A Sate Padang vision test and a Malaysia capital question show why the free, private model needs grounding and validation before it leaves prototyping.

Trained on six PathMNIST tissues with a small KEEP adapter, it handled added labels without retraining, though a linear classifier hit 88% on familiar ones.

Jev needs no training and at most three examples per intent; the open rebuild needs labels. Both still fall into the none-of-the-above trap.

A software tester ran six testing skills on Haiku, Sonnet, and Opus; pushy descriptions added only 5 points, and near-miss requests never fired a skill.

A Gemini Flash-Lite step pulls skills, seniority, and Nigeria eligibility from 18 RemoteOK postings, with Python checks and dbt tests guarding every answer.

A guard on negative enrollment days looked correct while the function it called overbilled a customer 11 days, and an invariant test would have caught it.

For anyone running an inference server: a zero-variance control, five traffic traces, and an engine-vs-scheduler split show where the gains come from.

Studies of short stories, metaphors, and 486 participants point one way, an 800-person experiment points the other, and the author weighs both.

A fisheries estimator from the 1890s, tested on 294 CUAD contracts, landed within 2 points of true recall for most agents with a 10% hand audit.

Saving favorite podcast moments from Spotify's live playback, in an app built on Replit

Tracking fried food from meal photos with the language model built into macOS

Catching the silent failures in apps coding agents build, without reading the code

Trained on six PathMNIST tissues with a small KEEP adapter, it handled added labels without retraining, though a linear classifier hit 88% on familiar ones.

Jev needs no training and at most three examples per intent; the open rebuild needs labels. Both still fall into the none-of-the-above trap.

A software tester ran six testing skills on Haiku, Sonnet, and Opus; pushy descriptions added only 5 points, and near-miss requests never fired a skill.

A guard on negative enrollment days looked correct while the function it called overbilled a customer 11 days, and an invariant test would have caught it.

For anyone running an inference server: a zero-variance control, five traffic traces, and an engine-vs-scheduler split show where the gains come from.

A fisheries estimator from the 1890s, tested on 294 CUAD contracts, landed within 2 points of true recall for most agents with a 10% hand audit.

Beyond basic ReAct loops: six production-oriented architectures for building autonomous, reliable, and scalable LLM agents.

How statistical moments connect the mean, the variance, and higher powers of a distribution

First in a series on probabilistic forecasting for physical signals. Next: what happens when you roll the forecast forward more than one step.

When Codex is the right shape for the problem, when Claude Code is, and how I split 5 specialist agents between them on dense AI capacity work.

Why the skill-inflation panic is aimed at the wrong thing, and what it costs to make agent knowledge a build artifact instead of a file.

A real Weave project that regression-tests three OpenAI models against the exact reply format your app depends on.