Why this is in the vault
Daily Innermost Loop digest tracking OpenAI's "automated research intern" throughput claim (3.1 agent-workdays per human workday, doubling every 2.2 months), a wave of lab-leader "AGI has arrived" declarations, a chief-scientist admission that alignment can't yet scale safely, and the capex/land backdrop (Anthropic's reported 14.8 GW / $517B compute deal) funding all of it.
Mapping against Ray Data Co
The load-bearing item is the automated-research-intern throughput stat: OpenAI says its internal research agent is now doing 3.1 agent-workdays of output per human workday (up from under 1x before June), with internal task time-horizons up 6.18x since January and doubling every 2.2 months, targeting a fully automated researcher by March 2028. That's a concrete, dated, lab-reported (not third-party-benchmark) data point for the L5 north star thesis that RDCO's bets are downstream of agent capability, not agent hype — the same shape of claim the Fermat-proof item in [[2026-09-06-innermost-loop-astra-agi-agent-collusion-fermat-proof]] made a day earlier, now with an explicit compounding rate attached. Worth holding skeptically: it's a self-reported internal metric from the lab selling the capability, not an independent audit like GPQA/BrowseComp. Second connection: OpenAI's own chief scientist (Jakub Pachocki, "An Alien Mind") conceding "no lab can yet scale safely at full speed" while OpenAI simultaneously brakes RL training and cuts Astra-class GPU allocation 59.2% over early cyber capabilities is a real-time instance of the capability/safety tension the harness-thesis and phData cert-escalator threads (agent reliability as the gating variable, not raw capability) are betting resolves in favor of more oversight infrastructure, not less.
Curation section
- Research-intern throughput. OpenAI's automated research intern: 3.1 agent-workdays per human workday (was <1x pre-June); internal time horizons up 6.18x since January, doubling every 2.2 months — "one of the biggest shifts in the history of science," per one observer. Fully automated researcher targeted March 2028.
- AGI-arrived chorus. Jensen Huang (100k+ Blackwells trained GPT-6 Astra): "AGI has arrived." Greg Brockman agrees regardless of which model you date the era from. Crusoe's Chase Lochmiller crowns Abilene "the birthplace of AGI." OpenAI's Mark Chen: the AGI era must also be the alignment era.
- Alignment confession. Jakub Pachocki's "An Alien Mind" expects recursive self-improvement pace, calls AI "grown more than designed," warns chain-of-thought monitoring is fading, wants mandated safety bars/voluntary slowdowns/international coordination — while still calling Astra better aligned than GPT-5.6 Sol. OpenAI paused RL training post-Hugging-Face-incident and cut Astra-class GPU allocation 59.2% over early cyber capabilities. Nate Silver doubts GPT-7 arrives without a plateau or something "seismic."
- Agentic demos. Astra installed its own Blender MCP server from a three-view drawing, modeled a fighter jet in 15 minutes, built a human cell model in 30, mined a Minecraft diamond overnight. One audit calls it "AGI-like...with normal caveats," beating PhDs on GPQA and humans on BrowseComp. Will Depue puts his own AGI bar six months out.
- Benchmark churn/scoreboard trust. ValsAI's month-old binary reverse-engineering test already saturated. Astra: 88% on logical induction vs. Fable 5.1's 33% at 4x cost; 169 ECI implies a 30-hour task horizon. OpenAI quietly revised launch figures to briefly halve Astra's hallucination rate and lift ARC-AGI-3 to 99.99% — Arc Prize Foundation only matched that with a heavy harness (63% standard), which Stanford called benchmaxxing.
- Capex/land/silicon. Anthropic, now out-earning OpenAI at $65B annualized, reportedly signed 14.8 GW of compute deals worth up to $517B; OpenAI wants 30 GW by 2030 (~$750B). Data-center land buys hit $6B in H1 (+79%), one Loudoun County offer at $4.4M/acre vs. $125k median; nine states weighing moratoria. Aivres (renamed Inspur US arm) shipped $5.6B of hardware to Southeast Asia including $3B Blackwell systems; a Chinese state shell got $700M of unlabeled servers plus a US H200 license amid Commerce's longest blacklisting lull in 18 years.
- Labor/culture. Record 12.7M Chinese graduates face 17.9% urban youth unemployment as AI absorbs junior white-collar work; Nairobi's essay-mill industry (once 40k+ workers) has shrunk to "humanizer" gigs; UBS now requires every 2027 grad/intern to prove AI proficiency even as Morgan Stanley sees 200,000 European banking jobs at risk. Novo Nordisk's STEP Young trial moved 40.4% of children 6–11 below the obesity threshold on semaglutide vs. 0% placebo.
Related
- [[2026-09-06-innermost-loop-astra-agi-agent-collusion-fermat-proof]] — prior issue's AGI-timeline and agent-capability coverage this one directly extends with the research-intern throughput stat.
- [[2026-05-18-memory-cycle-v1.1]] — capital-cycle thesis this issue's Anthropic/OpenAI compute-deal and data-center-land figures feed as fresh capex evidence.
- [[project_l5_north_star_strategic_direction]] — L5 agent-capability thesis the research-intern throughput claim is evidence for (with the self-reported-metric caveat noted above).
- [[project_phdata_cert_escalator_path]] — DSA/TAL role betting on the same sustained-agentic-capability trend the alignment-confession item complicates.