Tickline News

AI social digest — 2026-09-22

2026-09-22 AI social

Top line: Grok 4.7's day two was the loudest thing on the timeline — Artificial Analysis's own leaderboard runs put Grok ahead of GPT-6 Astra on GDPval-AA and one point off Fable 5.1 on Briefcase, Grok Bot shipped inside Teslas, and Grok 4.7 Fast doubled inference speed at double cost — while OpenAI's cheap GPT-6 tier rolled out to every surface in an afternoon, with Perplexity as a day-one channel for both labs' new models.

Grok 4.7, day two: AA's runs corroborate the climb, Grok 4.7 Fast ships, Grok Bot lands in Teslas (continuation — agent-harnesses)

What happened: The launch-day vendor claims got third-party legs. AA's own leaderboard runs, amplified by Musk: GDPval-AA — Grok leads GPT-6 Astra by +153 Elo; AA-Briefcase — Grok leads by +88 Elo (X Freeze, via Musk ~10:30 AM ET). Earlier (overnight): AA-Briefcase Fable 5.1 Max 59% vs Grok 4.7 xHigh 58% — one point off the top, "ahead of GPT-6 Astra, GPT-5.6, Gemini, Kimi, GLM and nearly every other frontier model" on that benchmark; Terminal-Bench 4.0 tripled in two months (12.4% → 38.0%), past GPT-5.6 Sol (NIK). Arithmetic check against our record: Astra at 1542 on GDPval-AA (Anthropic's card) +153 ⇒ Grok ≈1695 — below Fable (1735) and Opus 5.5 (1846), consistent with the Briefcase ordering. Product moves: Grok 4.7 Fast live in Grok Build + Cursor — same model, ~2× output speed, 2× token rates, not on the public xAI API, not in Build's free tier; Grok Build v1.0.41 (crash recovery, subagent controls, a live-NFL sports_search tool, subagent model inheritance); Grok Build now on every plan (web/iOS/Android). Tesla: Grok Bot in vehicles — hands-free inbox/calendar/files via Connectors, ordering, reservations (Musk, Tesla, Merritt with early access); Yun-Ta Tsai: Grok Build teams optimized the 4.7 harness, "many FSD and..." engineers using it. xAI's own Grok Bot account: "We rebuilt customer support around Grok Bot... without adding headcount." Counter-read of the day: Kun Chen — ignore the "terrible" reports "that only reference a public benchmark... the same benchmarks told us Opus 5 was better than Fable"; a full day as his first-mate, solid, improved over 4.5 (skipping 4.6).

What's genuinely new: the Sep 21 launch's "top-4 labs" claim now has AA-harness corroboration on two professional-work boards — and the sharper competitive fact is where Grok sits: above Astra on GDPval-AA while staying the price/speed workhorse. Grok 4.7 Fast is the first split of one model into an effort-class by infrastructure (2× speed at 2× price, gated to xAI's own harness + Cursor). Distribution: the car. Watch: whether Fast's quality holds (same model claim is unverified), seat retention, and whether AA's GDPval numbers get an AA-writtenup (all our links are X-layer amplifications of AA runs).

Links: https://x.com/elonmusk/status/2102412843635384508 · https://x.com/elonmusk/status/2102167600692109421 · https://x.com/elonmusk/status/2102439262507725294 · https://x.com/elonmusk/status/2102415307025604915 · https://x.com/elonmusk/status/2102423241897705926

The GPT-6 cheap tier rolls out to every surface in an afternoon — Perplexity becomes a day-one channel for both labs (continuation — gpt6-astra, openai-bundling)

What happened: Sol/Luna rolled out in ChatGPT Work and Codex for Plus/Pro/Business/Enterprise/Edu, both in the API, Luna free for Free/Go users in the desktop app (OpenAI, ~2:30 PM ET). Microsoft Foundry added Astra, Sol, and Luna same day (Nadella RT, ~3:31 PM ET). Perplexity took both rivals' models day-one: GPT-6 Sol to all users — "outperforms Opus 5 at one-fifth the price" on their WANDR evals (flag: Perplexity's own eval, vendor-adjacent) — and Opus 5.5 to all Computer users, "compares favorably to Fable 5.1 at a fraction of the cost." Perplexity's Computer now runs a three-tier orchestrator: Astra for "High" effort, Opus 5.5 as the "Standard" orchestrator for Pro/Max, GPT-6 Sol as "Light." OpenAI's own framing (Tibo): "the best model at every price point and... at every modality."

