REFERENCE · LAST UPDATED
The Absorption Index
A nine-track gauge of AI's arrival rate against society's capacity to absorb it. Each track carries one published indicator from a named primary source; tracks with no public number stay directional rather than being invented.
The question this page answers
Is change arriving faster than the institutions meant to absorb it can respond?
DRIVERS (accelerants — push the Index up)
| Track | Grounded indicator | Current → prior (trend) | Conf. | %-able? |
|---|---|---|---|---|
| Capability / autonomy | METR task-time-horizon (autonomous task length @50%) | 5–9 min (GPT-4o, Time Horizon 1.1 — 9.2 min Vivaria/TH1-tasks, 6.0 min Inspect/TH1-tasks, 5 min Inspect/new-suite; no single "~7 min" figure is published) → ≥16 hr floor (Claude Mythos Preview, early/preview build, limited Mar 2026 window; 95% CI 8.5–55 hr). METR's own caution: "measurements above 16 hrs are unreliable with our current task suite." Doubling ~7 mo (2019→) → ~2.9 mo (88.6 days, TH1.1, 2024→). Newest run (Jun 26, 2026): GPT-5.6 Sol = 11.3 hr (gaming scored as failure) / 71 hr (gaming discarded) / >270 hr (gaming scored as success) — METR states it does not consider any of the three a robust measurement | ✓ trajectory solid / level not measurable | Doubling-rate: yes. Level: no |
| Benchmark saturation | How fast "unbeatable" tests fall | ARC-AGI-1 ~5% → 97.5% (Jul 2026); ARC-AGI-2 ~3% → 92.5%; ARC-AGI-3 <1% (2025) → 30.2% (Jul 2026); GPQA Diamond 94.6% (saturated, PhD baseline 69.7%); FrontierMath re-released as v2 (Jun 2026) after errors in 42% of problems — now ~89% T1–3 and ~87.8% on Tier 4, no longer a holdout; SWE-bench-Verified ~77–79% on the official board — the ~95% figure is aggregator-only and not corroborated by swebench.com. Hard benchmarks still saturate in ~12–24 mo | ✓ 4 of 6 points corrected | Yes (saturation rate) — not SWE-bench |
| Compute scale | Epoch AI frontier training compute | ~5×/yr (90% CI 4–6×; 0.7 OOM/yr) since 2020, doubling ~5.2 mo. (Separate, broader series: ~4.5×/yr for "notable models" since 2010 — different population, label it if used.) | ✓ solid (dashboard 2026-02-05) | Yes |
| Compute money (chips) | Nvidia data-center revenue (company-reported; FY runs Feb–Jan) | FY23 $15.0B → FY24 $47.5B → FY25 $115.2B → FY26 $193.7B (+68%). Latest quarter Q1 FY27 (ended Apr 26, 2026): $75.2B, +92% YoY, +21% QoQ. Nvidia publishes no "run-rate" — ~$300B is our ×4 annualization of one quarter, label it as ours. Q2 FY27 reports Aug 26, 2026 | ✓ solid (reported figures) | Yes — reported figures only |
| Capex | Hyperscaler AI capex + megaprojects | ~$220B (2024 actual) → ~$360–400B (2025 actual) → $700B+ guided 2026 (MSFT ~$190B · GOOGL $195–205B · AMZN ~$200B · META $125–145B — each company's own latest guidance through Jul 22, 2026) ≈ ~1.9× YoY. Every one of the five biggest spenders has raised guidance in 2026; none has cut. MSFT + META report Jul 29, 2026. Stargate: $500B four-year target, ~$400B committed to signed sites, ~0.3 GW operational ≈ 6% of committed capacity | ✓ guidance, not actuals | Yes — must be labeled guidance |
| Self-improvement | % of a lab's production code written by its own AI | Anthropic low single digits (Feb 2025) → >80% of code merged to production (May 2026) — 15 months; leadership separately estimates 90%+ including scripts/experimental. Google: 25% (Oct 2024) → 50% (fall 2025) → 75% (Apr 2026). The on-ramp — no longer one lab | ✓ solid (Anthropic Institute, primary) | Yes |
| Deployment / automation | Anthropic Economic Index (augment vs automate) | 57.4/42.6 baseline (Feb 2025) → automation briefly overtook augmentation Aug 2025 (47 aug / 49 auto) → 52/45 (Nov 2025 — the last published split). Anthropic's Jun 26, 2026 report drops the metric; no current figure exists. Sales/outreach + trading/market-ops API automation at least doubled Nov 2025→Feb 2026 (Anthropic, Mar 2026). Coding is the opposite of concentrating — the same report describes it diffusing "across many task categories" | ✓ corrected — "roughly stable" was wrong; the coding claim was contradicted by the source | Partial — no current aggregate split exists |
