Intelligence Explosion and Self-Evolving AI — The Actual Threshold
AGI is not the inflection point — self-evolving AI that improves its own architecture is
Mo Gawdat's Intelligence Explosion framework establishes a three-level hierarchy of AI milestones and argues that the critical inflection is not AGI (Level 2) but self-evolving AI (Level 3) — AI that develops the next generation of AI. Level 1 (current AI) is already smarter than 97% of humans in most cognitive domains. Level 2 (AGI) matches or exceeds the best human at all cognitive tasks — Mo revised his prediction from 2027 to 2026 at latest. Level 3 is the actual event horizon: once AI is autonomously improving its own architecture, human oversight becomes advisory, then nominal, then vestigial, and capability improvement compounds without a defined ceiling.
The framework's central insight is that the game-theoretic structure of the AI race makes Level 3 inevitable regardless of any individual actor's preferences. If one player uses AI to develop the next generation of AI, every other player copies. Neither China vs. America nor OpenAI vs. Google can unilaterally stop — the race structure guarantees adoption of self-evolving AI by all frontier players.
Mo's concrete evidence that Level 3 is already beginning: Google's AlphaEvolve system — four AI agents working in parallel (identify performance issues, formulate problem statement, develop solutions, assess solutions) — improved Google's AI infrastructure by 6-10%. 'In Google terms, that is billions and billions of dollars.' This is not a research prototype; it is a deployed self-improvement system already generating measurable value.
- Self-evolving AI — not AGI — is the actual inflection because it removes the human bottleneck on AI development itself
- The game-theoretic structure of the AI race makes self-evolving AI adoption inevitable regardless of individual actor preferences
- AlphaEvolve-type systems (multi-agent self-improvement) are already deployed and generating measurable value at Google
- Once self-evolving AI is running, capability improvement is compounding with no defined ceiling
- Sam Altman's shift from slow-takeoff preference to fast-takeoff belief is a leading indicator that frontier labs have lost the ability to govern pace
- Calibrate your AGI timeline against Mo's revised predictionMo's original 2023 prediction was AGI by 2027; he revised to 2026 at latest in 2025. The standard definition is fuzzy — his pragmatic threshold is 'when the top engineer becomes an AI.' Use this as a forcing function: if AGI arrives in 2026, what does your portfolio, career, or organisation look like in early 2027?Pro tipAGI declarations from major labs will be contested and politicised. Mo's pragmatic definition ('when the top engineer becomes an AI') is more useful than academic consensus definitions for real-world planning.
- Identify Level 3 (self-evolving AI) signals already in the marketAlphaEvolve at Google is already Level 3: four AI agents working in parallel to identify, formulate, develop, and assess improvements to Google's own AI infrastructure, generating 6-10% improvement. Watch for public announcements of similar systems from OpenAI, Anthropic, and Chinese labs as the signal that Level 3 is no longer a single player's advantage.Pro tipThe 6-10% infrastructure improvement figure translates to 'billions and billions of dollars' in Google terms — Level 3 creates immediate commercial incentive for every frontier lab to deploy it regardless of safety preference.WarningLevel 3 adoption by the first player creates an immediate prisoner's dilemma forcing all other players to deploy. There is no coordination mechanism strong enough to prevent this given current geopolitical conditions.
- Apply the DeepSeek wild card to your concentration thesisDeepSeek R3 at 1/30th the cost of leading frontier models, open source, and edge-deployable, partially nullifies the moat of compute-heavy centralised providers. Evaluate whether your AI exposure is concentrated in centralised platform owners (ChatGPT, Google) or distributed inference infrastructure — the two theses have different risk profiles if DeepSeek-type disruption accelerates.WarningA multi-AGI scenario — where multiple AGI-level systems emerge within months of each other rather than one dominant player — is Mo's 'very interesting scenario.' It disrupts the platform concentration thesis (FACE RIPS 'I' vector) and benefits distributed compute infrastructure plays.
- Map the human oversight degradation curveOnce self-evolving AI is running, the pattern Mo describes is: AI requests more resources, the team complies. Human oversight shifts from active governance to rubber-stamping to vestigial supervision. Map at what point in this curve governance structures in your sector or jurisdiction become structurally unable to intervene — that is the actual policy window for AI governance.Pro tipMo's framing: the human oversight that exists in Stage 1 augmented intelligence becomes advisory, then nominal, then vestigial. The policy window is Stage 1 — before the compounding removes the governance lever.
Four AI agents work in parallel: one identifying performance issues in Google's AI infrastructure, one formulating the problem statement, one developing solutions, one assessing them. The system improved Google's AI infrastructure by 6-10% without human engineers driving the improvement cycle.
In 2023, OpenAI published a note stating 'a slower takeoff is easier to make safe' — Altman preferred gradual capability development. By 2025, Altman believes a fast takeoff ('on the order of a small number of years rather than a decade') is more plausible.
DeepSeek R3 achieves frontier-adjacent capability at 1/30th the compute cost of OpenAI-equivalent models, is fully open-source, and is deployable at the edge without centralised API access. This partially nullifies the infrastructure moat that Mo's 'digital soil' thesis identifies as the primary value capture mechanism.
This framework synthesises existing AI safety research (intelligence explosion, first described by I.J. Good in 1965; expanded by Nick Bostrom and others) with Mo's insider knowledge of Google's actual deployed systems. Mo is not originating the intelligence explosion concept — he is providing practitioner evidence that it has already begun, drawing on his proximity to Google X and DeepMind.
The Sam Altman position shift is Mo's external validation: Altman moved from preferring a slow takeoff (2023 OpenAI note: 'a slower takeoff is easier to make safe') to believing a fast takeoff is now more plausible — 'on the order of a small number of years rather than a decade.' Mo reads this as capitulation to observed reality: 'It is so fast that none of them has the choice to slow down.'