Skip to content
Slow Variables
Ask

Diffusion / bucket 5

Return arrow

Stock
Deployment feedback flowing back into invention — data, RL environments, continual learning.
Speed limit
Whether the deployment→training channel is open and how far it substitutes for external bottlenecks.
Valve measured
leak

Indicators

Lab recoupment ratio
emerging
0.27×as of 2026-09-30
Agent-workdays per human workday in frontier research (self-reported)
emerging
3.10×as of 2026-08-15
Human interventions on 4–8 hour agent tasks (self-reported)
emerging
50.0%as of 2026-07-31
Continual-learning ladder (highest production rung)
emerging
3.00 rungas of 2026-05-27
Active RL-environment vendors (directory count)
emerging
23as of 2026-10-05
Expert-data market run-rate (Mercor)
emerging
$2.0Bas of 2026-06-30
Vendors claiming shipped continual learning
emerging
2as of 2026-11-04
Fastest production refresh of a continually updated model (hours)
emerging
1.00 hoursas of 2026-05-27

Continual-learning ladder

How far products have climbed the ladder of learning on the jobfilled up to the current production level
  1. L6Six-month-employee test passed

    A frontier system passes a public benchmark for knowing an organisation the way a six-month employee would. What would show it: the benchmark exists and a frontier system passes it.

  2. L5Per-session weight updates that persist and pool

    Updates from individual sessions persist and pool without forgetting. What would show it: a lab release branded continual or online learning; a forgetting benchmark; per-user weight divergence.

  3. L4Organisation knowledge in weights

    Firm knowledge internalised in weights with measured retention. What would show it: cl_forgetting_retention; firm-knowledge lift; cost per query.

  4. L3Per-customer adapters on a cadencecurrent production level· as of 2026-05-27

    Adapters per deployment, refreshed hourly to weekly, routed and versioned. What would show it: cl_refresh_cadence_hours; cl_adapters_per_product; per-customer lift over time.

  5. L2Population-level online learning

    One objective, all users; the model updates from aggregate usage. What would show it: refresh cadence; documented lift.

  6. L1Gradient-free memory

    Feedback distilled into retrievable rules; no weight change. What would show it: pass^1 lift with memory; hold rate; transfer across models.

  7. L0Static snapshot

    Static snapshot; memory is files in context. What would show it: none.

Production rows are vendors describing their own systems (tier 7 unless independently evaluated); research rows never move the level. The rule: rungs 0 to 3 normal, 4 emerging, 5 and 6 fast.

Source: Cursor research blog; Baseten blog and research posts; Harvey research and insights posts; Engram blog; Google Research blog; Thinking Machines Connectionism posts; arXiv abstract pages (manual rows). Cursor (Anysphere). Baseten. Harvey. Engram. Google Research. Thinking Machines Lab. arXiv.

Crosswalk to the capture lens

  • Training input (return arrow) — Expert data, RL environments and continual-learning infra are the deployment→invention channel; lab recoupment ⇄ capex-to-revenue.
  • Model (return arrow) — Inside the labs: agent work on the labs' own research and lab recoupment are deployment feeding back into invention.