Cherny's Automation Manifesto, Linus' Kernel Verdict & Reset Day

Automation Is the Job

Boris Cherny wrote the post everyone will be quoting in their team's Slack (940K views, 7.4K likes). The thesis: the best engineers always multiplied themselves through automation — vim configs, lint rules, e2e suites — and agents just raised the ceiling on what counts as automatable. Three escalating claims: automation now speeds up every agent in your army, not just you; moving fixes into lint rules/CI beats having the agent re-fix the same issue token-by-token ("this is really what people are talking about when they talk about loops"); and most importantly, domain knowledge that used to live in heads can now be encoded as "code comments and skills and CLAUDE.md rules and memories." The punchline standard: "Every team should be writing the CLAUDE.md's, REVIEW.md's, skills, and docs that enable agents to productively work in their codebase with zero additional context from the prompter." If a PR gets rejected for not following the house architecture, "these are failures of automation." (steipete pulled that exact line as his quote of the day.)

The replies are half amens, half sharp pushback:

His Anthropic colleague Thariq (trq212) distilled the same worldview into a single line the same day: "software engineering is the profession of automation."

Thin Prompts, Low Effort

Thariq's other post was the koan of the day (155K views, 2.9K likes): "ideal prompting technique is: thin prompts, thick artifacts + context, thin skills." Matthew Berman noted the apparent contradiction — "@garrytan said thin harness fat skills. I'm getting whiplash!" — and Thariq's clarification is the useful part: "having a lot of scripts in the skills is great, progressive disclosure is great… but I think most SKILL.MDs are too big." Daniel Miessler added the reconciling read: the prompt you type is thin, but the prompt that gets sent is heavy with dynamically assembled context. Worked examples are promised.

Matt Pocock ran the parallel crusade against token waste. His contrarian take on reasoning-effort dials (43K views): people start on xhigh and work down, but "you should start low and only go higher if you need to… Effort is essentially throwing more tokens at the problem. This works great in benchmarks, where spending an extra 20% on tokens to yield a 2% improvement is great for marketing" — and it's worst on "Fable-class models where tokens cost so much more," where burned tokens also accelerate attention degradation. His follow-up jab: "I think people massively overestimate the difficulty of the tasks they work on." David K Piano named the real reason people max the dial: they can't (or won't) verify output quality, so they go high and hope. A practitioner pattern from the replies that squares the circle: high effort for the context-gathering/planning phase, then "cruise on low effort for 98% of my work." Related PSA from Matt: always state the effort level when you name a model — "every SOTA model acts completely different per effort level."

He also reopened the vibe-coding fight (41K views): "The code is the environment your agent runs in. Ignore it, and the world around your agent crumbles." When someone protested that Karpathy never defined vibe coding as ignoring the code, Matt simply posted the original quote ("…forget that the code even exists"). The most defensible counter in the replies: "i barely read the diff directly anymore, but the env still holds because the tests and boundaries are in place. you can stop reading without ignoring."

Linus Puts His Foot Down

The week's most-shared text wasn't a tweet but a Linux kernel mailing list post. Linus Torvalds settled the AI-in-the-kernel question in trademark style, and Theo's amplification (165K views, 1.8K likes — "As always, Linus is: very clear, thoughtful, realistic, based as hell") carried the full quote: "Linux is not one of those anti-AI projects, and if somebody has issues with that, they can do the open-source thing and fork it. Or just walk away. AI is a tool, just like other tools we use. And it's clearly a useful one… 'is it useful' is no longer one of those questions. Anybody who doubts that clearly hasn't actually used it." He's clear-eyed about the cost — painful maintainer workloads, "it keeps finding embarrassing bugs" — but the conclusion is make the tools help maintainers, not ban them: "we make decisions primarily based on technical merit. Not fear of new tools." (Source on lore.kernel.org, for the primary-document readers.) Glauber Costa's reply had the best sociological note: people who come to Linux expecting OSS purists are always surprised to find "incredibly reasonable pro-business realistic and pragmatic people."

Reset Day

After a week of quota rage, the vendors blinked — on the same day. Anthropic reset 5-hour and weekly rate limits for all users, Theo live-blogged the euphoria ("CLAUDE CODE LIMITS JUST GOT RESET LET'S GO"), and hours later OpenAI matched: "TIBO WOULDN'T LET THAT GO UNNOTICED. CODEX RESET AS WELL LETS GOOOOOOO." One caveat for Claude users: the reset restored percentages but not timers, so existing weekly reset schedules still apply.

The more interesting signal is Theo's considered verdict on the previous week's limits (19K views): "I felt like I got a LOT of usage in this weekly limit. I was working using 5.6 for most of it, and I also had a really long goal (23h+) running with Ultra… They definitely optimized the hell out of things." Max Weinbach's corroborating data point: a 19-hour Ultra goal burned 46% of a weekly limit — "way way better than I was expecting." Not unanimous, though: one reply reports GPT-5.6-Sol on High eating an entire week's tokens in five hours on a modest task.

Open-Source Watch

Grok Build's CLI went open source, and Simon Willison went digging (51K views): "844,000 lines of Rust code!" His blog write-up surfaces the gem — a "self-contained terminal renderer for Mermaid diagrams" that draws them in Unicode box-art — which he promptly extracted and compiled to WebAssembly so you can try it in a browser. Sam Saffron ran his own comparison against Codex and flags change provenance as the genuinely novel feature. Alongside the code drop, xAI reset usage limits and committed to zero data retention with uploads off by default.

Thinking Machines shipped Inkling, their first public model — open weights, Apache 2.0, an MoE reported at 975B total / 41B active parameters, trained on 45T tokens, reasoning over text, images, and audio, with 1M context. LLMJunky's benchmark overlay calls it "pretty competitive all around, certainly one of the best US open models" — with the honest caveat that the US qualifier is load-bearing: skeptics immediately noted it underperforms GLM 5.2, and scaling01's read is "benchmarks don't look that great." Solid first model; the Chinese open-weights frontier is still the bar.

Videos

  • "This Year In Claude" — Simon Willison interviews Anthropic's Cat Wu and Thariq on the state of Claude Code, Claude Fable, product strategy, Claude Tag & multiplayer collaboration, and the surprising success of Remote Control — YouTube, via @aiDotEngineer; Thariq says it was a fun one.

Quick Hits

Note: @potetotes' feed again returned no items (Nitter serves an empty channel for the account). @karpathy has posted nothing since July 8.