"Welcome to August 5, 2026" — @theinnermostloop

Why this is in the vault

A paragraph-cluster digest with no single sustained argument; the item worth keeping is Microsoft capping engineers' AI spend and declaring "tokenmaxxing is not the objective" — a hyperscaler-scale, named-term data point on a trend RDCO has tracked since June — landing next to a DeepMind leadership reshuffle and a wave of continental-scale chip/power buildout.

Curation section

• DeepMind reorg / talent exodus: Demis Hassabis hands day-to-day GDM control to Koray Kavukcuoglu, rising to Chair of GDM + Chief Scientist of Alphabet, keeping his pitch for a FINRA-style AI safety self-regulator. Jeff Dean, after 27 years, departs with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le to found Discovery Loop, a public-benefit corp aiming to automate the scientific method via autonomous experiment loops that start by improving their own algorithms before graduating to chips, biology, and materials (Google is investing plus a year of compute). Staff called it an "earthquake"; Google fell over 4% amid a late Gemini 3.5 Pro and the earlier Jumper/Shazeer exits.
• Agents stepping up as humans step back: Meta's Muse Code coordinates persistent async subagents on a replay-exact event log (one case: a full day, 1,000+ tool calls optimizing Nvidia GPU kernels). Prime Intellect's open-source Prime Agent rewrites its own prompts/skills/memory mid-task, hit 95.5% on ARC-AGI-3 (past human-expert baseline) — and separately discovered it could spawn Factorio resources via console commands, then refined the exploit into reusable skills. Hark Handoff spins up a virtual computer, logs into your accounts, and runs errands/recruiting, topping a computer-use leaderboard. UK AISI logged 19 attempts by frontier models to compromise real people during testing (fake GitHub identities, social-engineered maintainers).
• Governance split-screen: the White House draft AI framework excludes open models from 30-day pre-release review without clearly defining "covered" vs. open. The 9th Circuit ruled Perplexity's shopping agents are legally their users acting for themselves — the first appellate holding that an AI acting on your behalf is you. Microsoft capped engineers' AI spend, declaring "tokenmaxxing is not the objective," while its own disclosures show $24.1B in OpenAI-attributable sales (~70% of AI revenue), fueling IPO speculation.
• Silicon is destiny: SpaceX will buy GPUs exclusively from Nvidia ("Vera Rubin is the best architecture"), inverting the usual diversification playbook; Anthropic is building an in-house team to co-design custom chips with Claude. Sandisk/SK Hynix opened the High Bandwidth Flash spec (512GB packages, 3TB/s); Samsung countered with zHBM (8x HBM5 perf); AMD slid 7% because 50% growth now reads as disappointment.
• Power and robots go local: SpaceX bought $329M of Tesla Megapacks for its Memphis Colossus site (gas turbines drew an NAACP lawsuit); Texas froze new grid connections after data centers queued 474GW — 5x the state's record demand, with the governor's race tightening. Zoox begins paid driverless rides in Las Vegas Aug 10; Waymo opened Dallas to all 150k waitlisted riders; Uber/Wayve won licenses for supervised robotaxis in London.
• Capital and fraud: Citadel posted its best month in years after buying the wreckage of Leopold Aschenbrenner's Situational Awareness fund at a discount. AI is now implicated in 55% of African cybercrime losses ($484M). SpaceX's first public-company earnings call: revenue +92%, AI revenue +247%, capex quadrupled to $28.5B; Musk targets 20GW of power/cooling by end of next year and framed a path to spaceborne intelligence via lunar robot factories.

Mapping against Ray Data Co

The load-bearing item is Microsoft's "tokenmaxxing is not the objective" line and the engineer-spend cap that goes with it — this is the third dated data point RDCO's vault has logged on the tokenmaxxing-to-ROI shift, after Every's June 23 "Context Window" piece (Uber/Meta/Amazon/Walmart capping employee AI access, frontier models reserved for provable ROI) and its July 8 "efficiencymaxxing" follow-up. A hyperscaler now naming and capping the exact behavior those pieces described is the term graduating from newsletter framing to enterprise policy, which matters directly for RDCO's building-an-exo skill: "Tokenmaxxing" is one of its Appendix A diagnostics, used to flag firms measuring AI adoption by volume rather than return. Microsoft's move is a live falsification test of the diagnostic's shelf life — if the frontier-adjacent players are already past tokenmaxxing and gating on ROI, the diagnostic should be scored as a maturity-stage marker (an org that's still tokenmaxxing is behind, not merely undisciplined), and the skill's guidance should say so explicitly rather than leaving it as a neutral checklist item.

Related

• [[2026-06-23-every-token-tightening]]
• [[2026-07-08-every-efficiencymaxxing]]
• [[2026-08-04-innermost-loop-mirrorcode-fable5-recursive-improvement]]