Compare
The same evidence read against four families of claims: AI as Normal Technology, AI 2027, the labs' own timelines, and the capture theses about who pays for the build-out. Each claim sits in its family's column by the ledger it is filed under, citing the sourced claim on the prediction ledger. The last column is not an opinion: for a diffusion indicator it follows from the current status (faster than normal leans AI 2027; consistent with normal or slower leans Normal Technology; anything else is open). A capture indicator reads concentrating or dispersing, which neither lean describes.
7 lean Normal Technology · 2 lean AI 2027 · 6 open
| Indicator | Normal Technology expects | AI 2027 expects | Lab timelines say | Capture theses say | Leans |
|---|---|---|---|---|---|
| METR 50% time horizon faster than normal17.4 h(8.5 h–55.1 h)⚑disputed: METR: measurements above 16 hours are unreliable with the current task suiteas of 2026-04-07 | Capability can move fast; the paper's speed limits sit downstream of methods. A short doubling time does not contradict Normal Technology on its own. | Doubling every four months or faster from 2024, continuing to shorten, is the premise of the superhuman-coder timeline. | no claim | no claim | AI 2027 |
| METR 80% time horizon faster than normal3.1 h(1.6 h–6.6 h)as of 2026-04-07 | The 80% horizon is the number that matters for products; it should lag the 50% horizon by years because reliability fixes are application-specific. | 80% reliability on tasks taking humans years by March 2027 if the trend keeps accelerating. | no claim | no claim | AI 2027 |
| 50%/80% horizon ratio consistent with normal6.34×as of 2026-03-05(2 obs) | The capability-reliability gap stays wide (ratio of five or more); benchmarks are false summits. | The gap collapses toward two as agents become reliable enough to run unattended. | no claim | no claim | Normal Technology |
| Developer productivity uplift (METR RCTs) consistent with normal-4.0%⚑disputed: METR itself calls this 'very weak evidence' because developers refusing the no-AI arm biases the sample.as of 2026-02-24 | Modest, setting-dependent developer uplift; AI has not replaced software engineers and will not. | Broad, large uplift as coding agents take over the work. | Most of what software engineers do, end to end, within six to twelve months of January 2026 (Amodei). | no claim | Normal Technology |
| Enterprise pilots with measurable P&L impact consistent with normal5.0%⚑disputed: Denominator and sample disputed; Menlo reports ~47% of pilots reach production on a different definition. No MIT-hosted copy of the PDF exists; this is the mirror Menlo cites.as of 2025-06-30 | The proof-of-concept-to-product gap keeps most pilots from measurable P&L impact for years. | Agents that work reset the pilot economics; production share rises quickly. | no claim | no claim | Normal Technology |
| Share of work hours assisted by generative AI emerging5.3%as of 2026-09-30 | Single-digit share of work hours in year three, like PCs and the web at the same age. | A fifth or more of work hours assisted as agents diffuse through white-collar work. | no claim | no claim | open |
| US firms using AI (Census BTOS) consistent with normal24.6%(24.1%–25.1%)as of 2026-10-04 | Under a third of firms in year three; organisational speed, not capability speed. | More than half of firms as agents remove the integration cost. | no claim | no claim | Normal Technology |
| US businesses paying for AI (Ramp AI Index) emerging55.9%as of 2026-09-30 | Paid adoption saturates early among tech-forward firms and flattens as the hard integration work begins. | Paid adoption keeps rising toward universality as agents pay for themselves. | no claim | Trillions of dollars of AI revenue before 2030 (Amodei) would need paid adoption close to universal. | open |
| US labour productivity, year on year consistent with normal2.2%as of 2026-06-30 | Productivity at trend for years; the Solow paradox repeats (paper's entries 28, 29, 61). | no claim | no claim | Productivity growth well above trend by 2030 on the extreme path; under half a point of GDP growth on the modest one (Korinek et al.). | Normal Technology |
| US total factor productivity, private nonfarm business consistent with normal0.8%as of 2025-12-31(2 obs) | TFP invisible at annual resolution; Acemoglu's 0.66% over a decade is the normal reading. | no claim | no claim | TFP separates from trend only after 2027, and only on the substantial and extreme paths (Korinek et al.). | Normal Technology |
| US labour share of income, year on year emerging-3.4%as of 2026-06-30(2 obs) | Labour share moves within its cyclical range; no AI-attributable break before 2028. | no claim | no claim | Labour share falls from about 60% toward 45% by 2030 on the extreme path (Korinek et al.). | open |
| Entry-level employment shortfall in AI-exposed occupations emerging19.0%as of 2026-06-30 | No broad AI-attributable employment decline for years; canaries stay within youth-employment swings. | no claim | no claim | Cognitive employment down about 4% by 2030 on the substantial path; entry-level exposed work breaks first (Korinek et al.). | open |
| Cross-tracker concordance (labour) consistent with normal0as of 2026-09-30(7 obs) | The four monthly labour trackers agree on a null. | no claim | no claim | On the substantial path the trackers converge on a decline in exposed occupations by 2030 (Korinek et al.). | Normal Technology |
| Agent-workdays per human workday in frontier research (self-reported) emerging3.10×⚑disputed: OpenAI measuring its own research organisation and publishing the result itself; no independent check (tier 7).as of 2026-08-15 | Agent effort inside labs can rise without escaping external bottlenecks; self-reports do not resolve the question. | Agent effort passes human effort inside labs by 2026 and the R&D multiplier compounds from there. | An automated research intern by September 2026 and an automated AI researcher by March 2028 (OpenAI). | no claim | open |
| Human interventions on 4–8 hour agent tasks (self-reported) emerging50.0%⚑disputed: OpenAI measuring its own research organisation and publishing the result itself; no independent check (tier 7).as of 2026-07-31 | Humans keep steering most long agent runs; unreliability compounds over steps (entry 4). | Long runs go unattended as reliability catches up; intervention rates fall below a quarter. | An automated AI researcher by March 2028 implies long runs without intervention (OpenAI). | no claim | open |
| Semiconductors' share of stack operating income concentrating61.0%as of 2026-06-30(8 obs) | Rents at a chokepoint erode as users capture most of the surplus, the usual course for a general-purpose technology. | Whoever controls the frontier labs and their compute captures a rising share as capability compounds. | no claim | The chip layer's share of stack profit is the question the gap analyses turn on: spend far ahead of revenue (Cahn, Covello). | n/a (capture) |
| Hyperscaler capex to AI revenue (trailing year) dispersing4.38×as of 2026-09-30(34 obs) | Capex runs ahead of revenue for years, as in earlier infrastructure build-outs. | Capex is repaid quickly because capability growth turns into revenue fast. | no claim | Infrastructure spend many times AI revenue, needing revenue in the trillions to justify it (Cahn, Bain, Tunguz). | n/a (capture) |
| Lab recoupment ratio emerging0.27×as of 2026-09-30(12 obs) | no claim | Lab revenue compounds fast enough to fund the next training runs from operations. | Revenue growing about tenfold a year (Amodei). | Trillions of dollars of revenue before 2030 (Amodei) against the capital already raised. | n/a (capture) |
| Circular financing scale concentrating$998Bas of 2026-09-30(34 obs) | Vendor financing is a normal feature of a build-out and unwinds as end demand is proven. | no claim | no claim | Circular deals inflate reported demand and widen the gap between spend and end revenue (Cahn). | n/a (capture) |
Every prediction is scored in full on the prediction ledger.