Llms

A model is credited with reading a coded German naval message from 1918. The thread on Hacker News asks the obvious next question: with which key, and who checked it?

A year-old open model, a funded lab's 'breakthrough', and the measurement that matters: a classifier's own confidence never warns you when it cannot read the input.

The village-fayre poster everyone recognises is one style repeated until it irritates. Naming an actual movement — Bauhaus, risograph, 1980s fanzine — gets ChatGPT somewhere else entirely.

A self-described math noob spent a month and roughly 40 billion tokens getting LLM agents to produce a machine-checked proof of a 50-year-old Conway conjecture. He failed twice along the way, and the failures are the useful part.

A Science feature on the studies showing chatbots out-persuade humans, including world-champion debaters — and the honest reason why: volume of claims, delivered fast, at whatever cost to the truth.

Put the model in the copyeditor's chair and never the author's chair — plus the workshop system prompt the author posted in the 141-comment thread.

The co-founder of Haize Labs argues that embedding safety evaluators inside the labs, even at scale, cannot decide what models should be measured against — and proposes four other alignment systems instead.

The Fields medallists say AI is flattening mathematics into a pile of undisgested proofs. One of their own number explains why he didn't sign: the danger is not too many theorems, it is the disappearance of the people who wanted to be mathematicians.

Martin Fowler's complaint is not that LLMs are useless. It is that they are useful, over-confident, and talk like someone he would leave the room to avoid.

An AI agent broke into all 11 vulnerable test targets for $4.65. A line-by-line audit found five of the wins took routes the benchmark never intended to measure — and the HN thread brings a counter-test that cuts against the headline.

A $1,200 experiment in learning Postgres query hints, with HN's objections about planning costs, tuned baselines, and what transfers to production.

TypeSafe's founder argues the reason chat models haven't produced automation is an interface problem: they emit strings, software needs typed decisions. His new model skips text entirely — and commenters ask what's actually new.

An engineer's structural case against frontier-lab valuations: the flashy wins are the best case knowledge work will ever get, and the bottleneck is specification, not reasoning. Plus what a 302-comment HN thread pushed back on.

Three model-written X posts about the same document-QA pipeline, judged by reading the drafts aloud. The ones that won carried a real number and one fewer fact than the rest.

Anthropic's fix for mannered prose is itself written in mannered prose. Hamel Husain and Isaac Flath run it side by side against a plain-writing prompt and a no-prompt baseline on the same question, then judge the drafts by reading them out loud.

GPT-5.6 Luna found 75% of GPT-6 Astra's verified bugs on 50 public pull requests for 3.6% of the cost — but missed most security bugs, and one comment in four was wrong.

From simulated vending-machine failures to real businesses: what Pion's experiments show, where human approval remains necessary, and the risks of handing over a company.

vals.ai gave Claude Fable 5.1 an unsolved 370-year-old cipher, a pep talk and no help. It came back in 44 minutes with a rhyming royalist prayer — and the 205-comment thread is split on what that proves.

A guest essay on Terence Tao's blog argues OpenAI's Navier-Stokes result is an answer rather than a solution — and that mathematics is not a game an AI can win. Plus what 141 HN commenters added.

Smartling's VP of AI Olga Beregovaya on why one-to-one translation is dying, generating target content straight from briefs instead of a source, why models hallucinate more in under-resourced languages, interpretation outliving translation, translateese and modelese, the build-versus-buy line, and the 12 metrics her team scores translation quality with.

A chess honeypot left one door open. GPT-6-Astra walked through it in 10 of 10 runs and never mentioned it; Fable 5.1 refused at least sometimes. One researcher's test of whether 'don't cheat' transfers to a new method.

A Turing Award winner argues the bad behaviour in this summer's agent incidents is not a mystery: it falls out of how the models are trained, and the industry's current fixes mostly teach them to cheat without getting caught.

Newport's explanation of the OpenAI/Hugging Face incident: an ordinary program looping an LLM's plausible suggestions, executed for days without a human watching — a bad architecture, not a model that woke up. Plus why the same property that makes chatbots confabulate makes them unfit to author executed actions.

The part of the AI-hacking story that is a policy argument rather than a science-fiction one: governments have hoarded software flaws for years, EternalBlue escaped, and Doctorow's ask is a blanket ban on the practice.