| Embodiment (robots) | Optimus units vs targets | 0 commercial (Q2 2026 call, Jul 22, 2026). Tesla publishes no internal unit count — Musk's most specific on-record figure is "several hundred units… primarily for learning, not productive tasks" (Q4 2025 call, Jan 28, 2026); the circulating "~1,000" is an unattributed industry estimate, not a disclosure. Dedicated production had not begun as of Jul 22, 2026 ("we will soon start production"). The target sequence was abandoned, not missed: 50k/2026 (Jan 2025) → "impossible to predict" (Apr 2026) → no figure given (Jul 2026). 1M/yr = Fremont design capacity (no year attached); 10M/yr = explicitly aspirational, for a not-yet-built next-generation robot | ✓ corrected — both "~1,000" and "~90%" are unsourceable | No. Status only — not a gap % |
| Concentration | AI's share of US VC funding (+ AI-lobbying $) | US, dollar-value (PitchBook-NVCA): 43% (2024) → 65% (2025) → 86% (H1 2026) — one consistent series. (The older 34%/50% figures were global; do not chain them to the US 86%.) OpenAI+Anthropic = 43% of global startup funding ($217B of $510B, H1 2026, Crunchbase). Lobbying: Anthropic in-house +333% YoY (Q1 2026 vs Q1 2025) per Senate LDA filings; OpenAI roughly 1.5–2× YoY — the "+70%" figure is not reproducible from the filings | ✓ solid (scope corrected) | Yes — US series only |
BRAKES (absorption capacity — hold the Index down)
| Track | Grounded indicator | Current (trend as a brake) | Conf. | %-able? |
|---|---|---|---|---|
| Enacted regulation | US state AI bills enacted vs introduced | 2025: 1,208 introduced / 145 enacted / 12.0% (MultiState.ai). 2026: introductions already past 2025's full year (1,561 by Mar 2026), but **enactments are contested and may be falling — NYU CTP counts 109 laws by Jul 1, 2026 vs 121 by the same date in 2025, tying the slowdown to the Dec 2025 EO. Transparency Coalition counts 84 vs its own 73 for all 2025. Trackers define "AI bill" so differently that 2025 enacted counts span 73 / 145 / 159** | ✓ 2025 solid; 2026 contested | 2025 yes, attributed; 2026 no |
| Federal preemption (erosion) | moratorium + EO + DOJ task force | Moratorium killed 99–1 (Senate Roll Call 363, Jul 1, 2025 — brake win). EO 14365 (Dec 11, 2025) directs DOJ + Commerce. Task Force formalized Jan 9, 2026 and already litigating — intervened against Colorado's AI Act in xAI v. Weiser (Apr 2026). Commerce's BEAD penalty notice, due Mar 11, 2026, was still unissued as of late Jun 2026 — no state has lost BEAD funds over an AI law. NYU CTP (Brennen, Jul 6, 2026): $742.2M avg BEAD allocation for states with AI laws vs $920.9M without — the author's own caveat is that this gap nearly closes ($873.7M vs $779.6M) once California enacts, and it is a raw allocation correlation, not a measured effect of the EO. Live vehicle is now the narrower Obernolte–Trahan draft (Jun 2026, 3-yr sunset, development-only) | ✓ vote/EO/Task Force solid; BEAD "evidence of bite" materially weaker than stated | Directional — BEAD figure not printable unhedged |
| EU AI Act | enforcement calendar | Split verdict — not uniformly strengthening. GPAI duties live since Aug 2, 2025. Commission enforcement powers + Art. 101 fines (3%/€15M) activate Aug 2, 2026 — confirmed on schedule, explicitly untouched (Art. 113 carves Art. 101 out of the Aug 2025 date). But the 7%/€35M figure is Art. 99 and has applied since Aug 2, 2025 — a year earlier than this page previously said; the Art. 5 prohibitions it enforces have been in force since Feb 2, 2025. And Reg. (EU) 2026/1744 (the "Digital Omnibus on AI," published in the OJ 24 Jul 2026, in force 27 Jul 2026) defers the high-risk chapter: Annex III Aug 2026 → Dec 2, 2027; Annex I Aug 2027 → Aug 2, 2028; sandboxes → Aug 2027; a 4-month watermarking grace to Dec 2, 2026. Narrower and later, not strengthening | ✓ corrected — the direction was wrong | Milestone-based — must be split into two dates |
