Quick Hits

• Dario Amodei's new essay, "We Must Pace the Frontier," commits Anthropic to giving third-party evaluators permanent, employee-level access to its systems. Pushback came fast: @DavidSacks replied "go ahead" (via @levelsio's RT), and @sundeep, identified by @jun_song as an Nvidia VP, doubts "the entire world" will "align to stop innovating."
• @trydotworks argues a model reaches the frontier by excelling at any one of Quality, Cost, or Speed, citing Qwen 3.8 27b at 2,000 tokens per second and DeepSeek V4.1 Flash as profiles that OpenAI's and Anthropic's quality-only generalists don't cover.
• @rauchg demoed fx orchestrating subagents with different models and reasoning efforts, "Fable planning and Grok executing," with mid-task steering and no server-side routing required.
• @unclebobmartin had Grok build a Clojure Othello from a single prompt in roughly 40 minutes, then produce a layered object-oriented design report with a UML-style diagram in under a minute. His verdict: "These models are GOOD."
• An unverified rumor worth tracking: @kimmonismus relays leaker chatter that Google DeepMind has reached recursive self-improvement. Nothing in today's sources confirms it.

Anthropic asks the industry to slow down; the timeline declines

The day's biggest cluster orbits Amodei's essay. In the post quoted by @NetworkChuck, Amodei lays out a three-part plan for slowing down and says Anthropic is unilaterally taking the first step: evaluators get permanent, employee-level access to verify safety measures, report incidents, and assess alignment during training. @NetworkChuck's reaction is less policy than vertigo, noting it "isn't a movie plot," and closing with "I'm choosing to be optimistic" about global coordination.

The market-facing angle comes from @SamanthaLaDuc, whose "Market, show me you care" frames the essay as a repricing event. She quotes @AnatoliKopadze's dramatized reading, which claims the essay admits AI is already building the next generation of AI and warns of a rogue swarm causing hundreds of billions in damage within 6 to 12 months, with Elon Musk allegedly agreeing. Those specifics exist only in Kopadze's paraphrase, not in the quoted essay text, so treat them as one reader's interpretation.

The skepticism is blunt. @levelsio retweets @DavidSacks noting that Dario wrote we should "pace the frontier" and that Sam agreed, before offering his apparently permissive response: "go ahead." @jun_song supplies the context that @sundeep, who insists innovation won't stop worldwide, is a VP at Nvidia. And @kunchenguid delivers the day's gallows humor: quoting the StarCraft 2030 "DEFEND MANKIND" tease, he jokes that Blizzard is the first lab "to take real action in pacing their development."

Cutting against all the slowing-down talk, @kimmonismus surfaces rumors that Google DeepMind has reached RSI, citing leaker @lyraxana's cryptic "huge congRatulationS Indeed!" plus public signals: Google saying Demis Hassabis will focus his full attention on AGI, and Reuters reporting in August that Sergey Brin is directing resources toward RSI. All of it is rumor and the poster's synthesis, but it explains why nobody seems to be braking.

Open-weights discourse drifts toward bans and torrent fallbacks

@tszzl predicts open source models "will be banned before too long after some major disaster," hoping Kimi, DeepSeek, and peers keep building but keep weights behind monitored APIs. @gospaceport calls that "a dangerous take for openai, the USA and for freedom." Adjacent, @jaredctate asserts that NVDA has bought HuggingFace and expects censorship of open, uncensored models to accelerate, recommending "The Hugging Bay," the Pirate-Bay-style torrent index for LLM weights that @RoundtableSpace described. The acquisition is stated as fact by the poster and unverifiable from today's timeline-only sources; the ban talk is prediction, not policy.

Dispatch, don't poll: how people are actually running agents

The most practical thread of the day is orchestration. @pvncher shares a pattern for multi-agent work: have the planner dispatch tasks to another thread running the model of your choice, end its turn, and let the worker message back only when done. That avoids polling entirely, since the worker wakes the originator on its own. @daniel_mac8 calls Codex's ability to manage its own threads "massively underrated," suggesting the same idea is landing in shipped products.

@rauchg shows fx v0.0.9 doing this with mixed models: per-subagent model and reasoning effort, subagents that keep working while you steer and interrupt, mid-task feedback, and a Ctrl+P model switch that preserves your draft. Per @fazxes's release notes, it works with any model and any gateway, "no fancy server-side routing," just the harness doing the orchestrating.

The governance counterweight is @thdxr's: access makes everything easier, access creates risk, and humans blocked waiting on other humans is "10x worse now that we have agents." On the output side, @flaviocopes recommends concise output styles, hidden reasoning summaries, and short sticky output rules to stop Claude Code, Codex, and Cursor from writing too much. Capability evidence comes from @unclebobmartin's Grok-built Othello and @steveruizok's tribute to @rikarends, who is throwing agents at the entire Chromium codebase, currently eating 100GB of RAM ("Some optimisation required. Oops"). Steve's point: the antidote to AI-coding dread is watching someone aim agents at "ridiculously ambitious problems and having the most fun with it."

Speed and cost claim frontier status, and silicon backs them up

@trydotworks makes the day's spiciest model argument: excel at one of Quality, Cost, or Speed while holding baseline on the other two, and you're at the frontier. Qwen 3.8 27b at 2,000 tokens per second is "in its own class, and an entirely different way of working with agents," while DeepSeek V4.1 Flash (up to 400 tps) wins on cost-quality balance. The critique lands on OpenAI and Anthropic for "building the same generalist model over and over," with a call for domain-specific models that beat larger ones in narrow niches. @basecampbernie's field note, running that Qwen build on Cerebras, adds the catch: the bottleneck becomes reviewing a wall of output produced in seconds. Supporting the silicon angle, @bookwormengr boosts the underrated @siliconcodesign (under 1,500 followers), whose latest is an architecture breakdown of OpenAI's Jalapeno inference accelerator that goes past raw FLOPs into the surrounding network design.

Review tools, domain harnesses, and the humans still in the loop

@ChrisShort shares alibaba/open-code-review, a hybrid of deterministic pipelines and an LLM agent with precise line-level comments and fine-tuned rulesets for NPE, thread-safety, XSS, and SQL injection, compatible with OpenAI and Anthropic APIs. @RiverAi7z (posting in Chinese) recommends a review plugin from the author of pi, installable straight from git. @hwchase17 points to @sydneyrunkle's blog on building domain-specific harnesses, alongside @garrytan's aphorism: "Either you die a system of record or you live long enough to become a domain-specific harness."

On craft and careers: @jonathan_wilke celebrates a library of small UI animations as what separates a product from "AI slop," assuming you skip the purple gradients and border-everything look. @TheRealAdamG amplifies @colintjarvis on what makes a great OpenAI FDE: "always make the thing exist (and work) first," scale second, more domain experts, and success defined as replacing yourself in every customer problem. And @martin_casado retweets @mattyp's plan to speedrun a company with @poteto and @roshan_s, going from nothing to a business in 72 hours, streamed, next week.

Practical Takeaway

If you run multi-agent coding workflows, the actionable pattern is @pvncher's dispatch-and-wake loop: stop polling, let worker threads call back on completion, and assign planner and executor models separately the way fx and Codex threads now allow, trimming verbosity per @flaviocopes while you're at it. But @thdxr's post is the prerequisite: audit what your agents can touch, because every permission you withhold now blocks not just a teammate but the agent acting on their behalf. The question worth investigating this week is which of your approval gates become pointless, or newly dangerous, once an agent is the one waiting at them.