writing / instrument

Compression Scorecard

commitments dated June 2026 — scored January 2030 — updated 2026-07-23

thesis failsjagged middle ◂ currentconfirmed

Tracking since June 2026. First full readings expected late 2026.

indicatorfails ifconfirms ifreadingstatus
Entry-level postings, AI-exposed vs matched non-exposed occupations (2022 baseline)exposed occupations recover to within 15% of peers while overall employment stays healthyexposed occupations stay ≥30% below peers while sector output growsBaseline set; read as exposed-vs-peer difference (2022 was a bubble year for cognitive hiring). Exposure list pre-registered via a published exposure index (Eloundou et al. primary; Felten et al. robustness check). Sources: postings data (Lightcast/Indeed) and BLS OES. First reading due late 2026.PENDING
Wage premium: new-hire wages at fixed percentiles, mid-skill cognitive vs embodied service workstabilizes or widenskeeps compressingBaseline set; read from new-hire wages at fixed percentiles, not incumbent averages — survivor composition would otherwise mask compression. First reading due late 2026.PENDING
Major cognitive sectors running routine production above 10:1 supervision (σ method, enterprise boundary, exposure-index universe)none, outside demos and marketing≥2 sectors0 confirmed. Software is closest: Anthropic discloses >80% of merged code is model-written, but no sector-wide ratio series exists yet.Anthropic disclosure, June 2026MIXED
Firm signature: output per firm vs headcount per firm, cognitive industriesoutput and headcount move togetheroutput rising while headcount fallsAnecdotal divergence in software; no clean series identified yet.PENDING
New-task channel: employment and posting share in occupation titles that did not exist in 2022new categories absorb displaced cognitive workers at scalenew-title share stays marginal while exposed postings stay depressedThe scorecard's declared blind spot: none of the other indicators would see this. Requires SOC revisions and new-title postings tracking; first reading due 2027.PENDING
Takeoff signature: frontier-lab revenue and shipped output vs lab headcount, plus lab-internal supervision (leading indicator, added July 2026)stays quiet alongside quiet broad indicators — lab headcount and output rising together, lab-internal ratios stubborn (augmentation won everywhere; no third door)fires while the broad indicators stay quiet — lab revenue and output decouple from headcount, lab-internal σ collapses, machine work graded by machines (compression concentrated at the frontier; diffusion queued)Added 2026-07-23 as a dated amendment (see the essay's takeoff-route paragraph); the original five indicators and thresholds are unchanged. Early signal: >80% of merged code model-written (Anthropic, June 2026); insider reporting of lab revenue growth far outpacing headcount (Kokotajlo, July 2026). No clean series yet.Anthropic disclosure, June 2026; Kokotajlo, Diary of a CEO interview, July 2026MIXED

← back to the essay