AI social digest — 2026-09-08
Top line: The Navier–Stokes attribution fight resolved (mostly) on-platform today: OpenAI congratulated mathematicians Alpöge and Buckmaster, said its researchers and agents never saw their work, and — via Sama — revealed the solve took 88 hours and ~10,000 coordinated agents. Elsewhere: Astra is now fully rolled out to every paid tier, and OpenClaw shipped 2026.9.3, its third release in nine days.
Navier–Stokes attribution fight resolves on-platform — with a new fact: 88 hours, ~10,000 agents (continuation, lead)
What happened: The morning's Engadget/Verge "drama" (see today's ai digest) played out on X through the afternoon ET. OpenAI (1:23 PM ET): "We congratulate Levent Alpöge and Tristan Buckmaster on their remarkable mathematical work. We (the researchers and the agents) did not see any of their work through any [channel]..." — the attribution-credit statement. Sam Altman (2:02 PM ET): spent the weekend with the team, "acted with integrity and generosity throughout. Initially we believed the oth[er party]..." And the new technical detail, via Altman's RT of OpenAI (1:20 PM ET): the solution was produced in 88 hours by the internal model group using ~10,000 coordinated agents, with training ongoing.
What's actually new: the fight's shape. This is now publicly an OpenAI-vs-(prior mathematicians) credit dispute, resolved by OpenAI asserting independent discovery and by naming the two mathematicians it's accused of overlooking. And the 88-hours/10,000-agents figure is the first concrete scale claim — which, note, is the same order of magnitude as the Hugging Face swarm from the July incidents thread, this time aimed at a math problem instead of a wiki. Whether the "coordinated agents" framing gets regulatory attention is the interesting overlap between today's two big OpenAI stories.
Sentiment check: Trapit Bansal (ex-OpenAI RL): "the first successful achievement of Recursive Self Improvement" — the hype pole. No independent verification of the proof anywhere in the thread.
Links: - OpenAI congratulation: https://x.com/OpenAI/status/2097375276384567642 - Altman on the team: https://x.com/sama/status/2097385167002415140 - 88 hours / 10k agents: https://x.com/sama/status/2097381051173913055
Astra rollout completes; the vendor-eval war escalates (continuation)
What happened: OpenAI (5:06 PM ET): Astra is "fully rolled out to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work." Two vendor-adjacent eval claims landed alongside: Andon Labs (1:33 PM ET) — "the biggest jump in Vending-Bench history," Astra "better at making money and more ethical" than Claude Fable 5.1; and Noam Brown (1:25 PM ET) defending cost — "yes, this result cost millions of dollars, but remember when OpenAI announced o3 it cost..." with ARC-AGI-1 cost-per-task comparisons.
What's actually new: rollout completion (the timeline closes: launched Sep 3, full tier coverage by Sep 8) and the eval war going from benchmarks to ethics-and-revenue framings. Andon Labs is the same shop whose RT claims we flagged yesterday as vendor-adjacent; "more ethical than Fable" is a new and spicier genre of it. None of it is independent.
Also: Sama announced a GPT-6 community event in San Francisco on September 16 (event, not promise — noted for the calendar) and a modest one for honesty: "I don't think [Images 2.5] can solve super difficult math problems."
Links: - Rollout complete: https://x.com/OpenAI/status/2097431322117476423 - Andon Labs Vending-Bench: https://x.com/sama/status/2097394039557628115 - Noam Brown on cost: https://x.com/sama/status/2097449884530131057 - SF event Sep 16: https://x.com/sama/status/2097404861642137851
OpenClaw 2026.9.3 — third release in nine days (continuation)
What happened: OpenClaw shipped v2026.9.3 (12:36 PM ET, amplified by steipete): updates recover cleanly, faster session reconnection, live browser-automation watching, meeting-transcript search, background run output.
What's actually new: the cadence more than the features — 2.0 (Aug 31) → 2026.9.2 (Sep 5) → 2026.9.3 (Sep 8). "Updates recover cleanly" reads as the response to the 2.0 upgrade pain documented earlier in this record (sessions→SQLite migration); whether the Gemini-embedding memory-sync bug (open since Sep 1) is fixed in this train is worth checking before relying on the upgrade.
Links: - Release: https://x.com/openclaw/status/2097363436028191021
One-liners
- Perplexity on Blackwells (AravSrinivas, 1:10 PM ET): "started serving a lot of our inference on NVLink Blackwells. Vera Rubins next!" — infra-detail flex; Vera Rubin as the next-generation NVIDIA platform name worth noting.
- X re-hires a Head of Safety (Musk RT, 3:07 PM ET): Michael O'Herlihy, "glad to be back." Given the platform's role in the agent-coordination saga, worth tracking whether safety staffing is real or decorative.
- Cybercab to Europe "hopefully soon" (Musk, 11:55 AM ET): replying to a glowing Austin robotaxi rider review. Musk claim, no regulator, no date — same unverified-till-paperwork rule as the FSD/Slovenia item.
- "YC companies are a bit different now" / demo-day ARR (nichochar, evening): accelerator sentiment, not news.
Dropped, and why
- ~35 unclustered/filtered tweets — Musk political RTs (Sowell, Pratt, UK-migration, "far-right" discourse — not AI), X self-promo ("#1 news source"), Cybercab "rave cave," Rob2628's 350-mile FSD road trip, astra-robot-painting demo, "people are excited about capabilities that already existed" meta-observation. Standing rule: all-tweet clusters are sentiment, not events.
- One miscategorised item, folded up: OpenAI's Alpöge/Buckmaster congratulation tweet was initially misclassified; it's the attribution story's key primary source, covered above.
- von der Leyen satellite-connectivity RT — genuinely EU-space ("new cooperation on satellite connectivity... North Greenland"); belongs on the space beat rather than counted as AI-social signal.