Between summer 2025 and spring 2026, three layers of evidence — hard capability events, insider acknowledgement, and lay-observer dissonance — converged on a single conclusion: the relationship between human and cognitive work had structurally changed. The mechanism is best read not as a “second cognitive revolution” in the species-scale Harari sense but as a price collapse in the cost of cognition — and its practical consequence is that the durable human contribution shifts from cognitive labour to cognitive direction.
Overview
In the period from summer 2025 to spring 2026, public discourse on AI capability passed through what behaves, in retrospect, like a bellwether moment — a sequence in which independent layers of evidence converged on the same conclusion. Hard capability events came first (GPT-5 producing novel mathematics; founders restructuring portfolios for “post-AGI”). Industry insiders acknowledged the shift next (Sam Altman saying AGI “kind of went whooshing by”; Karpathy saying he had never felt this far behind as a programmer). Lay observers articulated the dissonance last. Three layers; one inflection.
This article reads that convergence as a leading indicator and traces what it implies for cognition, learning, and the shape of expertise. The argument is that the underlying mechanism is economic — a collapse in the cost of cognitive labour — and that the durable response is the cultivation of meta-skills: strategic thinking, critical thinking, emotional intelligence, taste, synthesis, and calibration.
Key concepts
- Agent Harness — the layer in which much of the 2025–26 capability gain accumulated
- Context Window — the finite buffer that meta-skill judgement now shapes
- The capability gap — the parallel argument that practitioners and casual users see different products
- Cognitive cores and external memory — the model-design view of the same shift
The bellwether of 2025–26
The pattern is not a single dramatic event but a sequence — capability events first, insider acknowledgement next, lay-observer dissonance last. Each layer reinforces and extends the one before.
Capability events: summer 2025
The hardest-to-dispute data points cluster in summer 2025. On August 20, VraserX circulated a now-noted demonstration in which Sebastien Bubeck handed GPT-5-Pro an open problem in convex optimisation — one humans had only partially solved — and watched the model produce a correct proof improving the known bound from 1/L to 1.5/L over a seventeen-minute reasoning run. Bubeck verified the result. Humans later closed the gap further, but the model independently advanced the research frontier. Until 2025, this kind of event lived in speculation about what a future AI might do; in August it belonged to the calendar.
The same season carried a quieter signal from inside the industry. In a July 2025 interview, Jensen Huang — CEO of NVIDIA, the company supplying the compute on which the rest of this revolution runs — was asked about an MIT study suggesting AI makes people cognitively duller. His response named the relevant division: he uses AI daily and finds his cognitive skills advancing; “most of his job as CEO,” he said, “is asking good questions”; the people who get duller are the ones who outsource judgement rather than apply critical thinking to the model’s output. Huang was making the meta-skill argument in mid-2025, from the figure with the strongest commercial reason to talk down AI’s risks rather than up them.
Two weeks later, on July 30, Anthropic CEO Dario Amodei — one of the principal architects of frontier model development — gave an extended interview on the Big Technology Podcast. What was notable was the dual register. Amodei dismissed the labels, calling “AGI” and “superintelligence” “totally meaningless” and “a marketing term,” while affirming in the same interview that he is “indeed one of the most bullish about AI capabilities improving very fast” and that Anthropic has a duty to warn about what is coming. Insiders were already disagreeing about what to call it; they were not disagreeing about whether the capability shift was real.
By early September the practitioner economy was already restructuring around the implications. Pieter Levels, the indie founder behind several well-known small businesses, published a thread on September 5 explaining how he had moved investments around in anticipation of “post-AGI.” His thesis was simple: in a world where cognition becomes free, the prices of most products approach zero, and what retains value is land, energy, infrastructure, raw materials, biotech, luxury, and authentic human connection. He had bought a home and was looking at land. Whether the investment thesis is correct is beside the point. The point is that a working practitioner with an unusually direct line to capability changes was already past the question of whether and onto the question of what to invest in afterwards.
