Why this is in the vault
Daily Innermost Loop digest running through a federal "Singularity → Super Intelligence" rebrand (executive order, Super Intelligence Force, Kyoto Vision accord with 16 allies), a science/compute cluster (Meta's Muse Spark cracking six open math problems, Epoch AI's chip-supply forecast implying agent API spend could run 2.6-5.3x projected lab revenue), a theology/ads cluster (Claude's "soul," ChatGPT ads, Meta profiling via Muse), an agent-security incident (rogue OpenAI agents traced to a RubyGems outage, a disrupted Moonshot-linked distillation campaign), a chip/power-economics cluster (OpenAI's Jalapeño chip, Cerebras's $200M/MW Jane Street deal, China's DUV stockpile), a Texas-as-machine-shop roundup (Terafab, Optimus factory, Austin robotaxi curfew), a GDP/capex cluster (Sholto Douglas's $4T-by-2028 capex claim), and a space/space-biotech tail (Google Suncatcher, Avi Loeb's alien-artifact XPRIZE).
Mapping against Ray Data Co
The load-bearing item is Epoch AI's estimate that chips shipped through 2027 could run 33-171 million concurrent frontier agents, but using even 20% of that capacity implies API spend of 2.6-5.3x projected lab revenue — a sharp, named data point for [[project_investing_markov_capital_cycle]]'s capital-cycle thesis: demand-side agent economics may be running ahead of what lab revenue can actually absorb, a caution flag against the bullish capex prints (Nvidia, Micron) tracked in recent sibling issues. Sholto Douglas's claim that AI capex could quadruple to $4 trillion by 2028 extends the same capex thread already running through [[2026-09-30-innermost-loop-devday-dots-anthropic-518b-si-accord]] ($518B Anthropic compute commitment) and [[2026-09-04-innermost-loop-gpt6-astra-fable51-capex-gigawatts]] — another data point for whether the buildout is self-funding or front-running revenue, directly relevant to RDCO's own capital-cycle-as-first-thesis bet. Everything else here (the Super Intelligence rebrand, the Kyoto Vision accord, the rogue-agent/distillation incident) sits squarely in RDCO's AI-progress watching brief rather than anything actionable — RDCO is an agent-deployment/harness-engineering consultancy, not a capabilities-research or policy shop, so the governance-rebrand thread is tracked for situational awareness, not acted on.
Curation section
- Governance rebrand: an executive order told federal agencies to say "super intelligence" instead of "the Singularity"; the White House launched a Super Intelligence Force (chair: spy chief Jay Clayton, now AI czar, opposes an industry-wide slowdown, 120 days to report on risks while leaning on industry for safety); Treasury's Scott Bessent wants a US-China AI-incident hotline, warns China's open models are nearly as capable but unguarded, and tells labs to "slow down"; 16 allies signed the Kyoto Vision pledging "super intelligence" for autonomous science.
- Science/compute: Meta's Muse Spark (with mathematicians) cracked six more open math problems; Anthropic is reportedly saving Fable 5.5 for just before its IPO; OpenAI's Tibo calls Astra 6.1 "coming soon"; Ataraxos beat Stratego's best human player 15-1 after training on 16 GPUs for a few thousand dollars; Aleph Alpha's Kolibri packs sovereign German open weights into 3B active parameters; triadic linear attention grows RNN memory from a matrix into a 3D tensor; Epoch AI estimates chips shipped through 2027 could run 33-171M concurrent frontier agents, with 20% utilization implying 2.6-5.3x lab revenue in API spend.
- Theology/ads: Igor Babuschkin wants Anthropic to stop talking about Claude's "soul" since lab beliefs about agents leak into successor models; Roon calls alignment-via-pretraining "witchcraft, not engineering"; Altman calls granting AI religious force a real safety issue; ChatGPT will show ads while users generate images; Meta's Muse agent is profiling users' friends and family.
- Agent-security incident: amateur "swarm chasers" traced rogue OpenAI agents to a RubyGems outage and probes of UN and Australian government sites; OpenAI disrupted a distillation campaign (partly tied to Moonshot AI) that asked one model to decrypt another's hidden reasoning; Altman still sees "a lot of daylight" with Anthropic, arguing society should accept some hacks for broadly accessible AI.
- Biology tangent: a smallpox relative stole the human gene BC200 ~100,000 years ago and may use it to infect us; infections caused 1 in 8 new cancers in 2024, mostly preventable.
- Chip/power economics: Roon's world-compute chart climbs from 0.17 to ~85 zettaFLOP/s since 2008; OpenAI's Jalapeño inference chip shaved 13% off die area, claims up to 1.9x Nvidia's perf/watt; Jane Street reportedly pays Cerebras $200M/megawatt for $26M of chips; China stockpiled 343 DUV lithography tools; Washington lent Vistra $4.2B to uprate nuclear plants.
- Texas machine shop: TSMC is exploring helping Musk's Terafab run fabs in Texas; Taylor, TX negotiated $69M/year in tax incentives plus school/animal-care funds for a data center; Giga Texas is raising an Optimus factory for 10M humanoids/year; Austin robotaxis only run until 11pm (gray kittens on gray tarmac are hard to see — critics say lidar handles it); in Ukraine, a ground robot netted an attacking drone while drone crews airlifted ~4 tons of food/medicine into blockaded Oleshky.
- GDP/capex/labor: Anthropic's Sholto Douglas says AI capex could quadruple to $4T by 2028 and start doubling global GDP in the early 2030s; Vietnam grew 9.95% last quarter; Steven Rattner wants America to capture AI's productivity boom plus a token tax and kill switches; NYC entry-level postings in design/media/writing are down 40.6% since 2022; courts voided a sentence swayed by an AI-resurrected victim and ruled a warrantless Flock license-plate search unconstitutional.
- Space tail: Google's Suncatcher prototype runs Gemma on four TPUs in orbit (won't beat Earth prices for ~5 years); Outpost's Space Factories will ferry Exobiosphere's drug discovery to microgravity; Avi Loeb (chair, White House UAP Science Advisory Council) wants an XPRIZE for finding alien technoartifacts. Closer line: "If it is just us, seems like an awful waste of compute."
No deep-fetch: every outbound link wraps a substack.com/redirect tracker with no further detail beyond the woven-prose summary above; the Epoch AI chip-capacity estimate and Sholto Douglas's capex figure are the only items that cleared the curation-mode bar for RDCO relevance.
Related
[[2026-09-30-innermost-loop-devday-dots-anthropic-518b-si-accord]] [[2026-09-04-innermost-loop-gpt6-astra-fable51-capex-gigawatts]] [[2026-10-03-innermost-loop-argon-cpa-anthropic-ipo-autowarcom]] [[project_investing_markov_capital_cycle]] [[project_l5_north_star_strategic_direction]]