be beaten. The signatories don’t dispute the capability claim. They accept that LLMs have improved dramatically in recent months and can now solve major outstanding problems. Their objection is about goals, not ability: in their words, “the goals of the AI companies and the goals of the mathematical community are severely misaligned.” ...

Mo Moustafa runs Olly, an AI assistant that lives in iMessage, on open-source models through OpenRouter — over 18 million messages to date, roughly a third of them on open models through the router. That is enough volume, as he puts it, to hit every edge case at least once. His post is the list of things he wishes he had known going in, and it is the most concrete public account of what a routing layer actually costs you. ...

The X account @stochasm — an inference and systems engineer with a technical following — spent yesterday morning posting a live read-through of DeepSeek’s V4.1-Flash tech report. It is 27 posts long, one observation at a time, with no tidy summary at the end, and it stops mid-pass at “gonna take a brunch break.” Which is exactly why it is useful: you get the reasoning as he encounters each design choice rather than a conclusion somebody else smoothed out. Here it is, with his judgments attributed to him and his uncertainty left in. ...

Daniel Hook, chief scientific officer at Holtzbrinck Group, took a Waymo across San Francisco and noticed something uncomfortable: as a self-confessed introvert, he found the empty driver’s seat delightful. He was in town to collaborate with Susan Winslow, chief executive of Macmillan Learning. Two people who had crossed an ocean and a continent to talk to each other were quietly pleased to be in a car whose defining feature is that you don’t have to talk to anyone. ...

Andy Balaam makes programming videos and has been writing code for decades. His short essay is not an argument about whether AI works, and not an ethics piece — it is about how the last couple of years have felt, and he is blunt that the sadness has been AI’s single biggest effect on his life. Bigger, he says, than any concrete change to his actual work. He traces that feeling to disrespect rather than fear: ...

The author spent 2006 building Flash microsites in Manhattan. Flash was objectively bad — inaccessible, unskippable intros, back button broken — and the web is better without it. But it dropped a powerful tool into the hands of people who weren’t “web programmers,” and a wave of weird, unexpected work came out the other side. Product configurators, drag-and-drop, loading screens as a creative medium. Some of it was bad. That was the point. ...

John D. Cook flags a detail he says the coverage of OpenAI’s Navier-Stokes announcement skipped: alongside the human-readable proof, OpenAI published a Lean 4 formal proof. Lean is a proof assistant — software that checks a proof step by step, so “verified” means a machine confirmed the reasoning, not that a reviewer found it persuasive. That combination — AI generating the proof, a checker certifying it — is becoming standard for AI-assisted mathematics. What makes it notable is the price. ...

Armin Ronacher gave GPT-6 Astra one ambitious goal — build a version of Python with a couple of long-wanted language features — and then deliberately stayed out of the way. The agent managed its own context, kept its own notes, and spun up its own subagents over a weekend. The results, in his accounting: 35 hours of unattended running ~1 billion tokens, roughly $1,200 in API costs 75,000 new lines of code across 79 commits (about $15.50 per commit) ~1,400 messages exchanged between agents Zero useful output, and no lesson about how to run a better “software factory” The interesting part is not that it failed. It’s what kind of code it wrote. ...

A “harness” is the program wrapped around a model: it holds the system prompt, hands the model its tools, and runs the loop. Nathan Sutton ran nine of them — opencode, cline, codex, goose, crush, dsh, pi, mini-swe-agent, plus his own chad — against a local model on a laptop instead of a data center, on eight small coding exercises each. The point isn’t which one wins. It’s that the reason local agents feel broken is structural. ...

Friday’s feed was agent-platform shaped: 10 keeps out of 945 gathered, and OpenAI shipped an Agents API the same day it put its full-duplex voice model in the API. The Agents API is the one to read closely — server-side sessions, subagent orchestration, and a self-hosted sandbox with a skills path, but US-only data residency and explicitly not ZDR-eligible, including when you bring your own sandbox. Around it: Cognition’s SWE-2 lands within a point of Fable 5.1 on its own benchmark at a claimed 64% lower cost, a 122B MoE trained to drive a real shell for 300+ turns, three agent papers including a prompt-injection detector that runs as a sidecar over existing logs, and OpenAI pausing new $200/month Pro signups while the API stays unaffected. ...

Andreas Thom is a mathematician whose work with Gabor Kun supplied one of the techniques behind a recent OpenAI-assisted result on non-sofic groups. He and a colleague had spent months discussing related problems with ChatGPT. So when the result appeared, he emailed two OpenAI researchers to ask a direct question. The question had two parts: Did our conversations end up in the training data — the material used to build later models? Were those conversations reachable by the system while it was solving the problem? The complete answer he received was one sentence: “Regarding your conversations with ChatGPT: that did not happen.” Thom argues it addressed only the second half, categorically, with no evidence and no reference to his account’s settings. ...

Thursday’s feed ran 10 items deep out of 929 gathered, and the biggest story is a config change. DeepSeek shipped V4.1-Flash — open weights, MIT, 1M context, built on KV-cache compression — and said it is retiring V4-Flash while phasing out V4-Pro, so all deepseek-v4-pro traffic reroutes on Sep 14 and a pinned model string can silently resolve to a different model. Around it: three agent-harness papers (subagents vs. skills, spec-first enforcement, and a 197-word distilled harness), a multimodal prompt-injection benchmark where model choice matters more than framework, 18 on-device models with a Swift/Kotlin SDK, and Anthropic handing ENISA test access while still withholding Mythos 5.1. ...

Rob Kaminski — co-founder of Fletch PMM, the consultancy that has run positioning work for roughly 600 B2B startups including Notion, GitLab and Superhuman — does a live teardown with host Angelina on TwoSetAI (1h02). The subject is a real AI agent startup: AgensFlow, Nicole Königstein’s coordination layer that sits above frameworks like LangChain and LlamaIndex and learns which agent and model to use for each task. ...

Anthropic’s public brand is the careful lab — the one that spent early 2026 in a standoff with the Pentagon over mass surveillance and autonomous weapons. The American Prospect’s Daniel Boguslaw reports that the same company is building an intelligence operation aimed at the people who oppose rapid AI development. Anthropic did not comment. The reporting draws on the company’s own job postings, a podcast interview with Anthropic’s security staff, company statements to the Wall Street Journal, and local coverage of police reports: ...

Anthropic’s economics team built a model of how AI could reshape the US economy through 2030, plus a public explorer where you plug in your own assumptions about AI capability, adoption, and how fast people find new work. The framing is simple: the economy is a pile of tasks, and AI can augment a task, automate it, leave it alone, or create new ones for people to do. They run three scenarios: ...

Cole Medin on the security gap in AI coding workflows (17 min, partnered with Sonar; Archon, the harness builder he uses, is his own open-source project). His premise: agents are already better and faster than humans at a lot of code — but security is where they fall short far more than reusability, documentation or performance, and almost nobody is building for it. ...