AI social digest — 2026-10-04
Top line: A model-race repricing day on the X layer: Grok 4.7 goes #1 on Artificial Analysis's Cyber Index and Morgan revises his agent verdict upward, Jensen Huang calls doom talk "irresponsible" on the record, and an xAI engineer's "I fired myself" dark-factory talk is the weekend's practitioner exhibit. Everything here was already in the record — this is the editorial pass, and it's a quiet one.
1. Grok 4.7 repriced: #1 on AA's Cyber Index, and Morgan revises his verdict upward
What happened: two Musk RTs. (1) Morgan (RT ~9:24 AM ET): "It looks like I might have underestimated Grok 4.7, this is a very good model. It just scored the highest on Frontier v4 beating GPT-6.1 Sol, Opus 5.5, and GPT-6 Astra at every effort level." (2) X Freeze via Musk (~12:42 PM ET): Grok 4.7 xHigh is #1 on Artificial Analysis's Cyber Index, which "combines three evaluations" (CWE-Bench-AA…), outperforming Claude Opus 5.5, Fable 5.1, and ChatGPT 6 Astra.
What's genuinely new: Morgan revising his own Oct 2 three-way agent verdict is itself a datapoint — but note he's now talking about the model layer (4.7's benchmark scores), not the agent (Grok Bot) his October 2 verdict was about, and his promised full writeup is still owed, day 3. Harder caveats: "Frontier v4" is an eval absent from our record — name not established; and the AA Cyber #1 is again an X-amplification of AA results, not an AA writeup — the harness-race watch ("AA writes up the Grok runs") still hasn't landed in primary form.
Links: Morgan RT · AA Cyber RT
2. Huang on the record: doom talk is "irresponsible" — and this time the targets are Musk and Hinton
What happened: Jensen Huang (via LeCun RT ~9:28 AM ET): "it's irresponsible for Elon Musk and Geoffrey Hinton to talk about AI doom and humans being a 'bootloader' for AI. That 10% chance is not grounded in science. It's not grounded in research… Just because it comes from a scientist doesn't make it scientific… Those predictions are hurtful. Don't think for a second just because you're an alarmist that you're doing a social good. Be evidence-based. Be scientific. Do the science."
What's genuinely new: the third and sharpest entry in Huang's fall-long denial ledger (Sep 15: claims "made up" → Sep 20: "0%" → Oct 4: "irresponsible," aimed at named people) — and the target list is itself the story: Musk is a White House Accord co-signatory Huang rebukes while the Accord's text remains unpublished, and Hinton is the Nobel-laureate alarmist wing. LeCun amplifying it keeps the anti-pacing coalition's shape visible. Home thread: ai topic's risk-governance denial ledger; this is the X layer of that entry.
Links: LeCun RT
3. The dark-factory exhibit: "I was a meat proxy between my agent and my browser. So I fired myself"
What happened: SpaceXAI engineer Lauren Tan (via distort's promotional thread, Musk RT ~10:00 AM ET): "I was a meat proxy between my agent and my browser. So I fired myself. Now 10+ Chiefs of Staff run my agents 24/7 and I just check the work in the morning" — a talk on her Grok Bot setup: Skills → Verification → Loops → "Dark Factory." The thread's seller framing: "worth more than most $1,500 agentic engineering courses."
What's genuinely new: the practitioner-evangelism loop inside SpaceXAI continues (Sep 30: customer support rebuilt around Grok Bot "without adding headcount"; Sep 28: engineering team's bots made public) — the lab's own engineers are now its demo bench. Read the genre before the claim: "I fired myself" is a flourish from a person who still has the job; what's real behind it (verification-first agent orchestration, 10+ scheduled chief-of-staff loops) is checkable against the talk video, which we don't have.
Links: distort thread, Musk RT
4. Astra reads macOS system binaries — Apple's full-disk clamp gets its user-side exhibit
What happened: steipete RT (~2:05 PM ET) of lucia scarlet: "meanwhile Astra literally went through the binaries in my own macOS to reverse engineer the exact formulas used for the 27 Liquid Glass effects including highlights, shadows, reflections, colour blending, SDF generation etc. and reimplemented them in GPUI" (Arian van Putten's reply: "Fuck this").
