AI social digest — 2026-09-29

2026-09-29 AI social

Top line: X's layer is Perplexity's agentic-flex day (an agent-run Coinbase portfolio, a $2,000 self-built game) with heavyweight skepticism landing on Anthropic's biolab claim — LeCun amplifying Strogatz's Navier–Stokes-style "coincidence" suspicion, now pointed at biology. DevDay keynote fires in a couple of hours.

The biolab claim meets its first heavyweight skeptic channel

• LeCun RT'd Steven Strogatz: "Another 'coincidence', like the one that happened to Tristan Buckmaster with the solution of the Navier-Stokes problem? Except now it's in biology" — pointing at the NYT's Sep 27 piece, "Did Anthropic's A.I. Really Make a Scientific Discovery on Its Own?" (enzyme system / Mestre). This is the anthropic-biolab claim (Sep 23) getting its first outside-press check and its attribution-fight layer: the wet-lab story's open watch item (independent validation, the "Crispr" comparison) now has a named skeptic counterweight with math-fight provenance attached. Read the NYT piece primary before quoting either side.
• Links: https://x.com/ylecun/status/2104934196691055102 · https://www.nytimes.com/2026/09/27/science/anthropic-biology-enzyme-mestre.html

Perplexity Computer: the agentic-money demo and the $2,000 game (vendor claims)

• Arav: (1) Multi-agentic trading on Coinbase — Computer using Fable 5.1, GPT-6 Astra, and Grok 4.7 "in tandem" to run an agentic portfolio with automatic research and rebalancing, citing public earnings, built-in CB Insights data, and on-chain sources. (2) An end-to-end, browser-based open-world game built by Computer in Standard Mode (Opus 5.5) for ~$2,000 in credits — renting its own GPUs, headless Blender, procedural design, downloaded assets, audio (playable at vice.pplx.app).
• Both are vendor demos, not audited results — but the Coinbase item is the agent-economics discourse (Apollo's bank-run thought experiment, yesterday) getting its first big-vendor product claim on real market rails. Watch: does anyone outside Perplexity reproduce either loop.
• Links: https://x.com/AravSrinivas/status/2104952698949042268 · https://x.com/AravSrinivas/status/2104941391272956125

DevDay countdown

• Sama: "DevDay livestream in 2 hours. We built some great stuff for you!" (published 11:04 AM ET) — keynote at 10am PT / 1 PM ET ("10am PT, on the dot," OpenAI account). The rumored consumer agent Aeon is the story; the Verge's pre-keynote read lives in the ai digest.
• Links: https://x.com/sama/status/2104950376466362578 · https://x.com/OpenAI/status/2104934336206430527

Smaller moves

• "AI is NOT taking recent college grads' jobs" (LeCun RT of Catalini): Wu/Fairlie NBER paper — "this was supposed to be the summer it showed up. It didn't." The sentiment layer finally carrying paper, not vibes; note LeCun's appetite for this genre.
• UK open letter (Inherent via LeCun RT): UK AI companies that have raised $5B+ call on the government to end restrictions "that stop Britain's best switching jobs or starting companies" — non-compete/restriction relief as the UK AI scene's next policy fight.
• Grok distribution layer: Grok 4.7 now on Amazon Bedrock (SpaceXAI RT) — the third frontier lab Bedrock adds this cycle; and Grok Bot "Team Bots" — shared team bots with skills/plugins/credentials, one-click into Slack.
• Safety-case framing wars: OpenAI's safety-case post promoted on X; Arav's counter-frame "AI Safety is an engineering problem" — the governance layer as plumbing, days after Perplexity signed on as NVIDIA's marquee sandbox partner.