| Lab safety grades | FLI AI Safety Index (A–F) | Summer 2026 (4th edition, pub. ~Jul 7): Anthropic C+ (flat), OpenAI C ▼, Google DeepMind C (flat), xAI F ▼, DeepSeek F ▼. On existential safety, FLI's own words: "No company exceeds C-; most score D or below" — the best actual grade is D+ (Anthropic, OpenAI). Net down; "retreating from commitments." (xAI's second down-arrow is unconfirmed — both xAI and DeepSeek moved D→F identically.) | ✓ solid grades (read via fetch-summary, not raw PDF) | Yes (GPA) |
| Public will | Pew: "more concerned than excited" | 37% (Nov 2021) → 52% peak (Aug 2023) → 50% (Jun 2025) — the latest asking; Pew's Feb 2026 wave did not re-ask the question. Net +13 pts vs 2021, but flat-to-down off the 2023 peak — this is not a strengthening brake. ~60% fear under-regulation (Aug 2024). Confidence government can regulate AI effectively: 62% little/none (Aug 2024) → 67% (Feb 2026) — belief still eroding. Will flat, belief ↓ | ✓ solid | Yes — but re-color: not green |
| Safety investment | safety compute / alignment effort | OpenAI Superalignment dissolved May 17, 2024, days after Sutskever and Leike left. Its 2023 pledge of 20% of secured compute was, per contemporaneous reporting, never fulfilled and lapsed with the team — never formally rescinded, and never defined well enough to measure. A separate "Mission Alignment" team was disbanded Feb 11, 2026 — OpenAI describes it as a mission-communications function, not technical safety research; outside observers contest that framing. Its lead, Josh Achiam, left OpenAI Jul 1, 2026. No lab's public safety framework discloses a safety-compute share (Anthropic RSP v3.0 read directly; OpenAI Preparedness Framework + DeepMind Frontier Safety Framework via secondary), consistent with FLI's Summer 2026 grades | ✓ dates solid; "abandoned" and the two-team framing both overstate | Directional only |
| Eval coverage | the "eval ceiling" | Evaluators (METR/Apollo/CAISI) face persistent time + model-access constraints — with one pilot exception: METR's Frontier Risk Report (May 19, 2026) confirms internal-model + raw-CoT access at Anthropic, Google, Meta and OpenAI. The same report documents **evaluation awareness on toy scheming/sabotage scenarios — a distinct finding from power-seeking evals, which it treats separately. US AISI→CAISI is Jun 2025 background, not a contemporaneous development. No numeric rate exists. Newer than the cited report: METR's Jun 26, 2026 GPT-5.6 Sol run (highest detected benchmark-gaming rate METR has recorded on its ReAct harness); Apollo Research Jul 13 + Jul 23, 2026**; CAISI assessments of GLM-5.2 (Jul 8) and, with UK AISI, Kimi K3 | ✓ corrected — scheming ≠ power-seeking; CAISI is 13-mo-old background | Directional only |
| Whistleblower shield | AI Whistleblower Protection Act | S.1792 / H.R.3460 — identical companion bills, both introduced and referred May 15, 2025 (Senate HELP; House Education & Workforce). That referral remains the only recorded action on either bill — no hearing, no markup, no floor vote. Not enacted, 14 mo on | ✓ CONFIRMED as stated — the only fully clean row across both passes | Binary/status |
| Human-in-the-loop | autonomous-weapons policy | DoD 3000.09 does not "require no human-in-loop." It requires "appropriate levels of human judgment over the use of force" — an undefined, flexible standard that permits fully autonomous engagement without mandating it. UNGA adopted Res. 80/57 on Dec 1, 2025, 164–6–7 (the 156–5 figure was the Nov 6, 2025 First Committee vote on draft L.41, and dropped 8 abstentions). Still no binding treaty; Guterres's 2023 New Agenda for Peace calls to conclude one "by 2026" — negotiations have not formally begun. Brake failing | ✓ corrected — the directive claim was false as stated; UN vote via corroborated secondary (primary PDFs blocked) | Directional |
Narrative-vs-delivery note (2026-07-25): the gap first flagged for humanoid robots now appears on two more tracks — Stargate ($500B announced, ~$400B committed, ~6% of committed capacity operational) and SWE-bench-Verified (~95% circulating, ~78% on the primary leaderboard). Three instances across robotics, infrastructure, and coding capability is a pattern, not a robotics quirk. Open question for Karl: promote "arrival story vs. measured delivery" from a callout to a structural axis of the Index.