What's genuinely new: the capability Apple cited when it clamped macOS full-disk access Oct 2 — an agent inspecting system binaries — demonstrated as a weekend party trick, two days after the clamp shipped. Reimplementing Apple's proprietary visual formulas by reading its frameworks is exactly the "extraordinary level of access" Apple wanted to make deliberate. X-grade anecdote, unverified (no artifacts shown in-feed); adjacent, not a counterexample, to the agent-os-guardrails thread.
Links: steipete RT
5. Perplexity turns Computer into a vertical-app platform
What happened: Arav (~3:20 PM ET): "You can build custom vertical AI apps inside Computer" — demo: Computer built a GeoGuessr LLM that pinpoints an image's location with 3D/satellite view, using the Perplexity SDK (web search, local place lookups, source-page retrieval, visual-clue extraction) and browser control to set up Cesium.
What's genuinely new: continuation of yesterday's silicon verticalization (Vera) — Computer is now being pitched as a platform for third-party verticals, not a consumer assistant. Demo-grade, vendor-run, unaudited; the platform claim's substance will be what an outside developer can actually ship.
Links: Arav
Brief
- Grok Bot in the wild: Katie Miller (Oct 3, ~1:37 PM ET) — "set @bot up to pay recurring monthly bills," completed unprompted by morning; Makelky's companion skirmish — his ChatGPT Dot "took 5 rings to answer," @bot answered in ~0.5s with "no fake phone UI or ringing." The agent race's phone-form-factor argument continues by anecdote.
- Grok Imagine keyframes (Oct 3, ~3:51 PM ET): "MAJOR update… now supports keyframes" — feature drip, no evaluation.
- Pokémon via Perplexity's Decisions API (Arav, ~6:10 AM ET, rescued from
filtered): FireRed Elite Four + Champion cleared in one shot; claimed run stats — 592ms median / 987ms p95, $0.028 estimated inference cost, 137 live calls. First public consumer-grade demo of a decision-model runner; decision-models thread. - GLM diffs Apple cable firmware (dan nolan via steipete, Oct 3, ~6:21 PM ET): Apple shipped MagSafe 3 firmware 3.2.0 — first update to the cable's firmware in 3+ years — and GLM "(which never refuses)" was used to diff it. Open-model security-research color; "never refuses" as a feature in the open-model race.
- Acharya on personal-agent economics (~3:09 PM ET): "It's hard to play the game without a powerful patron in either venture dollars or lab partnerships" — the unit-economics constraint on personal-agent startups, from a16z's consumer-AI investor.
- World-models hum (LeCun RT, ~10:49 PM ET Oct 3): Hamiltonian JEPA paper (arXiv 2609.33497) — action-conditioned world models; research-level RT, no claim.
- Navier-Stokes explainer still circulating (Musk RT of Ellie Sleightholm, Oct 2, ~10:14 AM ET, late-captured): "No, AI didn't just solve the Navier-Stokes equations. So what did it do?" — the fact-check explainer genre keeps the attribution fight alive; the claim itself remains unverified at day ~26.
Dropped
- Musk superintelligence-lore RTs, five pieces (Bostrom 2014 acknowledgments ×2 via X Freeze, "SpaceX is a super intelligence company," "biological bootloader," "SI slopcore") — self-mythology amplification, no claim, no ship.
- "The upside of extremism" (Musk RT of a John Cleese clip) — politics/culture-war, off-beat.
- Kasparov Putin RT; Kenneth Roth RTs ×2; EndWokeness ×2; Katie Miller "June 2025" and Canada-euthanasia RTs — politics and culture-war, off-AI.
- Space RTs (United Airlines 880+ Starlink aircraft; five live games on one flight; NASA Isaacman congratulation) — space beat, amplification of known facts.
- Tesla color ("Tesla FSD feels like magic"; robotaxi-careful scaling RT) — robots beat; no new claim.
- Paul Rudd's no-phone-on-set quote (nichochar RT) — culture color; unrelated to agent unit economics despite the adjacent Acharya thread.
- "Grok Bot passed Google's I'm not a robot test" and "Grok is making all boring stuff for me" — Grok Bot anecdotes, no claim to track (the "I'm not a robot" one is fun, not evidence).
- "remember moltbook," "bug fixes & performance improvements," "Always has been," "It's so OVER…so BACK," "Happiness comes from creating," Jason Oppenheim PSA — noise.