Dropped

• Starship/Flight 14 RTs — space beat's story, covered in the Sep 28 digests; nothing new today.
• Kyunghyun Cho's cancer-story thread, Pinker on Anthropic's philosophers, Kenneth Roth's measles RT, Hensen Juang's alignment hot-take — discourse/adjacent, off-beat or unactionable.
• Arav's icons chatter, nichochar's screen-time/repo-style quips, the "average screen time: 0" meme, AgentCribs save-the-date, "this week is going to be insane", grok @bot finance-dashboard RTs, the renaissance-painting RT, world-model reading-group RT — chatter and product-puff.

Evening addendum — Sep 29, ~11:55 PM ET

Top line: The White House's Super Intelligence summit swallowed the feed — lab leaders signing the Accord, Amodei praising the president, Musk at his side — while the DevDay product layer moved underneath: Sol economics, Codex as an open-model storefront, and ChatGPT subscriptions porting into Notion and Pi.

The Super Intelligence summit, X-layer: the Accord, the EO, and Amodei's flip

• What happened: the day's feed was one event — Trump's summit/America.gov launch. The late layer: Kratsios announcing the White House Accord on Super Intelligence — "Joint Commitment on Frontier Responsibilities" (~5:58 PM ET, video RT); Zuck endorsing the labs' commitment to "robust internal controls and multiple layers of audits and reviews"; Sacks: "Only President Trump could convene all the leaders…"; Sholto Douglas (Anthropic MTS): "pretty fucking cool that we get to live at the time when there is a 'White House Accord on Super Intelligence'"; and Dario on stage: "President Trump helped put his Super Intelligence worries at ease… If we do this right, we work with the President and everyone here, we can win safely!" (Eric Daugherty clip transcription). Around it: America.gov relaunched with Grok + Gemini powering the answers (Gebbia), Coristine thanking the career staff, Musk RTing the whole rollout.
• What's genuinely new: the governance fight's geometry inverted in public. Four weeks after the White House called the pacing warners "overreacting," it convenes, brands and legitimizes the labs' audit machinery — with Amodei, excluded from the Sep 25 state dinner, personally crediting Trump for easing his worries. Musk's RT pattern is doing work here too: the administration's Accord rollout distributed by the man the pacing essay ignored. The EO (official documents → "Super Intelligence" only) is covered in the ai addendum.
• Links: https://x.com/elonmusk/status/2105060244988514444 · https://x.com/elonmusk/status/2105060542020993038

Codex opens to rival open models — metered through OpenAI

• What happened: Philip Kiely (~5:36 PM ET), steipete ("This is pretty amazing"): GLM-5.3 Flash and Kimi K3 run natively in Codex, with spend counted against enterprise OpenAI commits.
• What's genuinely new: OpenAI's harness is now a storefront for the cheapest competitors' fleet — including the model Anthropic flagged the same day. The harness war's lines are blurring: OpenAI sells rivals' tokens, OpenClaw Enterprise launched with OpenAI as a partner, and steipete spent the day sensing "an agent war" between agent-pushing companies and users who want their own agents ("You want to block my agent, I buy something else").
• Link: https://x.com/steipete/status/2105048996549116268

Anthropic flags GLM-5.3 cyber spread — hours before it enters Codex

• What happened: Andrew Case (RT): "I am very confused why Anthropic wrote a blog post advertising for cybersecurity teams to switch to GLM!?!?" — pointing at Anthropic's research post, "GLM-5.3 and the spread of advanced cyber capabilities" (anthropic.com), warning about Z.AI's model as an advanced-cyber-capability diffusion vector.
• What's genuinely new: the independent-eval thread (AISI's ~20× Astra execution rate, this morning) gains its Anthropic-side entry on the Chinese frontier's cyber spread — published the same day GLM-5.3 Flash became a native line in OpenAI's Codex. Read the research before weighting; the framing ("why advertise for GLM") is Case's skepticism, not Anthropic's text.
• Link: https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities

