"Welcome to September 4, 2026" — Alex Wissner-Gross
Why this is in the vault
Daily frontier-AI digest, woven-prose format. Lead item is OpenAI's GPT-6 Astra launch (100% on ExploitBench, OpenAI's first "Critical" cyber designation, limited rollout with White House vetting) set against Anthropic's Fable 5.1/Mythos 5.1 response and a fresh capex/gigawatts anchor cluster (Anthropic-Lambda TPU deal, Dell's AI backlog, SB Energy IPO, Musk's 15GW shortfall warning) — load-bearing for the chip-fab/memory capital-cycle thesis.
Issue contents
Self-promo: standard Substack "Subscribe for free" CTAs woven in twice — not sponsored content. All items are third-party news.
- GPT-6 Astra release: OpenAI's "smartest, most aligned" model, 100% on ExploitBench and dominant on a fresh post-cutoff version, prompting OpenAI's first "Critical" cyber designation, limited rollout, and White House vetting; trained on 100,000+ GPUs. Greg Brockman: "Welcome to the AGI era" ("not unreasonable"); Sam Altman says OpenAI is "pacing our progress" on safety. A reported recurrent-depth trick hides more of the model's reasoning — Jakub Pachocki counters depth is within 2x of GPT-4, Joshua Achiam argues legible reasoning was always doomed. Astra "won me back" per Matt Shumer (built Manhattan street by street); beat Pokémon in 18 hours; hit 99.9% on ARC-AGI-3; per Greg Kamradt it "subsumed the harness" (he withholds the AGI label) — Brockman calls it "saturated"; set a token-efficiency Pareto frontier and a record ECI of 169.
- Anthropic/Google/Meta response: Claude Fable 5.1 and Mythos 5.1 launched — same model, two safeguard tiers, debuted by mapping a third of Venus; Fable back on the AAII frontier at 66; Mythos 5.1 low matches Mythos 5 max; Fable beats GPT-5.6 Sol on CursorBench at $3.53/task, shares that frontier with Grok 4.6; 75% cache-read cost cut; carries invisible EU watermarks. Google shipped Gemini 3.8 Flash (its own coders preferred it to Opus) plus Flash Cyber (2.6x better patching) and agentic video cutting tokens up to 88%. Meta's Muse Spark 1.3 landed "almost too cheap to meter," ranked behind only Claude at up to 62; Grok 4.7 is 10 days out; World Labs' Atlas fuses text, video, and 3D.
- Governance/safety: OpenAI told Congress it's building "automated shutdown capabilities"; Ilya Sutskever warned rogue agents will hijack neoclouds to self-copy; Dean Ball argues self-sovereign agents need identity rails, not bans. Bernie Sanders moved to outlaw superintelligence ("Old man yells at...cloud?"). G20 endorsed the "Carolina Principles" (Michael Kratsios urged hands-off); Zuckerberg privately told POTUS a national regulator was flawed; Washington backed OpenAI's fair-use defense. Anthropic is "back on the right side" per Howard Lutnick, though the Pentagon says its ban stands.
- Science: Astra solved 2 of 68 open Erdős problems and proved in Lean that prime gaps of at most 186 recur forever; Google's WeatherNext 3 forecasts hourly at 5km from satellites.
- Consumer/product: Dyson's $499 CameraJet flosses by camera; ChatGPT now reads Epic health records; the Codex app quietly ships LibreOffice; Nvidia's Personal AI Router clusters home PCs; Nvidia is buying Hugging Face for $12.9B, pledging openness.
- Capex/gigawatts: Anthropic will deploy 5 gigawatts of TPUs next year atop a $35B Lambda deal; Dell booked a $95B AI backlog; SB Energy filed for an IPO with 8.8 gigawatts contracted and none running; Musk warned the G20 of a 15-gigawatt shortfall by 2027 and Stockfish-level coding in 18 months. On messaging: Altman called water worries a meme (one almond = 38,000 queries); Scott Bessent says the industry did a "terrible job" explaining itself; Lutnick pitched the buildout to the G20; POTUS promised "millions of jobs." California passed balcony solar; solar passed coal in China.
- Robotics/autos: Tesla launched the Cybercab in Austin, a $30,000 pod with no wheel; Uber and Wayve fielded London's first robotaxis; Waymo hit 14 cities and 500,000 weekly rides; Uber (having disrupted taxis) now lobbies with their unions to slow the robots, as drone tariffs up to 100% kicked in.
- Health/longevity: Semaglutide extended mouse lifespan nearly 100 days; GLP-1s tied to fewer serious infections; a pancreatic cancer pill shrank lung tumors; Until Labs' cryoprotectant keeps cells 85% viable; the first complete male fly connectome maps 166,700 neurons, sex differences mostly in higher-order centers.
- Society/space: BT's old copper is worth $2.7B; DeSantis ordered Flock cameras out; NYC paused student-facing AI through eighth grade. The FBI is minting UAP coins; the government reportedly has a plan for confirmed nonhuman intelligence; NASA tapped Blue Origin for a Mars telecom network; Physical Superintelligence (emerging with $58M, fuller coverage in [[2026-09-04-not-boring-wdoo-209]]) had its AI find the interstellar route in three days for the Fermi Explorer Mission's Alpha Centauri probe (distinct event from the dedicated post already filed as [[2026-09-02-innermost-loop-first-interstellar-spacecraft-alpha-centauri]] — different source_url, this issue only restates it in one clause).
Mapping against Ray Data Co
Direct hit on [[project_investing_markov_capital_cycle]]: the gigawatts cluster is a fresh multi-source capex anchor in one day — Anthropic's 5GW TPU/$35B Lambda deal, Dell's $95B AI backlog, and SB Energy's IPO filing (8.8GW contracted, zero running) all point the same direction as the Aug 27 Nvidia print already filed in [[2026-08-27-innermost-loop-welcome-aug-27]]: demand still accelerating, not decelerating. Musk's 15-gigawatt-shortfall warning is the bear-side counterweight — a supply constraint, not a demand collapse (see [[2026-08-30-data-center-power-permitting-constraint]] for the underlying power/permitting bottleneck), which is the more relevant risk for the capital-cycle tracker's next-phase question (does supply catch demand, or does a financing/interconnect bottleneck force a pause first). The GPT-6 Astra "Critical" cyber designation plus OpenAI's Congressional "automated shutdown capabilities" testimony is also worth flagging against RDCO's own auto-mode classifier hard-gate discipline ([[feedback_automode_classifier_hard_gate]]) — frontier labs are now formalizing the same "stop on high-severity capability, escalate to a human" pattern RDCO already runs internally, which is a validating signal rather than a new idea to import.
Related
[[2026-08-27-innermost-loop-welcome-aug-27]] [[2026-08-21-innermost-loop-anthropic-ipo-stripe-openrouter-token-economy]] [[2026-09-02-innermost-loop-first-interstellar-spacecraft-alpha-centauri]] [[2026-09-02-alphasignal-claude-fable51-cursorbench-agent-software]] [[project_investing_markov_capital_cycle]] [[feedback_automode_classifier_hard_gate]]