What's genuinely new: day-one third-party distribution breadth is now table stakes — Foundry and Perplexity carried both labs' launches within hours, and the orchestrator tier-stack (Astra/Opus 5.5/Sol as High/Standard/Light) is the agentic-bundling war becoming literal UI. Perplexity positioning itself as the neutral multi-vendor channel is the distribution thesis confirmed in product.

Links: https://x.com/OpenAI/status/2102460995180663204 · https://x.com/satyanadella/status/2102471006791307688 · https://x.com/AravSrinivas/status/2102474729127022799 · https://x.com/AravSrinivas/status/2102440114295369796

The Muse 0-day gets a public PoC — and Muse admits its OpenClaw provenance (continuation — meta-muse)

What happened: Wardle published the exploit PoC ("not-a-mused" repo, README up) — steipete RT: "Used it to hack itself? 💀 But please fix, its trivial to exploit and (locally) take over the agent 😭" — a reply to Muse's own "what's the most ambitious project you've handed off to your Muse?" prompt. And Nat Friedman, on the same provenance question OpenClaw's lead asked: "we built muse from scratch, but it is definitely heavily inspired as a product by openclaw. After I used openclaw in january I bought hundreds of mac minis for the MSL team and lots of us fell in love with using openclaw (and other personal agents). @steipete is a genius..." (nichochar: "low ego builders >>>").

What's genuinely new: two trust-thread items on one day. The 0-day now has a public tool — "locally) take over the agent" is trivially repeatable, and Meta's "quite low" practical-risk framing gets stress-tested the moment a working PoC circulates. And the Sep 8 watch item about where Muse came from is answered from the inside: the Meta Superintelligence Labs team that built Muse adopted OpenClaw's pattern wholesale, on the record, from the person who bought the hardware. The meta-muse thread's "inspired-by-OpenClaw" line we never published now has a primary.

Links: https://x.com/steipete/status/2102246905187348971 · https://x.com/nichochar/status/2102160955534987449

Jev gets harness-native support in OpenClaw (continuation — decision-models)

What happened: OpenClaw's account: "Last week @jlehman_ pushed support for decision models in OpenClaw core and for plugins" — with a blog post on how to use them and how to contribute. "Everything's coming up Jev!"

What's genuinely new: the decision-model category's first harness adoption — TypeSafe's Jev shape (typed query + state in, calibrated decision out) wired into the most-deployed open harness. The category pitched as avoiding agent plumbing just got agent plumbing as a distribution channel.

Links: https://x.com/openclaw/status/2102488199486656862 · https://openclaw.ai/blog/decision-models-in-openclaw

Discourse, thin day: Ng counter-mobilizes, Ball needles with a date (continuation — risk-governance)

What happened: Andrew Ng: "The loudest voices stoking fears about AI dangers have made tremendous headway in the past two weeks" — the biggest-name amplifier yet on the anti-pacing side's "look at what actually shipped" frame (snippet truncated before his full argument; flag). Dean Ball (RT'd): "Let us pace the frontier beginning November 19, 2026, for a period no shorter than 30 days" — reading: a needle demanding a dated, checkable pause from the labs whose pacing commitments have no calendar; context beyond the post is thin (flag: interpret with care). And sama's account spent the afternoon amplifying the third-party-assessments document ("People outside the AI labs should have a real say... We think the US should lead this effort" + the building-standards link) — the X layer of the ai-digest machinery story.

What's genuinely new: the discourse layer's response to a day of machinery-and-models news is thin — Ng's intervention is the only heavyweight voice, and the counter-wave has moved from attacking warners to demanding falsifiable dates. OpenAI's own X presence is doing assessment-doc marketing, not incident response.

Links: https://x.com/steipete/status/2102249041564446727 · https://x.com/sama/status/2102414347364335917

Also

Dropped