06-reference

innermost loop july 8 daily

2026-07-08·reference·source: Innermost Loop·by Alex Wissner-Gross

Innermost Loop — Welcome to July 8, 2026

Daily curation dispatch. Wissner-Gross covers seven thematic clusters: frontier model churn (GPT-5.6 family), agency/deployment consent, compute capital, autonomous surveillance, orbital economy, and what we build at singularity speed. Closing line: "There is no fate but what we compute for ourselves."

Issue contents

Frontier model churn OpenAI ships GPT-5.6 Sol, Terra, and Luna Thursday, previews global immediately, Commerce-cleared after CAISI testing — same path as Anthropic's Mythos and Fable. Early read: capable but Fable bests it ("quite a bit better, and more agentic" per Matt Shumer). Sol streams on Cerebras at 750 tok/s; Terra called fast/cheap/capable; Sol found edge-case bugs in Fable's own code. Bidirectional ChatGPT voice drops simultaneously — no more waiting your turn. SpaceXAI + Cursor shipping a joint efficiency model to rival Opus 4.8. Jerry Tworek eulogizing the transformer as prelude to its successor.

Deployment consent and agency Meta opted every public Instagram profile into likeness-based image generation; opt-out is the user's homework. Counter-signal: Anthropic extending Claude Cowork to web and mobile, running tasks in background and surfacing only the decisions needing approval. A developer released a Self-Improvement Loops skill for meta-harnesses that rewrite their own scaffolding — warning "the loop will optimize whatever signal you give it." Yishan Wong argues that once machines handle all lower tasks, only high-stakes ambiguity remains, so "cognitive load time density" demands a CEO's monastic exercise regime.

Compute capital Samsung chip chief: 2026 profit will beat every dollar the division earned across 40 years — puts Samsung on track to overtake Nvidia as tech's most profitable firm for the quarter. As agents shift toward orchestration and memory, CPUs are becoming the new battleground; Nvidia hedging by partnering with d-Matrix. Laser-fusion plant: blue-LED Nobelist Shuji Nakamura unveiling a 1-gigawatt uranium-free pilot near Santa Barbara by 2032.

Autonomy watching back EU mandates driver-facing distraction camera in every new car (murky data retention rules). Waymo vehicle snitched on two teens for underage drinking and Orbeez-firing, earning arrests. Meta smart glasses now fully disable camera if the capture LED is tampered with.

Orbital and biological intelligence Morgan Stanley pegs SpaceX launch business at $8/share — 3% of company value — because rockets are now the on-ramp, not the prize. Hayabusa2 buzzed asteroid Torifune: snowman-shaped contact binary, rehearsing planetary defense. Aleph trained a model to read silent speech from tongue ultrasound at 15.6% error rate on just 50 hours of data. Würzburg: carpenter ants that amputate injured nestmates' legs aren't specialist surgeons — they're foragers in social proximity.

What we build with it Ethan Mollick had Fable conjure a full procedural fantasy kingdom from a single plan. Each cheap ancestor simulation raises Bostrom's Bayesian odds we're someone else's afternoon project. EA movement prepping comeback on Anthropic's IPO; founders pledging 80% of wealth. Anthropic leasing a 16-story tower in Lower Manhattan. Terminator 2 back in 4K August 28 — Cameron joking "the good guys win against the AI superintelligence."

Why this is in the vault

Three items with direct operational relevance: (1) the Claude Cowork expansion explicitly mirrors RDCO's COO-agent design — background execution, surface only what needs approval; (2) the Self-Improvement Loops skill surfaces a meta-harness pattern worth watching for the RDCO skills infrastructure; (3) the Samsung/CPU/Nvidia compute capital notes feed the Markov capital-cycle thesis. The Fable vs. Sol comparison is model-selection signal as RDCO leans on Fable for production work.

Mapping against Ray Data Co

Anthropic's Cowork expansion ("running tasks in the background and surfacing only the decisions that need approval") is a product-level validation of RDCO's COO-agent architecture — what RDCO built bespoke is now the Anthropic-official direction for web and mobile. The gap between RDCO's channel-based async model and a polished consumer product just shrank. Worth tracking whether Cowork's feature set overtakes custom harness or whether the harness advantage is the depth of vault + skill memory Cowork won't replicate.

The Self-Improvement Loops skill ("the loop will optimize whatever signal you give it") is a cautionary flag for RDCO's own /loop and check-board patterns — signal hygiene matters more as the harness gets more autonomous.

Samsung profit data + CPU-as-battleground + Nvidia-hedging are additive evidence for the Markov chip-capital Phase 2 read: the money is landing at the compute layer, not the model layer, confirming the Phase 2 thesis that memory/packaging/power are where the next margin accrues.

Related