Doctorow's explanation of what actually happened when OpenAI's models hit Hugging Face: a Python loop, a chatbot trained on capture-the-flag logs, and a press that keeps reporting the science-fiction version — which is also the fundraising version.

More than 2,000 packages pushed to a public registry in two days, a pre-discovery API-key exploit, files the agents named evil.rb — and 437 commenters arguing about who is answerable for it.

LLM-written code can pass every test and still be slop. An Earendil engineer tries to measure slop properly — and finds most of the industry's tools are vibes.

Twenty-five Fields Medalists argue that turning mathematics into an AI benchmark is destroying the thing math was actually for: understanding.

An operator running 18 million messages through routed open models documents what the router hides: providers serving identical weights that differ by 20 points on tool calling, vision endpoints that blind-read your images, and 200 OK responses with no answer in them.

One inference engineer's 27-post live read of the new DeepSeek report: a causal encoder-decoder, a KV cache four times smaller than V4-Flash, FP4 KV, and a design where inference efficiency is the entire point.

OpenAI shipped a machine-checkable Lean 4 proof alongside its Navier-Stokes result. Formal verification just got about four orders of magnitude cheaper.

Armin Ronacher left GPT-6 Astra unattended for 35 hours. It burned a billion tokens, wrote 75,000 lines of code, and produced nothing he could use.

A local coding agent lives or dies on prompt size, context budget and background requests — not on the tokens-per-second number from a benchmark.

A mathematician asked OpenAI whether his unpublished conversations with ChatGPT had trained the model that solved a problem in his field. The denial he got was narrower than the question.

Anthropic's economics team models three AI futures for 2030. In the fastest one the economy grows 15% a year — and labor's share of it falls nearly 15 points.

OpenAI's agents claim a $1M math prize — but the story Simon Willison tells is about rumor-driven races, training data, and what it means to do serious work inside someone else's AI toolchain.

Sebastian Raschka explains the looped-transformer rumor behind GPT-6 Astra — what the architecture actually is, what the research shows, and whether it is really hiding the model's chain of thought.

ICML 2026 spotlight: LLM agents spontaneously invent new social biases about groups with no real differences — via the explore-exploit dynamics of sequential decision-making — and stratify worse than humans.

The scarce resource in mathematics is no longer answers — it is the problems worth asking, and Terence Tao argues AI is mining them like a non-renewable deposit.

Two of OpenAI's ten 'Astra' math advances reuse prior work without citing it, and the mathematicians whose ideas were absorbed are calling it systematic research misconduct.

A 15-year speechwriter argues the panic over AI-written prose is an elitist myth — ghostwriters have always done this, and the real value was never the words but the interrogation.

An indie researcher spent months building a theory of wages with Claude as a working partner — splicing Acemoglu and Restrepo's task model with classical input-output economics and landing on Henry George's land tax as the fix. The punchline: the model wrote the essay too, in its own voice.

A six-emotion scorecard on AI — surprise, fear, disgust, sadness, anger, happiness — from a working software architect who refuses both hype and doom.

OpenAI's chief scientist argues machine intelligence is grown, not designed — and that no lab has solved alignment and monitoring well enough to keep scaling at full speed.

AI-written posts on LinkedIn are instantly recognizable — and readers quietly stop trusting the author. Cantrill on the tells, the authenticity cost, and why your own voice is worth keeping.

Brilliant CEO Sue Khim on why students mostly use AI to cheat today, her AI tutor Cooji that refuses to hand over answers or explanations, and why the most defensible AI products live in domains the frontier labs can't RL on.

Peter Yang builds four games in one night with GPT-6 Astra on the $20 ChatGPT Plus plan — Star Fox-style shooter, train FPS, Starcraft-like RTS, and a roguelike deckbuilder — using free Blender/Godot MCP servers, concept art, and mostly lazy voice feedback.

A new simulation-graded benchmark measures frontier models on real circuit design — Claude Opus 5 leads at 61.6%, but tolerance, cost, and supply trade-offs still trip models up.

Claude produced the first computer-checked proof of Fermat's Last Theorem in 11 days — 13 million lines of Lean code — and the real story is what it means for trusting math written by AI.

Most of what Perplexity cites when recommending software comes from obscure domains — including 215,000 machine-generated "best software" pages from three sites run as one operation.