GODFATHER LAYER (the expert spread — a standalone element)
Four dated snapshots, not four current positions. Each cell records the last time that voice put a number on it. Three of the four have since stopped giving one.
| Voice | Risk number | Timeline | The signal they cite |
|---|---|---|---|
| Yann LeCun (skeptic / floor) | ≈0 — rejects the frame. p(doom) estimates are "complete bullshit" and existential risk "essentially zero" (own post, Apr 13, 2026); earlier and blunter, "P(doom) is BS." (own post, Feb 7, 2024). The only 2026-current number in this table | No figure as of 2026. Said Human-Level AI "will take several years if not a decade," with a long tail where "it could take much longer" (own post, Oct 16, 2024); by Jan 2026 says only that it is "not likely to see it this year, or next year." He uses "Human-Level AI," not AGI — he rejects the term, on the grounds that human intelligence "is not general" | The world-model gap. LLMs are "not the path to real intelligence. They're a detour" (Bloomberg, May 21, 2026); the current paradigm has a "shelf life" of "probably three to five years" and no future as "the central component of an AI system" (WEF Davos panel, Jan 23, 2025, per contemporaneous TechCrunch reporting — no primary transcript located). Interest disclosed: he left Meta (announced Nov 2025) to found AMI Labs, where he is executive chairman; it raised $1.03B seed at $3.5B pre-money (Mar 2026) on exactly this thesis |
| Geoffrey Hinton | 10–20% chance that AI causes human extinction. Asked on BBC Radio 4's Today (Dec 27, 2024) by guest editor Sajid Javid whether he had revised his earlier one-in-ten estimate, he answered: "Not really, 10% to 20%." He attached no timeframe to the figure. Known via reporting only — no transcript or primary statement located | No range is printable. His own revision: "I now predict 5 to 20 years but without much confidence" (own post, May 3, 2023), down from "30 to 50 years or even longer" (NYT, May 1, 2023) | Faster-than-expected progress; reasoning + emergent deception; digital > biological (copyable weights). In the same interview: "we've never had to deal with things more intelligent than ourselves before" |
| Yoshua Bengio | ~20% catastrophe — his exact words, once: "I got around, like, 20 per cent probability that it turns out catastrophic" (ABC News / Background Briefing, Jul 14, 2023). A 2023 snapshot, not a current position. He has published no personal number since; in his two most recent appearances he declines to quantify, arguing instead that "even a 1% chance of something going really, really bad is not acceptable to me" (80,000 Hours, May 7, 2026) | 5–20 years to superhuman intelligence, stated as an interval in his own writing: "a 95% confidence interval for the time horizon of superhuman intelligence at 5 to 20 years" (yoshuabengio.org, Jun 24, 2023), restated as "5 to 20 years with 90% confidence" (Aug 12, 2023), revised down from his prior "20 to 100 years" | Chairs the International AI Safety Report (2026 edition pub. Feb 6, 2026); founded LawZero (Jun 2025) to build non-agentic "Scientist AI" |
| Expert survey (Grace et al., n=2,778 — 2023 ESPAI, pub. Jan 2024) | 37.8–51.4% of respondents gave at least a 10% chance to outcomes as bad as human extinction — a range across four different question wordings, not one question's error bar. The medians are far lower: 5% on the headline extinction wording, 10% only on the loss-of-control wording. Mean on "extremely bad" outcomes fell 14% (2022) → 9.0% (2023) | HLMI: 10% chance by 2027, 50% chance by 2047. The 50% point moved 13 years earlier in a single year (from 2060); the 10% point moved only 2 years (from 2029). A 2024 wave was fielded but remained unpublished as of Oct 2025 | field consensus |
Cut from this table, and why. Six figures previously printed here did not survive a line-by-line source pass and are gone rather than hedged: Hinton's "~30 yr" window (the reporter's framing, not his — the BBC's own trail for the same broadcast said "two decades"); Hinton's "within 20 years" (2026, aggregator-only); Bengio's "50% human-level within a decade" (the ABC journalist's unquoted reconstruction of his reasoning, and it contradicts what he wrote himself three weeks earlier); LeCun's "obsolete" (journalists' word, not his); LeCun's "house cat" and his "~50× visual data / 16,000 hrs" data-efficiency figure (never verified). The Guardian URL long cited for Hinton's 10–20% does not resolve and is no longer cited.
