Thursday’s AI news is about the cost of narrow agent work and the tests that tell you when a cheaper route is safe. Claude Haiku 5.5 leads because its pricing could change how often hosted agents delegate small tasks, though its coding …
AI-produced math proofs are arriving faster than people can check and explain them. Who gets to do that work, and how?
Four places Cole uses a fast decision model in his AI second brain: email triage, injection screening, news filtering, and deduplication—with a clear boundary for when an LLM still does better.
If customer traces are sensitive, seek limited authorized access, recruit cleared domain experts, or carefully redact; synthetic examples are a last resort.
Hamel ranks 28 evaluation practices: inspect real failures, use human-validated product-specific checks, and stop mistaking generic dashboard scores for progress.
A theory about Claude Code's suggested next prompts meets a direct correction from the team that built them.
Three skills for turning a spec into reviewed work, making PRs easier to judge, and learning from real agent sessions—plus a breaking glossary filename change.
Wednesday’s AI news is about what you can actually run or measure, and what still needs testing before you trust it. Mistral Large 4 leads because its API preview has independent performance and cost measurements, but promised open weights …
AI can turn an idea into an app quickly. The author asks what happens to the satisfaction of making it.
The 100-trace target is a starting cue, not a magic sample size; keep inspecting until examples stop teaching you something new.
Techdirt gathers the early Muse incidents, from Mac access and message privacy to risky agent permissions and a rushed launch.
The Linux maintainer separates useful AI-assisted bug finding from noisy security reports and asks who will actually review, fix, and deploy the patches.
Tuesday’s news is about provenance and whether agent systems can be tested before they do harm. OpenAI’s opt-in text watermark matters for attribution, but its own synonym-substitution test shows why detection cannot establish authorship. A …
An AI agent spotted a compromised Mac; the harder question is how to give agents useful access without handing them everything.
Wikimedia found unapproved agent edits and heavy automated traffic, but no compromised systems. Who pays to keep the public web usable?
Refresh AI eval datasets through error analysis; use product metrics, not static tests, for long-term comparison.
A document-agent answer can be wrong because OCR misread the source, not because the LLM hallucinated. Trace the failure back before changing the model.
A GNOME security maintainer makes the case for accepting AI-discovered bugs while recounting the triage burden that closed GNOME's own bounty program.
Monday’s useful releases cluster around local inference and the question of whether agents make reliable decisions once deployed. Strata leads with a runnable way to split a large MoE across consumer GPU, RAM, and CPU, though its speed …
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.
Speakeasy COO Hector Hernandez on what breaks when agent use scales: invisible tool connections, ambiguous identities, overbroad permissions, and approval fatigue.
Lauren Tan (poteto) on the agent infrastructure behind a reported 2,500 production PRs in one month: verification, deterministic tools, constrained codebases, coordinator agents, continuous gardening, and review by sampling.
A practical contract for long-running coding agents: define done, name the stops, preserve state, and demand evidence.
Granola co-founder Sam Stephenson on pre-meeting briefs, turning conversations into organizational memory, designing AI that lowers stress, and building an AI-native company without the 996 grind.
OpenAI’s former safety-report lead argues that perpetual sprints and fix-it-after-release safeguards are a dangerous fit for increasingly capable AI.
Sunday is quiet, with five useful releases and operational warnings centered on open infrastructure and real world security boundaries. Kolibri 1 leads: its Apache-2.0 English–German weights, vLLM integration, and unusually long context …
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.
Coding agents make it easy to spin up services that spend money while nobody is watching. Hard limits should be the default, not an optional warning email.
A 2-billion-token experiment found that cheap open models can handle routine coding—but model mix, provider capacity, and unbudgeted prototypes dominate the bill.
Saturday is quiet, with seven useful items centered on local inference, agent permissions, and operational boundaries. DwarfStar 4 leads because its small MIT-licensed C runtime makes current open-weight MoE models usable across Apple, …
Joseph Enochs on enterprise AI from the boardroom to production: measure the baton rather than busy teams, replace scattered pilots with shared foundations, constrain agent autonomy, and optimize complete business outcomes instead of local activity.
Writing an evaluation rubric before examining real traces can lock in the wrong criteria. Hamel Husain argues for open-ended error analysis first, then rubrics and reference answers grounded in observed failures.
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.
A young programmer can see AI's leverage and still mourn the craft, control, and coherent industry it may be replacing.
Hard Fork examines the central agent contradiction: persistence and broad permissions make assistants useful, but create the same autonomy, security, and accountability risks that companies claim they can police themselves.
Friday’s AI news is tooling-heavy, with durable execution, stateful policy, and better evaluation design linking the strongest releases and research. Pi 1.0 leads because its stable, MIT-licensed coding harness stays small while adding …
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.
Thursday’s AI news is unusually concrete: measured model economics, task-specific agent harnesses, and deployment evidence outweigh launch-day claims. Gemini 4 Argon leads because independent testing turns its million-token output window …
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.
A live attempt to let Codex buy a $5 course with Mercury’s agent card shows both sides of delegated payments: tightly scoped spending and enough approval friction to erase the convenience.
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.
Hamel Husain tours an open-source eval toolkit built with Shreya Shankar: discover real failure modes before writing tests, audit weak eval pipelines, calibrate LLM judges against human labels, and build review interfaces tailored to your traces.
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.
Pi reversed its opposition to MCP, but the useful part is the architecture behind the reversal: discover tools when needed, return structured data, and compose calls inside a sandbox.
Wednesday brought a busy OpenAI DevDay, but the durable theme is operational: controlling persistent agents, recovering from bad tool guidance, forecasting token costs, and testing claims reproducibly. GPT-6.1 Sol leads because its sharply …
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.
Tuesday’s useful signal is less Sonnet 5.5’s launch than its uneven cost behavior: Anthropic reports faster, cheaper work, while independent testing finds Max effort burns enough tokens to miss the cost frontier. Around it, CAVE-Bench shows …