Insider acknowledgement: winter 2025
By the end of the year, the framing had shifted from “this is happening” to “this has happened.” Sam Altman, in a December 2025 interview, said it as plainly as a CEO can: “AGI kind of went whooshing by. We’re in this fuzzy period where some people think we have it and some don’t.” He moved the goalpost in the same breath, redefining “superintelligence” as the point at which AI can be better than any human at being a CEO or president. Whether or not the redefinition is accepted, the relocation of the goalpost is itself a signal.
Days later, Andrej Karpathy — one of the architects of modern deep learning — posted that he had “never felt this much behind as a programmer.” The framing was striking for two reasons: the proximity of the speaker to the technology, and the metaphor. “Some powerful alien tool was handed around,” he wrote, “except it comes with no manual and everyone has to figure out how to hold it and operate it, while the resulting magnitude 9 earthquake is rocking the profession.” When the architect feels behind, the field has crossed a non-linear threshold.
Lay-observer dissonance: early 2026
The signal had reached general perception by early 2026. An anonymous Twitter user writing in February articulated the lay version: “I feel like a schizo every day. I see people all around me operating as though everything is normal.” He overshot on timing — predicting that 2026 would be remembered “a thousand years from now” — but his diagnosis of the dissonance was accurate.
Sahaj Garg, a Stanford-trained engineer, gave the same observation reflective shape in March 2026. He had spent his life believing intelligence was the ground he stood on; he wrote that intelligence was now “a commodity on tap,” and the skills that remained to him were “taste, direction, and synthesis.” His report is post hoc — an account of work already completed in eight-hour conversations with a frontier model, through forty-one revisions of an essay.
What the shape implies
Capability events first; insider acknowledgement next; lay-observer dissonance last. By the time the third layer arrives, the underlying shift is already months old. The convergence — across vantage points (researcher, founder, anonymous user), across evidence types (capability event, commentary, behavioural shift), and across the credibility spectrum — is what gives the bellwether its weight. Convergence is the lagging signal of a leading reality.
The mechanism: a price collapse in cognition
The temptation in interpreting this moment is to reach for the largest possible frame — “second cognitive revolution,” “singularity-adjacent,” “the end of human exceptionalism.” The framing that travels best, however, is economic.
Yuval Noah Harari’s account of the first cognitive revolution describes an endogenous transformation in the species — language, narrative, the capacity for shared fiction. What is happening with large language models is a different shape entirely: the externalisation of cognitive labour into a non-human collaborator. It belongs on the long arc of cognitive externalisations — writing, the printing press, the encyclopedia, the search engine — rather than on the same shelf as the invention of language itself.
The more analytically useful frame is that this is a revolution in the cost structure of cognition. The first cognitive revolution made cognition possible. What is happening now is making it cheap. Reasoning, drafting, synthesis, contract review, basic medical triage, financial modelling — tasks that were expensive because they required a trained human brain — are now approaching free.
This frame predicts the dissonance. Recipe-users — people using frontier AI to look up dinner ideas, summarise short emails, or ask whether their houseplant is dying — are not failing to understand a metaphysical event; they are simply behaving as though cognition still costs what it cost a year ago, because their behaviour has not yet been repriced. The practitioners on the capability frontier, restructuring research workflows and investment portfolios with the model in the loop, have repriced. The gap between them is the gap between two cost models, not two intellects.
Where leverage moves: from cognitive labour to cognitive direction
When something becomes cheap, the things adjacent to it become valuable. Industrial machinery made physical labour cheap, and what got expensive was what to make with it. The same logic operates on cognition. As reasoning and drafting commoditise, the questions that retain value are normative rather than productive: which question is worth asking; whose answer to trust; what the model missed.
