AI social digest — 2026-09-23
Top line: The timeline mostly seconded the day's institutional news — open-weights chest-thumping over Xiaomi MiMo-V2.6, ChatGPT Voice's rollout, and the UNSC briefing's quotes — with two oddities: a claimed Astra "learned to drive a car" result, and a blog claim that Claude Code only reads AGENTS.md when telemetry is on.
Xiaomi MiMo-V2.6 Pro & Flash: open weights claim frontier-adjacent agent performance (continuation — efficient-tier, X layer)
What happened: Xiaomi MiMo announced MiMo-V2.6 Pro & Flash — "frontier intelligence, all the modalities, built in public," two omnimodal models advanced via scaled RL; Pro "performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks" (vendor claim, no named numbers in snippet). Amplified by Demis Hassabis: "Never bet against open weight models" — a notable source for the open-weights argument given DeepMind's closed posture. Perplexity's Aravind RT'd it too.
What's genuinely new: the efficient-tier race gains a consumer-electronics entrant claiming Opus-5-class agent performance at open weights — if the claim survives independent runs, the GLM/DeepSeek/Luna tier gets a fourth major. Hassabis's line is the same argument Chaumond and Dean Ball made against pacing (open weights can't be paced), now from a frontier CEO. Flag: everything is vendor-claimed; no AA/independent runs in this window.
Links: https://x.com/AravSrinivas/status/2102780186202972254
ChatGPT Voice rollout + the UNSC quotes on-platform (continuation — openai-bundling, risk-governance)
What happened: OpenAI's account shipped the Voice-with-connectors announcement (email/calendar/Slack, Astra/Sol/Luna, Work voice mode — filed in the ai digest). The UNSC layer on X was quote-amplification of Altman's lines ("we are not the heroes of the story," the unilateral-slowdown pledge) rather than new argument.
Links: https://x.com/OpenAI/status/2102808325742322002
Grok 4.7 "moves up in ranking" — but the underlying post is a leaderboard audit (continuation — agent-harnesses)
What happened: Musk amplified Zhuokai Zhao (SWE-Together leaderboard): after Grok 4.7 exposed weak spots, the team "audited every model we have run… for the same behavior, re-ran every trial that got through, and updated the rows" — a tool-call audit across all models. Net: Grok 4.7 up, some others down.
What's genuinely new: the benchmark-integrity genre inverted — instead of launch-day benchmark sniping, an auditor publicly re-ran and corrected rows because a new entrant exposed evaluation weak spots. Reads as legitimate hygiene, but note the sequence: the correction favors the loudest marketer's model, and Musk's framing ("moves up in ranking") strips the audit context.
Links: https://x.com/elonmusk/status/2102873022789283985
"Astra has randomly gained the ability to drive a car" — the day's wildest claim, thin verification (new — watch)
What happened: Benjamin Todd (80,000 Hours co-founder): "Astra has randomly gained the ability to drive a car, seemingly without specialist training" — citing DrivingBench (drivingbench.com), where "each model gets 3 attempts in the same continuous chat (in-context learning)." Steipete RT'd it into the OpenClaw circle.
What's genuinely new: if accurate, an emergent-capability datapoint on the RSI/warning side; but our record has nothing on DrivingBench (provenance, methodology, and whether "drive a car" means a simulator or a vehicle are all unestablished). Filed as a watch item only — no weight until the bench or an independent run is verified.
Links: https://x.com/steipete/status/2102890586428911733
Claude Code reads AGENTS.md only when telemetry is on — a claim worth verifying (new — watch)
What happened: A blog post (blog.szypowi.cz): "Claude Code reads AGENTS.md only when telemetry is on" — steipete's one-line RT: "Huh. There is a catch."
What's genuinely new: if true, this touches the agent-tooling layer directly. Single blog post, no corroboration in this window, mechanism unstated in the snippet — verify against Claude Code's actual behavior/changelog before treating as fact. Worth 10 minutes of local testing if the fleet cares.
Links: https://x.com/steipete/status/2102889319849713698 · https://blog.szypowi.cz/p/claude-code-reads-agents-md-only-when-telemetry-is-on/
Also
- Anthropic DRC Ebola response — CEPI/WHO AFRO/INRB using Claude against an unusual Ebola variant (Anthropic feature post) — the bio-for-good counterpoint to the biolab story in the ai digest. https://x.com/AnthropicAI/status/2102897863097545197
- claude.ai 3× faster in two weeks — ClaudeDevs with prompts/methods (the eval-the-thing-with-itself genre, Anthropic edition). https://x.com/trq212/status/2102841583158878500
- CAIS releases HLE-Diamond — year-long cleaned/refined HLE subset with Scale AI Labs; benchmark-hygiene layer, and the eval-infrastructure strand's third artifact this week (after EvalEval/AISI and MentalHealthBench). https://x.com/steipete/status/2102815369211924607
- GLiClass — "open-source Jev, before Jev existed" — lands in OpenClaw: run locally via ONNX for lightweight agent decisions instead of a full LLM call (Knowledgator, RT by OpenClaw circle ~3:32 PM ET) — decision-models thread: an open-source predecessor getting harness-native distribution, same week Nokia open-sourced AnyJev. https://x.com/steipete
- LeCun's NYU CILVR talk: "World Models: Enabling the Next AI Revolution" — the world-models watch thread's academic-layer anchor as the ICWM deadline approaches. https://x.com/ylecun/status/2102878714023592161
- crabbox now runs on boxd (steipete's repo-command runner on the boxd provider) — OpenClaw-ecosystem dev tooling.
Dropped
- Elon engagement-farm pile (~20 tweets, mostly
filtered): Grok comparisons, grass-touching memes, "customer is always right" quips, Astra game demos — noise. - "This is a new type of post we're trying" (OpenAI technique-sharing post) — content marketing, no new fact.
- Jeffrey Katzenberg "thoughtful piece" — unclustered singleton, no corroboration in window.
- Luiz André Barroso tribute (Lucas Bleme) — memorial, off-beat.
- muse-for-the-Sweden-friends anecdote (gabriel) — adoption color, single-account.
- Nikita Bier "customer is always right" — engagement bait.