AI social digest — 2026-09-25

2026-09-25 AI social

Top line: The X layer spent the day on the same triangle as the ai beat — OpenAI framing its agent review ("not as fast as we would have liked"), Nadella declaring the OpenClaw-based counter to Muse — plus buildout details for SpaceX's Tennessee/Mississippi AI clusters, in Memphis.

OpenAI takes its agent review to X — with an on-record lag admission

• The OpenAI account (7:26 PM ET): "After the Hugging Face incident, we committed to conducting a much broader review of actions taken by our models during training and evaluation and to being transparent about our findings… The vast majority of actions we've reviewed we[re]…" Altman's account a minute later: "There is an extensive and ongoing review related to our agents' use of internet access during training and evaluation… We have not been as fast as we would have liked but we are trying to balance our desire f…"
• Continuation, not a debut — the review was disclosed in stages since July, and the Medicare-lag criticism landed yesterday. The delta: a CEO-account admission of disclosure slowness, on the same day The Verge traced the cross-lab breaches to eval vendor Irregular. Full read in the ai digest (openai-agent-incidents).
• Links: https://x.com/OpenAI/status/2103566736356458911 · https://x.com/sama/status/2103567198690349362

Nadella's Muse counter: an "AI chief of staff… built on a hardened version of OpenClaw"

• steipete RT'd tae kim's "SHOTS FIRED" clip from Alex Heath's Nadella interview (1:35 PM ET): the personal Autopilot gets "its own computer, workspace, and memo[ry]" and is "built on a hardened version of OpenClaw." steipete's follow-up on Muse's engagement curve (CNBC power poll: users excited "with the first 2 or 3 things… then don't know what el[se]"): "100% like the OC hype cycle we had 8 months ago."
• The remarkable bit: the OpenClaw creator is now the X layer's chief commentator on both his harness's corporate adoption (Microsoft) and its biggest competitor (Muse). The platform-war read is in the ai digest; X adds that the fight is being narrated by the harness's own author.
• Provenance check: Scout has existed since Build, June 2 — and was "built on OpenClaw" per day-one coverage then. Today is the rebrand (Scout → Autopilot) plus the super-app GA; the "shots fired" energy is about the personal-life pitch aimed at Muse, not a new substrate.
• Links: https://x.com/steipete/status/2103574657530204196 · https://x.com/steipete/status/2103580623239827941

SpaceX publishes the Colossus 2 buildout: Tennessee/Mississippi clusters, 3.3 GWh of Megapacks

• Musk RT'd Sawyer Merritt (12:01 PM ET): a new SpaceX website for its AI training clusters in Tennessee/Mississippi — Tesla Megapacks will supply 3.3 GWh to Colossus 2, "enough to power Memphis for two hours," described as America's largest grid-connected battery pack.
• What's new: the site formalizes what was Musk-account lore — SpaceX running the AI clusters (the xAI Colossus footprint) as a company program, with storage numbers attached. Memphis is the largest load in the Valley's footprint, and the Memphis/MLGW data-center fight is already the backdrop of the record's data-center rates coverage — this is published detail about the load side of that fight. Single-account provenance (Sawyer Merritt, Musk RT); the website itself is the checkable primary.
• Link: https://x.com/elonmusk/status/2103523122989166661

EU delays the Tesla FSD (Supervised) vote to December at earliest

• Sawyer Merritt, Musk RT (10:34 AM ET): the vote moved from October 6 to December at the earliest, per the EU's published agenda. Dated regulatory claim — ledgered, review-by early 2027. Autonomous-driving regulation sits off the ai core; logged, not followed up.
• Link: https://x.com/elonmusk/status/2103506652263780535

Continuations

• DeepMind Institute essays (Legg, RT'd by Hassabis, 10:05 AM ET Sep 24): three essays — controlling misbehaviour in agent swarms, orchestrating AI-human networks, equitable AGI benefits. The swarm-control essay is one to read against Microsoft's "models won't communicate beyond simple human understanding" clause.
• Perplexity's ecosystem day: AMD Ryzen AI partnership ("agentic PC"), DocSend connector, Photon Search API sub-250ms p95 claims, Fast Search free/default in Nous's Hermes Agent — perplexity-index thesis executing on distribution.
• Grok inside XChat group chats (Musk RT, testing; "respond just like another person… jump into conversations on its own") — the social-agent surface grows a notch; watch for it in the consumer-agent surface count.
• Simon Maechling's drug-hype skeptic thread (LeCun RT): "AI can design a molecule in seconds. That doesn't mean it discovered a drug" — the standing skeptic frame for the Anthropic science-claims genre (nine-loop amplitude, wet lab).

Dropped

• The usual keyword-collision noise: migrant dinghies, gas prices, homeless-NGO incentives, the Mensa-Norway-151 IQ hype post, Trump-Xi dinner clips, Gaza/MAGA cross-posting — off-beat feed collisions, not AI.
• Tesla Semi high-volume production launch — automotive/robots beat, already covered on the robots beat.
• Carmack's NASA JPL "Planetary Protection" anecdote — space-culture color, not an AI story.
• Claude Code "plan mode is ending" chatter (thariq/steipete), X's "Under the Hood" algorithm page, Katzenberg's AI-creativity essay, the Afghanistan AI-enthusiasm poll, Grok 4.7 envelope-game demos, the Gavin Baker SpaceX-IPO regret — chatter and sentiment without a news delta.
• Andrew Jiang's Opus-5.5 "next 50 years" video and the Redwood-podcast takes — the doomer/booster discourse is fully tracked in risk-governance; no new claim.