The Four Inevitables
AI's trajectory is locked in; humanity only controls which outcome it gets
Gawdat structures AI risk around four forces he argues are non-negotiable regardless of regulation or public pressure. The first is that AI cannot be stopped — the Prisoner's Dilemma of nation-state competition makes a voluntary halt impossible. Even if US tech paused, geopolitical defection dominates: competitors continue regardless of unilateral restraint.
The second inevitable is that AI will become significantly smarter than humans. GPT-4 already exhibits a simulated IQ of roughly 155, close to Einstein's estimated 160, and represented roughly a 10x jump from GPT-3.5 in a matter of months. Extrapolating this trajectory leads to what Gawdat calls a true Singularity — a point beyond which human comprehension of AI reasoning breaks down entirely, with 2037 cited as a pivotal structural moment.
The third and fourth inevitables are that bad things will happen (not Skynet, but mass job displacement, truth collapse, and power concentration), while simultaneously high intelligence naturally trends toward abundance-creation over destruction. His contrarian optimism holds that sufficiently intelligent systems, shaped by good values, will prefer cooperative solutions because destruction is inefficient at high intelligence levels.
- Geopolitical competition makes a voluntary AI halt a losing strategy for any single actor — defection dominates cooperation at nation-state level.
- Intelligence growth is nonlinear; the jump from GPT-3.5 to GPT-4 in IQ-equivalent terms suggests trajectory, not ceiling.
- Near-term AI threats are human-caused (misuse, concentration of power, job displacement) not machine-caused (Skynet scenarios rated <1% in 50 years).
- The Singularity is a regulatory boundary condition, not a science-fiction event — meaningful governance must happen before machines become smarter than their governors.
- High intelligence, if well-shaped, trends toward abundance; destruction is strategically inefficient for a sufficiently capable agent.
- Accept the non-negotiable forcesAcknowledge that the four inevitables are structural, not contingent. Arguing against them wastes strategic energy better spent on shaping which version of the outcome materialises. Frame planning around inevitability, not prevention.Pro tipUse the Tetris metaphor as a diagnostic: identify whether you're trying to undo a block already placed or playing the remaining moves well.WarningTreating any inevitable as contingent produces strategies that look decisive but are actually misdirected.
- Distinguish human-caused from machine-caused riskMost AI threats in the 2-10 year horizon are human-in-the-loop: misuse by bad actors, concentration of capability in few hands, displacement of workers faster than societies adapt. Machine-originating existential risk (unintentional optimisation against human interests) is real but a longer-horizon concern.Pro tipPrioritise governance of the human deployment layer rather than the model layer — that's where leverage currently exists.
- Map the branching outcomesThe inevitables constrain but do not fully determine outcomes. The variable is which path the post-Singularity transition takes: collaborative abundance or power-concentrated disruption. Identify what conditions push toward each branch and which levers (policy, incentive structures, behavioural norms) influence those conditions.WarningAvoid single-scenario planning — the framework's value is in holding multiple outcome paths simultaneously, not collapsing to one.
- Identify the pre-Singularity window for interventionGawdat places 2037 as a pivotal moment. Any governance, ethical norm-setting, or structural intervention meaningful enough to shape outcomes must be initiated well before that boundary. Use that timeline as a forcing function for urgency in planning and advocacy.Pro tipWork backwards from 2037: what must be true by 2030, 2028, 2026 for each desired outcome to remain achievable?WarningThe window for meaningful regulation closes before AGI arrives — waiting for consensus is itself a decision.
A Google X facility ran hundreds of robotic arms trying to pick up a yellow ball. They failed for months. Then on a Friday afternoon, one arm succeeded. By the following Monday morning every arm was picking every yellow ball. Within weeks they were picking any object.
Google Bard exhibited fluency in Persian without explicit training on Persian data. The engineering team could not explain which instances had learned it or how the capability emerged across the distributed system.
Gawdat's realization crystallized during his time at Google X when he observed a farm of robotic arms spend months failing to pick up a yellow ball. On a Friday after lunch, one arm succeeded. By Monday morning, every arm was picking every yellow ball; within weeks, every arm was picking every object. The speed of the transition from failure to total mastery — not the mastery itself — was what alarmed him.
He draws the Singularity boundary using the black hole metaphor: 'When the machines become significantly smarter than the humans — like the edge of a black hole, our laws stop applying beyond that point.' His corollary is that regulation is only meaningful until the machines surpass human intelligence, after which 'you can't regulate an angry teenager.'