DAILY DIGEST
Anthropic says you can change Claude Code’s behavior, customize its interface or add features with mods written in TypeScript—or ask Claude to build them. Mods ship as plugins installable with /plugin in the CLI or desktop app. The getting-started guide shows examples including context use, command previews and summaries of recent changes.
Look for a repeated question or risky command in your workflow and prototype a mod for it.
Andrej Karpathy recommends asking models for more than prose: constrained writing using ASD-STE100, diagrams, interactive HTML pages or bespoke explainer videos. He says the controlled language can improve readability and that generated web pages and videos are starting to work well. Video narration may require an audio API key or local alternative.
For a hard-to-review topic, ask for a diagram or small interactive explainer instead of another long answer.
Pi announced Pi 1.0 with Pi Durable. A post about the release describes codemode: instead of making many tool calls and placing raw results in context, the agent writes a script, runs the work and returns the answer. The post says this is faster, cheaper and keeps context cleaner; details on Pi Durable weren’t captured.
Try script-based tool work when repeated calls are flooding an agent’s context.
Cloudflare announced two trained decision models, Clef and Clef-flash. A post describing Clef says it uses a frozen Qwen model for one prefill pass, then a small schema head scores answer options in parallel without generating text. The poster claims 4× speed and 2× accuracy on some benchmarks; the models are described as open source.
Consider testing the released models for structured decisions, but verify performance on your own workload.
A post from a ChatGPT team member says a global reset for all paid accounts is planned for tomorrow at 10 a.m. Pacific. The author also says GPT-6.1 Sol returned to expected speeds after a load spike during its first two days.
If you rely on paid-account limits, check your account after the announced reset.
The New York Post reports that the FTC is investigating OpenAI, Anthropic and other AI firms, with officials saying the agency is preparing subpoena-like demands for documents and testimony. Former FTC technologist Neil Chilson notes an investigation establishes no wrongdoing: the agency could close it, seek a settlement or litigate.
Keep records of model evaluations and safeguards; scrutiny of testing practices is a live policy issue.
A captured account of a Senate hearing describes bipartisan discussion of rogue agents, lateral movement between data centers, developer liability and new legislation; that account is one observer’s assessment. Separately, a report describes a proposed federal prelaunch-testing bill, a California law creating third-party evaluators and a New York law due in January requiring lab registration and quarterly risk reports.
Track the state requirements and federal proposals if you build or deploy frontier AI products.
Updated 2026-10-03 10:55 UTC
A compact digest when Nick publishes. No extra noise.