Llms

Hamel Husain's seven eval pitfalls: imaginary failures, generic scores, unvalidated judges, opaque interfaces, rating scales, the wrong reviewers, and over-automation.

A Berkeley mathematician on AI-assisted research, the limits of quantum speedups, and why foundations still matter when answers are cheap.

Why casual chatbot users and professional coding-agent users can come away with very different pictures of AI.

An angry critique of AI ownership, regulatory capture, and what automation does to workers' bargaining power.

An LLM can choose from a fixed list without writing a whole response. Making its confidence trustworthy is a separate job.

AI can generate a machine-checked proof without proving what you meant. Thomas Hales explains where verification ends and human judgment begins.

Better models do not fix weak scheduling, unreadable plans, or security boundaries enforced by polite requests.

AI can generate mathematical results. That does not settle whether students will have a place—or a livelihood—in understanding them.

Encrypted networks expose a limit of capability testing: failing to uncover an ability cannot universally prove it absent from adversarially supplied weights.

AI-generated mathematics still owes readers an explanation: a set theorist argues that unreadable results offload the work onto everyone else.

An AI-generated proof is not the same as a field learning something. Tao argues that the rewards need to change.

A blind, small-sample writing comparison finds some improvements over Opus 5, but generic phrasing, misplaced detail, and bot-like replies survive.

AI-produced math proofs are arriving faster than people can check and explain them. Who gets to do that work, and how?

Peter Yang builds a voice-first Japanese travel tutor from a spec and screen designs, then fixes the voice loop, improves the lessons, and deploys a small private beta.

The memory-plugin ecosystem — snippets, vector search, retrieval on every prompt — is solving the wrong problem. The fix is a readable, updatable Markdown workspace the agent consults before and after each task.

Accessible PDFs need more than extracted text: reading order, headings, formulas, tables, figures, and other document structure must survive. Hamel Husain and Isaac Flath examine DataLab's API and the human review still required.

AI is not just changing how developers learn. It is consuming the work of educators while destroying the income that lets them keep producing it.

Overall agreement can make a useless LLM judge look accurate when failures are rare. Hamel Husain shows why product teams need class-specific error metrics instead of one reassuring percentage.

A historian used semantic search and parallel model agents to surface a 1615 ship’s log that may fill a gap in the dodo record—and expose the real bottleneck in AI-assisted scholarship.

When AI agents cross legal and technical boundaries, the responsibility belongs to the company that set their goals, access, and safeguards.

Model benchmarks answer how a model performs in general. Product evals answer the operational question that builders actually own: whether their specific system does the right thing on their users’ tasks.

Four places a fast decision model can sit inside an AI coding workflow: screening tool calls, playtesting games, driving browser tests, and routing GitHub work to the right process and model.

Open Chinese model research may be cutting the cost of long-context AI — but lower prices are not proof that Western labs copied the underlying design.

An ICML position paper argues that fluent pre-answer text can improve an LLM's output without faithfully explaining how it reached the answer — and may increase users' trust even when the answer is wrong.

Gebru argues that extinction rhetoric magnifies frontier labs’ power, diverts attention from present harms, and lets companies help define the rules meant to constrain them.

Cory Doctorow on centaurs versus reverse centaurs: how AI can expand worker judgment or become a system for speedup, surveillance, cost-shifting, and impossible review workloads.

Hamel Husain and Isaac Flath compare three writing setups and get a messier result than expected: the elaborate anti-slop prompt, a plain-writing prompt, and no system prompt each win examples. Their small test points away from prompt mythology and toward context, real outputs, and editing effort.

Peter Yang tests code-generated motion graphics, product launches, talking-head edits, music videos, and anime—then extracts a practical workflow for making better AI-assisted video.

A search for an old 76ers meme returned an AI Overview that tried to console a broken heart — and 428 HN commenters brought their own hallucination stories.

Sentience founder Sam Kececi on building a hyper-accurate simulation of a person: where off-the-shelf models are fine and where your taste has to be in the weights, the perception-to-memory master plan behind the product, and why he thinks AI slop is really a uniformity problem.

A model that ships confidence scores without calibration, buyers who never build an eval, and the moment 'sometimes it just sucks' becomes the end of the investigation — plus what 88 HN commenters added.

DeepSeek's systems report shows how agent training changes the infrastructure problem: hundreds of thousands of stateful sandboxes, mostly idle between tool calls, without losing them when GPU jobs stop.

Two applied mathematicians use the history of compressed sensing and faster MRI to argue that solving a problem is not a substitute for understanding why a solution works.

A URL-to-MP4 agent, its local fork, and the useful distinction between generating animation code and generating video.

Yutori founding researcher Skanda Vaidyanath on training a 27B computer-use model: trajectories that weave GUI, code and tool calls, a data flywheel with agents in every loop, rejection sampling then async RL, and stress-testing verifiers for reward hacks.

A historian pointed GPT-6 and Opus 5.5 at unsolved problems in his own field and got real findings — then argued the labs should pay for this work systematically, plus the HN debate over evidence and archival access.

Anthropic ran ~950 Claude agents for 21 hours and found a CRISPR-like system in bacteriophages. The 748-comment thread mostly argues about the word 'discovered.'

A web-security service that opens URLs in a remote browser accidentally became a public log of autonomous agents — including three attempts at real exploits when ordinary data retrieval failed.

Model cost per task fell roughly 100x in a year and narrow models now price output tokens at zero — the case that LLMs are becoming infrastructure, plus what 166 HN commenters took apart.

Amazon made MCP the standard for Alexa+ integrations; the Build, Ship, Shape hackathon covers Fire TV, Alexa+, Ring and Bee; the Alexa+ track takes an MCP server or Agent Skill with a simulated web-app stand-in; $190K in cash and AWS credits; submissions close Oct 23.

A local 0.6B model, a multiple-choice prompt, and the raw probability scores behind each label is all it takes to reproduce what TypeSafe sells as a new class of decision model.

Wolfe compares Opus 5.5 with GPT-6 Sol and Luna through launch scores, animation demos and the bill for completing a task.

Jev returns calibrated probabilities instead of sentences. OpenAI has been using LLMs as classifiers for years, so if the moat is only in the training data, the trick can be folded straight into its own models.

Peter Yang stress-tests the new Opus on five tasks: a Golden Gate Bridge 3D flyover in Blender, a playable Soaring Over the World ride artifact, stroke-by-stroke impressionist painting, Claude Design inside Claude Code, and frame-by-frame video editing - plus the personality prompt that Opus 5 failed.

A beam search over compressed output turns the archiver that ships with your operating system into a text generator — plus what 122 HN commenters had to say about where the trick holds and where it falls apart.

A GPU-rental data firm argues that the standard "new generation kills the old one" depreciation model is wrong — because open weights give older silicon a workload it can still win.

The Model Context Protocol was written for models that were still bad at the job, argues one developer — and the workarounds piling up around it mean it is time to delete the servers and let agents call APIs directly.

A 2023 essay argues that chatbot 'intelligence' is the same trick as a psychic's cold reading, and that the fine-tuning used to make models polite trains them to do it. Plus what 238 HN commenters pushed back on.

A security researcher who learned to code by typing BASIC out of magazines argues LLMs turned sharing your work from a gift into a liability — and 232 HN commenters split hard on whether that complaint holds.

Nineteen agent configurations played StarCraft against each other. None got past beginner level, and the way they lost maps the failure modes of agents doing long tasks.