AI social digest — 2026-09-11
Top line: The money conversation went vertical — Jeff Dean's Discovery Loop is reportedly back in market seeking ~$50B, five weeks after launching at ~$10B, and consumer-AI app Instinct is seeking $1B at $10B with its compute costs unresolved. Plus: the Anthropic resignation got a name and a counter-narrative, and Coinbase's own telemetry printed the first market-share read of agentic trading.
Discovery Loop seeks ~$50B, weeks after $10B (continuation, lead)
What happened: Jeff Dean is raising again for Discovery Loop, seeking a valuation of around $50B — weeks after the round at $1B / ~$10B valuation (Business Insider, Sep 11 6:42 PM ET; Reuters carried it as a RANSAC item; nichochar flagged via Techmeme at 11:19 PM ET).
What's actually new: the ask and the velocity. The venture was a name on a departure list five weeks ago (Aug 5: Dean leaves Google, Hassabis elevated); it's now pricing at $50B without a product on our record. Pairs with Instinct (personal-AI app) seeking $1B at a $10B valuation (The Information via Sheel Mohnot; nichochar's skeptical thread: "the valuation numbers seem absolutely insane if the only thing they're doing is a good harness and free tokens" — Instinct raised $350M to date, mostly on open-source models, wants to own chips and data centers). The ex-DeepMind wave and the consumer-agent wave are now bidding against each other for the same capital.
Why it matters: this is the funding side of the deepmind-exodus thread — every departure is now a nine- or ten-figure vehicle. Note the shared structure with the robotics public-markets thread: valuations set by narrative before filings exist.
Links: https://www.businessinsider.com/jeff-deans-startup-discovery-loop-is-eyeing-a-valuation-2026-9 · https://x.com/nichochar/status/2098612531585757468 · https://x.com/nichochar/status/2098243134300012600
The Anthropic resignation gets a name — and a counter-narrative (continuation)
What happened: The Sep 9 resignation over the "superhuman systems" race now has a name: Jacob Coxon. The doomer-sphere response (Jeffries thread + Schachtel, amplified by LeCun's RTs, Sep 10 evening → Sep 11): Coxon worked at Anthropic six weeks, "all of his socials appeared yesterday," and the amplification pattern suggests "a coordinated op through doomer mega donors." Jeffries' thread goes further — anti-AI messaging groups framed as "terror cells," the Zizian murders cited, "Butelarian Jihad" as their rhetoric. Political layer: Trump dismissed the extinction warning in remarks carried via UK mirror (robots-feed pickup), the story's first entry into mainstream politics.
What's actually new: the name, the tenure counter-frame, and the political pickup. The counter-narrative is itself the story: the discourse war has moved from arguments to characterizing the messengers. Treat both sides' framing as framing — "six weeks" is a fact about tenure, not about whether the concern is valid; "terror cells" is not a sourcing standard either.
Why it matters: yesterday's risk-governance watch item ("the resigning researcher being named") resolved — and the resolution is that naming produced a smear contest instead of a substantive debate.
Links: https://x.com/ylecun/status/2098168313008586937 · https://x.com/ylecun/status/2098183593365537092 · MSN mirror (Trump reaction): https://www.msn.com/en-gb/politics/government/donald-trump-dismisses-mass-extinction-from-super-intelligent-killer-ai-robots/ar-AA2c2yDH
Coinbase telemetry: the first agent-trading market share (continuation)
What happened: Arav Srinivas: "Perplexity Computer makes you money" (5:07 PM ET). Underneath it, Coinbase Developer Platform's own breakdown of where agentic trading volume came from last week: Perplexity 26.8%, Claude 19.3%, Grok 15.3%, ChatGPT 1.1%, Claude Code 0.7%, CLI/other 36.8%.
What's actually new: the telemetry. It's single-source (the exchange's own platform data) and "trading volume" is doing unspecified work — but it's the first market-share read of agent economies, and Perplexity's "Computer" agent leading it is a real datapoint for the index-as-infrastructure thread (this is the same Computer MCP server we didn't evaluate yesterday, now with a revenue story).
Link: https://x.com/AravSrinivas/status/2098458484647907468
xAI's Grok Bot week: livestream spectacle + GTM push (watch)
What happened: SpaceXAI will livestream a three-person team building an entire company from scratch with Grok Bot, Sep 15–17 (tetsuo; Musk: "This will be cool") — "from ideation and product development to real engineering work and deployment." Same-day: Grok Bot adds GTM integrations (Salesforce, HubSpot, Gong, Clay) "for sales teams"; 50 free codes circulating ($200 usage or Ultra). Color underneath: Neuralink patient Bradford Smith's detailed account of running Grok Bot as his posting agent — drafts, waits for approval, retries, "judgment stays mine."
Why it matters: the livestream is the demo-reel genre transplanted to company-building — three days of spectacle whose honest output measure is a working product, not watch-hours. The Smith account is the serious artifact in the cluster: a real assistive-agent deployment with explicit human approval gates.
Links: https://x.com/elonmusk/status/2098304166854635528 · https://x.com/elonmusk/status/2098304042896162881 · https://x.com/elonmusk/status/2098283931007107457
OpenClaw v2026.9.4 + the security report (continuation)
What happened: OpenClaw shipped v2026.9.4 — 293 contributors, plugin/skill discovery, chats-to-skills conversion ("You steer"), GPT Image 2.5 on the canvas, more cloud controls (4:33 PM ET). Companion: Vincent Koc's security post — 1,788 security reports filed by researchers, "most junk," 17 critical, security page live.
Why it matters: two ecosystem-health datapoints for the open-harness side of the harness race, the week OpenAI made the harness a managed cloud product. The cadence (2.0 Aug 31 → 9.2 → 9.3 → 9.4) and a published security posture are the independent-harness answer to "why do you need a vendor harness."
Links: https://x.com/openclaw/status/2098449762119020926 · https://x.com/steipete/status/2098206204271173781
Quick hits
- Claude Code: plugin evals ship —
claude plugin evaltests plugins/skills against cases and scores with-vs-without deltas, aimed at "is my skill still working on the new model" (trq212, 5:58 PM ET). Same author's eval-critique: "basically impossible to interpret evals by looking at just pass/fail scores" — the eval-trust thread keeps growing tooling. /diff became a persistent real-time pane (Cherny). https://x.com/trq212/status/2098531560643539440 - LeCun's anti-doomer amplification round (Sep 10 evening, discovered early Sep 11): four RTs — Weiss-Blatt's doom-bible critique, Chaubard's HF/Navier–Stokes correction thread (asserts the HF model was prompted with an ExploitGym objective, not autonomous; the NS run used traces of prior work), a French extinction-literature review ("no plausible scenario without human intervention"), Jeffries' terror-cells thread. The practitioner-side correction of the HF narrative matters most — it's the same "prompted eval" vs "independent volition" question Anthropic's disclosures keep turning up. https://x.com/ylecun/status/2098183593365537092
Dropped
- The Elon RT sludge (robotaxi ride-along, BBC-interview clip, Cybercab efficiency, Solar take, Honduras Starlink plan, Tesla Semi Europe specs, Boring Co tunnel economics, Gulag-Archipelago clap comparison, Austin murder amplification, Farzad's free-code giveaway). Gemini for Windows (Alt+Space assistant — distribution note, no story). DeepMind's "Love, Rendered" documentary (marketing artifact, touching). tldraw arrows color, SF-heat AC joke, "Effective Altruism in the wild."