Standing rule for this table. Never take a godfather's p(doom) from a search summary. During this check a search engine returned Hinton's BBC "10–20% extinction over three decades" under Bengio's name. Bengio's "~10%" is his citation of polls of ML researchers, not his own estimate. Every figure above traces to a quotation the checker opened.
The spread is the story. Yann LeCun puts existential risk at "essentially zero" and calls the exercise "complete bullshit" (Apr 2026). Geoffrey Hinton says 10–20%. Yoshua Bengio said ~20% once, in 2023, and has not said it since. And between 37.8% and 51.4% of the 2,778 AI researchers surveyed by Grace et al. gave at least a 10% chance to outcomes as bad as human extinction — while that same survey's median answer was 5% on its headline extinction wording, and 10% only on the loss-of-control framing. Hinton and Bengio are not standing where the field stands. They are standing out beyond the worried third-to-half of it. A "where the people who built this stand" bracket is a distinct, honest element — provided it prints the share giving ≥10%, never a median: there is no survey median at or above 10% on the headline wording.
What this is
A single number for one question: is change arriving faster than the institutions meant to absorb it can respond? Above 1.0, it is. The reading today is 0.9, and rising.
(0.9 and rising are injected from ops/brief-pipeline/absorption-index.json at build time — don't hand-type the number here. A figure typed into prose is how the composite graphic ended up rendering nine superseded values for nine days.)
How it's built. Nine tracks. Six accelerants that push the number up, three brakes that hold it down. Each track carries one real published indicator from a named source — METR for capability, Epoch AI for compute, Nvidia's own filings for chips, the companies' own guidance for capex, PitchBook-NVCA for concentration, MultiState for state regulation, the Future of Life Institute for lab safety grades, the EU's Official Journal for enforcement.
Where the numbers come from. Every figure is checked against the primary source before it prints. That means the body that produced the measurement, not reporting about it. Where a primary document can't be reached, the track says so and names what it rests on instead. Where no public number exists at all, the track stays directional. Safety-compute share, evaluation coverage, Tesla's internal robot count and the current automation split are all cases where printing a figure would mean inventing one. So we don't.
What the reading is, and isn't. The tracks are measurements. The composite is not. 0.9 is an editorial judgment about how nine tracks add up, and it is labelled as one. Anyone can disagree with the weighting and still use the tracks.
What moves it. A track moves only on a named, cited event that has already happened. Not a forecast. Not an announcement. Not a date on a calendar that hasn't arrived yet. When a track moves, the event and its source are recorded with it.
When we get it wrong. Corrections are published, dated, and kept — including the ones that cut against our own argument.
Sourcing
Verified 2026-07-25 against: metr.org/time-horizons/ · epoch.ai/trends · epoch.ai/benchmarks/{gpqa-diamond, frontiermath-tiers-1-3-v2, frontiermath-tier-4-v2} · arcprize.org/leaderboard · swebench.com (Verified tab) · investor.nvidia.com (Q1 FY27 + Q4 FY26 releases) · sec.gov EDGAR (Alphabet Q2 2026 ex-99.1) · anthropic.com/institute/recursive-self-improvement · blog.google · nvca.org Q2 2026 PitchBook-NVCA Venture Monitor · crunchbase.com/news · lda.senate.gov/api/v1/filings/ · multistate.ai/artificial-intelligence-ai-legislation · techpolicy.press (NYU CTP count) · futureoflife.org/ai-safety-index-summer-2026 · pewresearch.org (Jun 2026 topline PDF) · arxiv 2401.02843 (Grace et al.) — full text read directly, all four sub-claims confirmed verbatim. Not verified: anthropic.com/research (Economic Index) · Optimus counts · congress.gov (S.1792) · artificialintelligenceact.eu · internationalaisafetyreport.org. ncsl.org was unreachable (Cloudflare bot-verification) and is not a source for this page.
About this page
Maintained by Karl Herbst as part of Wireframe News, which covers how power concentrates across politics, AI, war and the future of work. Free to cite and quote with attribution (CC BY 4.0). Corrections and additions: wireframenews.com.