06-reference

innermost loop dreaming memory rsi campaign issue

2026-11-09·reference·source: The Innermost Loop·by Alex Wissner-Gross
agent-memoryrecursive-self-improvementai-safety-evalsanthropic-ipodata-center-moratoriumsuperintelligence

Why this is in the vault

Daily digest woven around a Jensen Huang personal-scandal lede, then moving through RSI-as-midterm-campaign-issue, a METR/Ajeya Cotra capability-doubling read, Anthropic's "dreaming" memory-consolidation feature, a Nous finding on cheap calibrated judges matching frontier models, and the usual IPO/compute/space tail.

Mapping against Ray Data Co

The sharpest concrete connection is Anthropic's "dreaming" feature, which shrank Claude's own memory from 413 files down to nine. That is the exact problem RDCO's own ~/AGENTS.md Compaction section and ~/.claude/state/working-context.md exist to solve — periodic consolidation of accumulated session state down to the load-bearing subset (thread context, open decisions, Errors & Corrections) rather than letting memory sprawl unbounded. Anthropic shipping this as a first-party feature is a signal the pattern has graduated from a bespoke RDCO SOP to table-stakes agent infrastructure; worth checking whether a future Claude memory primitive could replace or simplify the hand-rolled compaction discipline rather than just running alongside it. Extends the same thread as [[2026-09-08-stratechery-write-things-down-harness-memory]] (externalize memory as text) and [[2026-09-21-indy-dev-dan-self-compact-pi-agent-zero-hype-devlog]] (self-triggered compaction loops) — this is the first sighting of a frontier lab doing it to its own production agent at the scale (413→9 files) RDCO's vault operates at.

Second, RSI formally arrived as a 2026 midterm campaign issue — Rep. Casar's "Ban RSI," POTUS declaring "100% in favor of Super Intelligence," Bloomington's local ban, and a predicted blue-wave frontier-AI-law wave by spring — which is dated, concrete continuation of the exact moratorium arc already tracked in [[2026-11-06-innermost-loop-post-midterm-moratorium-rsi-sandbagging]] and [[2026-11-04-stratechery-ai-on-the-ballot-data-center-backlash]]. Same "political permitting is now the binding constraint" phase marker for [[project_investing_markov_capital_cycle]], now with the RSI question itself (not just data-center siting) becoming partisan-coded.

Third, METR's finding — read through Ajeya Cotra as ~10x uplift and "superintelligence within a year," with time horizons doubling every two months toward century-long tasks, and internal lab models reportedly outrunning the AI-2027 forecast curve (85% accurate so far) — is the same underlying METR capability-doubling data point already logged via AlphaSignal and this same newsletter on 2026-11-06 ([[2026-11-06-alphasignal-gpt6-cyber-metr-capability-doubling]]), now with a named interpreter attaching a specific timeline claim to it. Third-party cross-confirmation across three separate issues in four days is worth one more mark toward [[project_l5_north_star_strategic_direction]]'s L4→L5 runway question, but the "within a year" framing is Cotra's extrapolation, not METR's own claim — flag the distinction if this gets cited forward.

Fourth, smaller but notable: a Nous finding that a cheap calibrated judge model matched Opus 5.5's output quality at a quarter the cost is a live data point for [[feedback_jev_for_typed_set_problems]] — the general pattern of routing a classification/judgment task to a small typed model instead of a frontier model-as-judge is getting empirical support outside RDCO's own tooling.

Curation section

No deep-fetch: every outbound link wraps a substack.com/redirect tracker with no blurb detail beyond what's already captured above in the woven prose — nothing cleared the curation-mode bar of a specific, RDCO-relevant hook not already summarized here.

Related