Quick Hits

• Per @arcprize (quoted by @theo), GPT-6 Astra scores 63% on ARC-AGI-3, 99% with a new provider adapter harness, and surpasses human performance on 96% of levels. @theo, who says he "never thought we would even break like 30%" on a bench he calls almost malicious, considers it saturated.
• @thsottiaux (retweeted by @romainhuet) says OpenAI will bank one usage reset for every day paid ChatGPT users lack Astra access, starting today, a policy landing amid growing complaints about hype outrunning availability.
• @andykonwinski says the Marin team is about 15% through training the largest fully open model yet: 535B parameters (23B active) over 18T tokens, with data, logs, and decisions all public, on Coreweave compute funded by the Jen-Hsun and Lori Huang Foundation.
• @finkd (Mark Zuckerberg) rolled out Muse Spark 1.3, claiming the biggest jump yet on coding and agentic work at performance "almost too cheap to meter," with open weights releases teased next.
• @penlume posted an integer he says divides RSA-260. If it verifies, it's landmark cryptography news; for now it's a single unverified post.

Astra posts record benchmark numbers while early users demo 3D worlds

The Astra posts split between record scores and creative demos, nearly all from people with early or insider access. @theo relayed the @arcprize announcement, which also claims Astra "builds the most precise symbolic model of novel environments we've seen." On the demo side, @skirano says he has early access to GPT-6 and that it "will completely shatter your understanding of what's possible"; his first test is a 3D model and animation generated in code from an image, the start of a six-part thread.

The 3D theme repeats. @Dimillian (quoted by @pvncher) calls Astra "very good at 3D modeling" and describes building the launch post's demo house from a Blender scene into a walkable Unreal Engine 5 experience. @pvncher says he and Thomas have been pushing Astra on 3D asset assembly and modeling, something he would not have believed an LLM could do three months ago. @mattshumer_ separately claims GPT-6 Astra built a Manhattan world in Unreal Engine over the course of a week, going "street by street."

The hype runs into an access wall

The day's sharpest tension is between celebration and lockout. @thorstenball, quoting @petergyang, highlights the contradiction: @petergyang lives in Codex and calls it the best software shipped in the past five years, but says influencers posting non-stop about Astra while paid users cannot get access is "pretty rough," and that he's back to "good ol' Sol and what's left of my Fable limits." @thorstenball adds his own read: the hype choir starts on cue, the "well, ..." caveats whisper in the replies, nobody else has access, and "it feels like a play we all pretend isn't a play."

The banked-reset policy from @thsottiaux reads as the official countermove, though it compensates for wait time rather than shortening it. Subscription strain isn't OpenAI-only: @Shpigford shares the top of his CLAUDE.md file for routing work between Opus and Sonnet to stretch limits, after @johnwebber noted that Fable "chews through a 5hr limit fast."

Open weights gather steam: Marin's 535B run, Muse Spark 1.3, and Base Labs

Three separate posts push the openness story forward. @finkd says Muse Spark 1.3 is available now in Muse Code and the API, and teases what's next: a cryptic 🍉 project and Muse Spark open weights "coming soon." @andykonwinski argues "nobody has been this transparent in a hero run before" on Marin, urging people to follow the run live; @percyliang (quoted) has it at 13% through and thanks @JensenHuang for supporting open models.

The biggest structural bet is Base Labs. @oneill_c announces the Baseten-backed lab with one rule: publish everything, including failures. The agenda is continual learning (which he calls "several problems wearing one name"), open RL environments and training data, a safety stack for open deployments, and cheaper serving. His framing is that closed labs aren't villains but can't do slow public science when the job is best model by end of quarter; his own path ran through parsedlabs, acquired by Baseten. The @baselabs announcement adds the BaseHub Data Foundry and "post-post training" of open models, and both posts say the lab is hiring.

Agent plumbing: manager loops, event logs, hooks, and background Linux control

The most concrete engineering content of the day is about controlling agents rather than marveling at them. @mattshumer_ details his "Manager Loop" for long-horizon Astra builds: a manager agent builds a massive checklist, splits it into phases, and spawns a separate implementer agent; both run in /goal mode and message each other until every phase completes. Two details he flags: asking for each phase done "extremely well" rather than "perfectly" avoided the model bogging down in minutiae, and having the implementer maintain an HTML checklist page with a counter and chart forced it to notice stalls. He also ran 96 sub-agents at a time.

On the framework side, @arjunkmrm (shared by @yoheinakajima) introduces Tardigrade, which treats the agent harness as typed state machines over an immutable event log, "think react for harness," built on Effect-TS and self-hostable as durable objects on Cloudflare or celld. @yoheinakajima notes it's another framework deriving agent state from an immutable log, @activegraphai style. @bcherny surfaces @ClaudeDevs' Function Hooks proposal for extending Claude Code; it hasn't shipped, feedback is on GitHub, and he calls it "a little crazy, and very exciting."

Desktop control advances too. @trycua (quoted by @LLMJunky) demoed an agent working across apps in the background with a "true multi-synthetic pointer" at the Omarchy SF Meetup, and wants Hyprland-native support for isolated background input. @LLMJunky calls it solving background computer use on Linux, credits @francedot as the best in the world at this, and predicts a major lab acquires them within three months, his prediction, not a fact. Elsewhere, @sawyerhood points to @get_bb_app's Kernel Browser plugin, which lets agents create and control remote browsers with an in-app preview, proving how extensible bb is. @dhh notes @ThePrimeagen is now spearheading automated QA for Omarchy.

One self-reported enterprise receipt, via @alex_prompter recommending @mardehaym: a delivery-orchestration core that FedEx estimated needed 7 to 8 months was rebuilt in 3.5 months by 2 engineers, 122 merged PRs in the first 90 days, production by month 6, using a knowledge-graph-first onboarding step, a six-step define-to-document loop, roughly 90% agent-generated code behind a human "V.U.E." review gate, and about $200 per developer per month in AI compute.

Cheap and off the frontier: transcription, tiny models, and tooling odds and ends

@OpenRouter (quoted by @altryne) says Microsoft MAI-Transcribe 2 is live, ranked #1 on FLEURS, with 60 languages, automatic language ID, code switching, diarization, and word-level timestamps. @altryne claims it transcribed over 90 minutes of audio with diarization in 15 seconds, and says people are still defaulting to Whisper-3-large despite worse accuracy. @liquidai retweets @IQReactorAI on 500M-1.