first-hand human knowledge

Attention is all that matters.

We read the podcasts, essays and interviews — and hand you the arguments: who claims what, who rebuts, and the original voice one click away.

Banning general-purpose computation in favor of specialized anal AI agents' success rate on end-to-end, economically valuable fre Slowing or banning the open-source AI model ecosystem is an effe Frontier AI agents' reliability on long-horizon, multi-step, mul

Popular moments

This week This month All time

all topics benchmarks · 17 claude-code · 9 automation · 7 sqlite-utils · 7 policy · 7 agents · 7

Mechanism◆Article

Newer Anthropic models have been specifically trained, presumably via reinforcement learning, to better use the edit tools baked into Claude Code, and this makes them more likely to misuse the differently-shaped custom edit tools of other coding harnesses like Pi.

Armin's theory for the regression: Anthropic has likely RL-trained recent models heavily on Claude Code's own edit tool, which inadvertently degrades their ability to correctly use differently-designed edit tools in third-party harnesses like Pi.

Armin Ronacher · Simon Willison's WeblogExtends · 2 Examples · 1 Supports · 2

Claim◆Article

Fable, and to a certain extent Opus, work best when allowed to use their own judgement about how to do a task rather than being given dictated, rigid instructions.

A tip from the Claude Code team: let Fable and Opus decide how to approach a task themselves instead of prescribing exact rules.

Cat Wu and Thariq Shihipar · Simon Willison's WeblogExamples · 3

Claim◆Article

To avoid burning through valuable Fable tokens before the price increase, tell Fable to hand off smaller tasks to other models, letting it use its own judgement about which model to use.

Jesse Vincent's tip: have Fable judge when a task is small enough to route to a cheaper model, saving tokens ahead of an upcoming price rise.

Jesse Vincent · Simon Willison's Weblog

MechanismArticle

Implementation work rarely needs the top-tier model, so coding tasks should be spawned as subagents with a model override (Sonnet for substantive implementation, Haiku for trivial/mechanical edits), while judgment, review, and synthesis stay with the main loop.

The saved memory file spells out the rationale and rule: hand off routine coding to cheaper subagent models, but keep judgment-heavy work (design, review, synthesis) in the main model.

Claude · Simon Willison's Weblog

ExampleArticle

Rather than giving Fable an explicit rule like 'only use automated testing for larger features, don't update and run tests for small copy or design changes,' it's better to tell Fable to use its own judgement about when to write tests.

Testing was the concrete example: instead of hard-coding a rule about when to test, just tell Fable to judge for itself.

Cat Wu and Thariq Shihipar · Simon Willison's Weblog

MechanismArticle

Claude Code's much higher token usage compared to Codex comes from its harness re-feeding a large accumulated context history into the model at each turn, not from generating longer outputs.

Investigating why Claude Code burns far more tokens than Codex or Qwen-Code, Raschka found the difference is almost entirely on the input side—one run used 578k input tokens versus just 4.5k output tokens—suggesting Claude's harness re-accumulates prior context every turn.

Sebastian Raschka · Ahead of AI

DefinitionArticle

Claude Tag is Claude Code made multiplayer, async, and proactive across the whole team, while Claude Code itself remains the fastest mode for solo, synchronous work.

Anthropic's Claude Code team positions Claude Tag as the multiplayer, asynchronous, proactive counterpart to Claude Code, which stays the fastest tool for solo synchronous coding.

Anthropic Claude Code team (@ClaudeDevs) · Latent SpaceSupports · 1 Rebuttals · 2

FactArticle

At Anthropic, traditional product requirement documents have been replaced by prototypes, and the vast majority of code — including nearly all of Claude Code itself — is now generated by Claude.

Boris Cherny revealed that Anthropic has largely abandoned PRDs in favor of prototypes, with Claude generating close to 100% of Claude Code and 70-90% of all internal code.

Boris Cherny · The Pragmatic EngineerSupports · 1 Examples · 1 Extends · 1

AnecdoteArticle

GLM-5.2 is the first open-weight model that feels right as a general coding agent in real harness use, despite minor integration bugs like image inputs bricking a Fireworks API session.

Testing GLM-5.2 hands-on via Fireworks' API inside Claude Code to build course content, the author found setup easy and capabilities immediately convincing, aside from a bug where sending images would brick the API session.

Nathan Lambert · InterconnectsSupports · 1