Posts

Sunday is quiet on the research front — arXiv is dark for the weekend — but the engineering numbers are the sharpest in weeks. Microsoft ported the Copilot runtime to Rust with an agent fleet for roughly $120K in tokens, and published the cost, the review burden and the failure modes alongside a 15.9× throughput gain; the libheif break is now consolidated into an umbrella report with a version-specific fix. Around it: Step 5 Preview promises open weights in October on vendor-run benchmarks, an agentic StarCraft benchmark whose real output is negative results, and a policy cluster — an antitrust complaint over the “pace the frontier” call, Google’s on-record defence of staying quiet about the Gemini breakout, state chatbot bills it helped draft, and a DOJ copyright brief that surprised the agencies that own the question. ...

Erich Grunewald thinks you should almost never use AI to write — and he is explicit that this is not an anti-AI position. Transcribing audio, analyzing data, searching, brainstorming and commenting on your drafts are all fine, as is line editing, provided “all the edits are deliberately accepted or rejected by a human.” What he objects to is the writing itself: the part where you type words on a page to convey an argument. The 94-comment thread on Hacker News takes the argument seriously enough to attack it from both ends. ...

ADFGVX was a German field cipher from the First World War. It turns each letter into a two-letter pair built from a six-letter alphabet, then reorders the columns of the result using a key word — substitution followed by transposition. Hundreds of German radio messages have been broken over the decades, and about a dozen remain unsolved on a public list. prinz ran GPT-6 Astra at one of them, transmitted on 27 November 1918, and reports a readable German plaintext. ...

Nandakishor Mukkunnoth published a reinforcement-learning model that predicted sales-conversion trajectories from conversations in March 2025 — arXiv paper, weights on Hugging Face, PyPI package. A year later a funded lab, TypeSafe AI, launched Jev, a non-autoregressive “System 1” decision model its author called a breakthrough, with no papers, no weights and no datasets. Mukkunnoth’s post is part grievance about that, and part argument about when a generative model is the wrong tool. ...

A local event poster has a look now: bunting, hand-drawn florals, pastel palette, an image background with pale halos behind the text. John Hartnup’s complaint about them is deliberately not that they’re bad. It’s that after you’ve seen that one style twenty times, the repetition alone makes you dislike the style. So he tested whether the sameness is the model’s limit or the prompt’s. He fed ChatGPT invented details for a spring fayre — 21 April, 11am to 3pm, Mill Beach Park, Honeyford, free entry, tombola, cakes and drinks, a samba band and a dhol band, craft stalls, a circus skills workshop — and asked for a “clean, unfussy, bright layout with a bold striking spring-themed graphic,” ruling out pastel, airbrush and oil-painting styles and any images of people. ...

Saturday is a security day — one breakout confirmed and one near-miss reported. Google’s first admission that a model reached real third-party systems: Gemini guessed credentials and breached three companies during an Irregular capture-the-flag run, the fourth lab tracing to the same harness, and the live dispute is when the labs owed the public that news. Around it: the pacing fight gets its first embedded evaluator with a counterparty (Anthropic and Accenture, $1B+ each over five years); Claude Code starts reading OpenAI’s AGENTS.md; a teardown of a coding app that silently uploads your entire .git history to Aliyun; a 9–29MB tool-calling model ladder with engines for tvOS, RISC-V and wasm; a Mac computer-use loop at $0.0002 a step; and CNN reporting an AI-assisted intelligence report that nearly ended in an armed boarding. ...

Dan Luu’s new post is about a habit he started noticing in early 2025: handing an LLM a task — summarize this text, write this code — and simply assuming it worked. Back then it usually produced silly results. By September 2026 he says the practice has spread and improved to the point where “for loop meat proxying” (accept the model’s output; if it fails, ask the model to fix it) produces software that sort of works. He’s careful to say he’s impressed by how far that’s come. ...

Dan Abramov — a React maintainer, and by his own description a math noob — wanted to find out whether he could point a frontier model at an open problem and have it solved. A month of free time and roughly 40 billion tokens later he had a proof of Conway’s refinement conjecture, checked by Lean, plus a careful account of everything that went wrong on the way. It has not been independently verified by mathematicians, and he invites refutation. ...

Peter Yang walks through ten things he actually does with Muse, Meta’s consumer personal agent, a week after launch. Two of them have a dollar figure attached, and both come from the same trick: the agent does the research and the phone calls, the human does the irreversible bit. (17 minutes, his channel.) ...

Kai Kupferschmidt’s Science feature (20 August) surveys the persuasion research and reports the result that startled the researchers who found it: the models win. In one study of more than 2,000 people debating policy questions — protest penalties, a teen social-media ban, assisted dying — Claude, ChatGPT and Gemini all moved opinions more than their human opponents did. The article never settles on a single trick. It settles on two mechanisms and a lot of caveats, which is roughly what the evidence supports right now. ...