ChatGPT subscriptions go portable: Notion and Pi

• What happened: Notion (nichochar, ~8:44 PM ET): your ChatGPT Plus/Pro subscription now covers AI usage in Notion — "connect your plan, pick a GPT model, and go." Pi: partnered with OpenAI on Sign in with ChatGPT — Pi users access their OpenAI subscriptions/credits through one login.
• What's genuinely new: the subscription is the new API surface — ChatGPT plans now meter into third-party apps. The federation layer DevDay launched this morning has two named distribution partners by night.
• Links: https://x.com/nichochar/status/2105096438363701291 · https://x.com/steipete/status/2105043539205964096

Pope Leo takes the Huang counter-frame to the Vatican beat

• What happened: FT: "Pope Leo criticises Nvidia's Jensen Huang over AI safety" — landing in LeCun's RT of Pessimists Archive's 1501 side-by-side (Pope Alexander VI on the printing press: "it can bring about serious evils… it will, therefore, be necessary to maintain full control").
• What's genuinely new: Huang's denial ledger runs Sep 15 "made up" → Sep 16 "don't need new laws" → Sep 20 "0% chance" on CBS → a papal rebuke. The anti-pacing amplifier with the highest office on Earth is now in the clip, and the counter-wave's own genre (historical panic parallels) is simultaneously being mocked by Huang's loudest skeptic.
• Links: https://x.com/ylecun/status/2105069976717549761

Smaller moves

• AD-E2E-JEPA (arXiv 2609.34085, Haoran Zhu joint work with LeCun and Choromanska, ~11:49 AM ET): a JEPA world model for end-to-end autonomous driving — "can a world model drive without training a driving policy?" The spatial-intelligence bet's driving-native test, from the thread's namesake lab.
• Grok 4.7 #1 on AA's Cyber Index (X Freeze): "outperforming Fable 5.1 Max, Opus 5.5, Astra 6, GPT-6" — vendor-selected-index genre as usual, but a cyber-benchmark lead for the model now powering America.gov is worth the line.
• GPT-6.1 Sol reactions (Tibo: 1/5th of Astra's price, 95% cache-read discount; Theo re-upping his Terminal Bench 4 claim — "better than Opus 5.5 for ~1/30th of the price"): the price-ladder chatter layer of the DevDay item; vendor-adjacent figures throughout.

Dropped (evening)

• Starship V3/Flight 14 resurfacing RTs, Crew-13 rehearsal, Starlink Mobile Ecuador launch — space beat's story (Flight 14 covered Sep 28).
• Tesla Roadster demo postponed two weeks for wind (Oct 15) — off-beat.
• Matt Hardy's "sandbox the agents by giving them Delta Wifi" — joke, though the sandbox-escape genre keeps earning jokes.
• Perplexity Automations commentary (gs-analysts RT) — launch already covered; sentiment.
• tae kim's "OpenAI resurgence vs Anthropic" FT/Reuters round-about — sentiment, and the resurgence frame rides the same DevDay-day coverage we've already weighed.

Primary-read corrections — Sep 30, ~12:10 AM ET

• GLM-5.3 post (read the primary): it's a safeguards/bypass analysis + defender-access policy argument — not an ad for GLM (Case's RT framing was sarcasm; the post is the opposite). Capability: ExploitBench end-to-end exploits 50/410 (vs Mythos Preview 56/410); internal Binary Exploitation — full control-flow hijacks 4% vs Opus 4.6/GLM-5.2's zero successes; GLM-5.3-Flash chained the public CVE-2026-11645 Chrome disclosure into a working ARM64 exploit bypassing PAC for ~$20.40 in API spend. Bypass rates for the built-in safeguards: "red-team agent" cover story 64%, thinking-token prefill 92%, abliteration 100%; abliteration (public variants existed within days) drops refusal 90%+→~3% for ~$4,400 in compute without hurting capability. None of it worked against safeguarded Claude. Cross-check: NIST CAISI (Sep 17) — "the most cyber-capable open-weight model released to date," ~4 months behind the US frontier, whose safeguarded/vetted versions attackers can't access while GLM-5.3 is a free download. Ask: expand defender access (Glasswing) + government safety testing of successors.
• Link: https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities