AI progress isn't one S-curve, it's a stack of them — each one starting its steep climb while the layer below it is already flattening. That's why "pretraining is hitting diminishing returns" doesn't mean AI progress stops. It means the locus of progress moved up the stack.
Ilya Sutskever, accepting NeurIPS 2024's "Test of Time" award, stated: "Pre-training as we know it will unquestionably end." That's a citable data point from the founder of the scaling-law era, not a critic's speculation.
That's one curve out of at least seven running at once. Reasoning and inference-time compute started its own steep climb almost exactly as pretraining's curve began to bend. Agentic orchestration started its climb the year after that. Four much slower curves — vertical adoption, physical robotics, scientific validation, and economic restructuring — are each just beginning multi-year-to-multi-decade climbs of their own, bottlenecked by things that have never moved at software speed: trust, biology, hardware, and institutions.
Press play below and watch these seven climbs overlap.
Started roughly 2012, climbed steeply through the GPT-3 era, and is now the one curve in this stack that the field's own leadership has explicitly called flattening. Ilya Sutskever, accepting NeurIPS 2024's "Test of Time" award: "Pre-training as we know it will unquestionably end." "The 2010s were the age of scaling, now we're back in the age of wonder and discovery."
Source: NeurIPS 2024, Dec 2024, Vancouver — techmeme.com/241213/p33
Started almost exactly as curve 1 began to bend — OpenAI's o1 (Sept 2024), then DeepSeek-R1 (Jan 20, 2025) proved a second, independent lab could reach frontier reasoning via reinforcement learning rather than more parameters. (The pure-RL characterization applies to the R1-Zero research variant specifically; the shipped R1 model used a cold-start supervised fine-tuning phase before RL.) McKinsey projects inference will become more than half of AI compute demand by 2030 — not the "75%" figure sometimes quoted, which doesn't trace to McKinsey. Separately, McKinsey also projects $7 trillion in AI data-center infrastructure investment, from a different report.
Sources: Fireworks.ai DeepSeek-R1 deep dive · McKinsey, Future of AI Workloads · McKinsey, $7T Data Center Build-Out
Gartner reports: "40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025" — an 8x jump in enterprise-app penetration in a single year. McKinsey Global Institute separately projects AI agents and robots could generate roughly $2.9 trillion/year in US economic value by 2030, at ~27% average automation of current work hours. This is the one curve in the stack where adoption and capability are visibly moving together, fast.
Sources: Gartner, Aug 26 2025 · McKinsey Global Institute, Nov 2025
No single clean statistic covers "verticals" as one category — by nature, this curve fragments per industry. Coding is furthest along (digital-native, low regulatory friction, roughly 2022 onward); medicine and finance lag hardest, gated by liability, compliance, and legacy-system integration. Curves 5 and 6 below are this curve's two most measurable sub-cases.
2026 is being explicitly called a "validation year, not proof of mainstream adoption" in current trade coverage. Roughly $6 billion went into world-model companies in Q1 2026 alone — world models being synthetic, physics-aware training data, since unlike text, "how physical matter actually behaves" doesn't already exist pre-scraped on the internet. China's National Venture Capital Guidance Fund targets roughly $138B for AI/robotics, but as a 20-year commitment, not a lump sum. However, most humanoid robots today run in the 90-minute-to-8-hour range per charge (not a single universal figure), and policies hitting 95% success in the lab commonly drop to roughly 60% in the real world.
Source: TechTimes, June 2026 · China's $138B fund, AI Insider
The design phase is transformed; the validation phase is untouched — and that split is the cleanest "capability curve ≠ value curve" example in the whole stack. AlphaFold raised proteome structural coverage from roughly 48% to 76% — a large, uncontested acceleration of the design phase. (A commonly repeated claim that AlphaFold cut experimental structure-determination work by 60-70% is contradicted by the peer-reviewed literature; experimental determination rates have stayed almost unchanged.) As of this writing, zero FDA approvals exist for AI-discovered drugs, and Phase II success rates for AI-derived candidates (~40%) are running at historical norms, not above them. The bright spot: Insilico Medicine's rentosertib is the first molecule combining an AI-discovered target with an AI-designed drug to show a Phase IIa efficacy signal (Nature Medicine, June 2025), and AI-native programs are running roughly 80-90% Phase I success versus a historical baseline near 52%. The gap that remains is upstream of what AI actually accelerates — it was never the bottleneck AI is good at removing.
Sources: PNAS, AlphaFold structural coverage · Insilico Medicine, Nature Medicine publication
This is now a named, modeled phenomenon, not just a vibe: current economic commentary explicitly frames it as Solow's productivity paradox with "computer age" swapped for "artificial intelligence." Global AI investment exceeds $500B/year (Goldman Sachs, hyperscaler capex); OECD total factor productivity growth remains flat (roughly 0.4% in 2024) — two separate, real findings, not one combined study. The historical precedent has a specific, well-sourced lag: Edison's Pearl Street Station opened September 4, 1882, and the productivity impact of electrification didn't show up in US manufacturing data until the 1920s — a documented ~40-year lag (Paul David, The Dynamo and the Computer, American Economic Review, 1990), because factories initially just swapped steam engines for electric motors in the same layout, and only later redesigned around distributed power. The named framework for why: the Productivity J-Curve (Brynjolfsson, Rock & Syverson, NBER 2018, published in American Economic Journal: Macroeconomics 2021) — value capture requires expensive complementary investment that initially masks a technology's true potential rather than revealing it. Most firms today have bought the AI technology but not done the harder organizational transformation, so "AI isn't paying off yet" may reflect deferred value, not absent value.
Sources: Paul David, 1990, AER · Brynjolfsson, Rock, Syverson, NBER 25148 · OECD Compendium 2026
Every curve in the fast software stack (1 to 2 to 3) is capability moving at software speed. Every curve in the slow diffusion stack (4 through 7) is capability running into something that has never moved at software speed — biology, hardware, or institutions — and it won't start now just because the underlying model got smarter. Pretraining flattening is real. It's also one row in a seven-row chart, and the doom cycle and the hype cycle are both reading from the same single row.
Deployment velocity, resistance, and adoption dependence already have dedicated fields in our 30-industry AI-effects matrix — this framework is the lens the whole engine is built around. AI Stock Market Impacts models how AI reshapes relative value across 30 industries and roughly 180 cross-industry effects, recalibrated continuously since February 2026, with every parameter change logged in public.
See the 30-industry matrixPart of the same honesty-first family as Hype vs. Reality, The Human Bottleneck, and What Abundance Theory Ignores — this page is the timing layer underneath all three.