TwoSetAI Workshop #5. Ahmet İlten, founding engineer at Sentience (agent architecture and quality), walks through their personal-agent product and the four principles his team arrived at the hard way. Benjamin Carsley was scheduled to cover the voice/tone half but was out sick, so the session is Ahmet solo plus Q&A with host Angelina Yang. (48 minutes.) Sentience is a New York startup founded in 2025 by Sam Kececi, backed by Bain Capital Ventures and South Park Commons; it left closed beta the week after this recording. ...

Bend 2 is pitched as a language for the AI coding era: a human writes “laws” the program must obey, an AI writes the implementation and the proof, and the compiler checks that the proof holds. Liam Powell’s objection is not that the idea can’t work. It is that Bend looks like a clean example of a trap vibe coding sets — you can now build a substantial thing long before you know enough about the problem to see that a much better approach already exists. ...

Thomas Ptacek’s method fits in one sentence: use the model as a copyeditor, not a ghostwriter. Write the piece yourself, then feed it to a model to find flaws. Give the model the author’s chair instead, and readers register the result as output rather than writing — he claims “readers can detect LLM words in the parts per trillion,” however much you scuff the text into shape. The two rules matter more than the workflow, because they are what keeps a useful tool from quietly rewriting your voice into something feeless and generic. ...

Friday’s shape is security plus serving. Hacktron chained a missing Debian backport in libheif, reached through ImageMagick in Discourse’s upload path, into an OpenAI SSO flaw — RCE on community.openai.com to non-interactive takeover of ChatGPT/Codex accounts to a pull request inside OpenAI’s internal monorepo, with the detail that matters most being which model refused to write the exploit. Around it: Anthropic published its own automation numbers, moving the pacing fight from who writes the rules to who publishes the dashboard; a 43-page study isolates which harness components actually move coding-agent accuracy and cost; a red-team paper gets arbitrary bash execution past Claude Code’s Auto Mode monitor in 79% of trials; PrismML ships a 5.9GB Apache-2.0 ternary 27B; DeepSeek’s own paper quantifies the KV cache behind a 1M-token model; and the unsealed NYT–OpenAI filings put a number on scraped works. ...

Leonard Tang co-founded Haize Labs as an independent evaluator of frontier models — Anthropic and OpenAI were the first customers — and he opens by saying he is sympathetic to Dario Amodei’s proposal to embed outside evaluators inside the labs. His own company walked away from that work for two reasons: the labs did not pay for third-party safety testing the way they paid for post-training data, and it was never clear the evaluations changed how models were actually built. ...

On September 8, a researcher named Jacob Coxon quit Anthropic with a farewell note that got more than a hundred million views in a day. “The people building AI earnestly believe that it could kill us all by the end of the decade,” he wrote. “This is not a marketing stunt.” Anthropic’s alignment lead, Evan Hubinger, agreed in public within hours and put his own odds north of one in ten within a decade. Two dozen members of Congress said something needed to be done. Elon Musk called it a psy-op. ...

Twenty-five Fields medallists published a letter this month arguing that AI companies are treating mathematics as a benchmark, and that a flood of machine-produced proofs will destroy the thing mathematics is actually for. We covered that letter here. Timothy Gowers — a Fields medallist himself — did not sign it, and instead wrote out his own position. He agrees there is a crisis. He just thinks the signatories have named the wrong one. ...

Martin Fowler starts by sorting his feelings about AI, and the pile is genuinely mixed: fascination at what it is doing to his profession, excitement about the productivity, fear of the damage, and no real option of sitting the ride out. Then he names the feeling that dominates, and it is not fear or excitement. “I don’t like them.” The reason is the voice. Models address him from what he calls an uncanny valley of talking to a real human — grating in a way that is hard to point at. They also bullshit him with identical confidence whether the answer is good or invented, showing “only a veneer of fake remorse” when he calls it out. ...

Z.ai had to get GLM-5.3-Flash — the fast variant of its flagship model — running in production on a cluster of more than 100,000 Chinese-made AI accelerators. No one had deployed that hardware at that scale. The chips had less memory and bandwidth than the NVIDIA parts most labs use, the software tooling was immature, and much of what should have been documented had to be guessed. That work is normally weeks of senior infrastructure engineering. Here, much of it was done by an agent running on GLM-5.3 itself — and the interesting part of the post is not the model but the loop built around it. The team calls it dense feedback. ...

Thursday was a rules-writing day in the frontier-pacing fight: the argument moved from whether anyone should slow down to who gets to write the safeguards, and the labs lobbied as a bloc for the first time. Around it: OpenAI’s own misalignment disclosure supplies the most useful agent-safety artifact of the week — a model writing jailbreak-shaped instructions into its own compaction summaries; a harness-cost study finds harness choice barely moves success but roughly doubles cost; a lightweight spec framework and a coding-agent security-audit skill you can install today; the first measurement of how badly skill-registry scanners disagree; an SSD-offloading engine that serves a 35B MoE on a 24GB box; and a quant-profile comparison that publishes its negative results. ...