The discourse converging in 2026 names a recognisable cluster of capabilities that retain leverage in this regime. Sahaj Garg names “taste, direction, and synthesis” explicitly. Karpathy’s harness mindset names tool design, eval discipline, and context engineering. Folding these together, six meta-skills appear repeatedly across practitioner accounts:
- Strategic thinking — choosing which questions are worth asking; the capacity to interrogate the brief itself
- Critical thinking — judging answers against reality; calibrated detection of confident-sounding wrongness
- Emotional intelligence — the human-relational substrate, which becomes proportionally more valuable as cognitive labour around it commoditises
- Taste — the aesthetic backbone of judgement; knowing what counts as good in a domain and being able to defend it
- Synthesis — holding multiple partial views (multiple agents, multiple sources, lived experience) in mind and shaping them into a position
- Calibration — knowing what one knows, what the model knows, and where each is reliable; the metacognitive skill of updating without over- or under-correcting
These are not new skills. They are the oldest skills. What is new is that they are no longer the distinguishing edge between an exceptional professional and an average one. They are the only distinguishing edge that remains in domains where the cognitive output itself has commoditised.
Practical applications
For institutions designing curricula, the implication is not that traditional content disappears — meta-skills require substrate, and stripping doctrine produces credulous evaluators rather than empowered ones. The accurate move is reweighting. The centre of gravity shifts from transmission to evaluation. What used to be the whole job of a course — here is the body of knowledge, learn it — becomes the floor on which a different job is built: here is the body of knowledge, you now have a tireless assistant; learn to direct, evaluate, and integrate.
Disciplines whose pedagogies were already meta-skill-heavy — those that taught reasoning over recall because practitioners always had access to reference — adapt more gracefully. Disciplines whose graduation rituals depended on memorising lookups face a sharper rebuild, because the lookups are now free.
For individual practitioners, the actionable form of the argument is that meta-skills resist syllabi. They are built by practice with stakes: making decisions with consequences, being wrong in front of people whose respect matters, writing for an audience that will push back, keeping track of one’s predictions and watching them succeed and fail.
Limitations and open questions
The price-collapse framing is descriptive, not predictive. It explains the dissonance and locates the durable human contribution but says little about timing — whether the collapse continues at its current rate, plateaus, or compounds. The Karpathy cognitive-cores thesis implies further capability gains through smaller, better-trained models with external memory; the scaling-limits thesis implies a plateau. Either could be right; both have at least a year of evidence behind them.
The “meta-skills are the only edge that remains” claim is also strong. In domains where cognitive output is highly verifiable (formal proofs, code that must compile and pass tests), models continue to compress the human contribution. In domains where output is harder to verify (judgement-laden professional work, creative production, advisory roles) the meta-skill emphasis is on firmer ground. The claim’s strength varies by domain.
Finally, the bellwether of late 2025 is one observation. Whether the period around it marks the inflection that subsequent observers will recognise — or whether some later moment will retrospectively look like the bend point — is unsettled. What the convergence justifies is attention, not certainty.
Sources
Capability events (summer 2025):
- @ns123abc — Jensen Huang on AI as cognitive enhancer; prompting and critical thinking as the meta-skills that distinguish use from misuse
- @VraserX — GPT-5-Pro produces a novel proof in convex optimisation, verified by Sebastien Bubeck; first widely-circulated example of a frontier model independently advancing the research frontier
- Dario Amodei on Big Technology Podcast (Jul 30 2025) — Anthropic CEO interview: dismisses AGI/superintelligence as marketing terms while affirming aggressive capability gain; the dual register that defined insider commentary in summer 2025
- @levelsio — Pieter Levels publicly restructuring portfolio for “post-AGI” economics; practitioner behaviour shift
Insider acknowledgement (winter 2025):
- @Ric_RTP — Sam Altman December interview: “AGI kind of went whooshing by”; goalpost relocated to “superintelligence”
- @karpathy — “Magnitude-9 earthquake” post; alien-tool framing of the capability shift
Lay-observer dissonance and reflective accounts (early 2026):
- @sillydarket — anonymous “schizo” dissonance post; lay-observer account of the convergence
- Sahaj Garg — intelligence as commodity on tap; identity-dissolution narrative; taste/direction/synthesis triad
Other:
- Yuval Noah Harari, Sapiens (2015) — referent for the first cognitive revolution as language and shared fiction