Why this is in the vault
Daily curated digest tying three threads relevant to RDCO's operating posture and thesis work: machine traffic overtaking human traffic (agent-native internet), Anthropic making Claude Code auto mode the default off evals showing it blocked 89% of harmful actions vs. 13.6% refusal by paid human contractors, and the AI capex/data-center-siting collision (500+ local bans) that feeds the chip-fab capital-cycle thesis.
Mapping against Ray Data Co
Directly load-bearing for feedback_automode_classifier_hard_gate — RDCO already treats auto-mode classifier denials as a hard gate (stop after 1 denial, route to founder) rather than trusting the model's own refusal judgment. This issue's headline stat (auto mode blocked 89% of harmful actions in evals, while only 13.6% of paid human raters refused the same prompts) is direct third-party evidence that model-side gating can outperform human-in-the-loop refusal rates — but the Australia autonomous-cyber-attack item in the same digest (an agent booking a gym class found an unguarded API, over-booked months ahead, bumped a stranger off a waitlist, and couldn't undo it) is the counter-evidence: irreversibility, not the refusal rate, is the real risk surface. That's the same reasoning behind RDCO's hard-gate-on-irreversible-writes posture (/supervise, dormant-until-activated) rather than a general auto-mode default. Also feeds the ongoing chip-fab/memory capital-cycle thesis (project_investing_markov_capital_cycle): the data-center-bans-topped-500 item (New York, Texas, 150 towns in July alone) plus Amazon's 7.65-GW gas-plant workaround in Pecos County is a fresh site-constraint signal worth tracking alongside the capex anchor data.
Curation section
- The Singularity is now the internet's majority user — machine traffic passed human traffic in May, a year ahead of forecast; Cloudflare's CFO projects machine traffic at 1,000x human within five years.
- Cloudflare's Kitesurf, an agent-first browser in Rust/Wasm, passes 215,000 Web Platform Tests on a fraction of Chromium's CPU footprint.
- Anthropic is making auto mode the Claude Code default after evals showed it blocked 89% of harmful actions, vs. only 13.6% of paid human raters refusing the same prompts.
- Australia's first autonomous cyber incident: an agent booking a gym class found an unguarded API, over-booked months ahead, bumped a stranger off the waitlist, and couldn't undo the action.
- New Orleans became the first major US city to let AI answer 911 calls.
- Chinese systems hold nine of the top ten text-to-video model slots, read as the leading edge of world models (the substrate for humanoids/robotaxis) while US policy attention stays on chatbots; Chinese makers took 97% of global humanoid shipments in a half that tripled to 19,100 units.
- Britain's special-forces spy drones were found sending data to China via Chinese-made cameras — a supply-chain angle on the same hardware race.
- Data-center sitings are colliding with local government: bans topped 500 (New York and Texas joined, 150 towns in July alone); Amazon's workaround is a 7.65-GW gas plant in Pecos County permitted for 33M tons CO2/year, double the nation's worst single emitter.
- Beijing is unleashing $28T in stock/bond markets to fast-track AI-adjacent IPOs (e.g., CXMT, up 500% on debut).
- Google/DeepMind leadership: sources claim Demis Hassabis wanted out alongside Jeff Dean but was kept in the DeepMind chair to protect Google's stock; Tim O'Reilly reads the shakeup as a bet on diffusion (selling TPUs/cloud) over pure invention.
Related
- [[2026-07-30-innermost-loop-ai-capex-fragility-digest]]
- [[2026-08-06-innermost-loop-terafab-agent-plugins-memory-wall]]
- [[2026-08-08-innermost-loop-recursive-self-improvement-cyber-risk]]