Google DeepMind Releases AGI Roadmap: 100 Million Human-Level AIs Would Constitute "ASI"
Alina Collins
DeepMind released a 57-page report arguing that even if a single AI never surpasses human-level cognition, 100 million instances working together would cross the threshold into superintelligence — the bottleneck is scale, not smarts.
What is this report actually saying?
The report, *From AGI to ASI*, was led by DeepMind co-founder Shane Legg and AIXI inventor Marcus Hutter, completed by a 14-person team.
It defines three tiers of intelligence: AGI (matching median human performance on most cognitive tasks), ASI (consistently outperforming tens of thousands of top experts collaborating for a decade), and UAI/AIXI (the theoretical absolute ceiling).
This means → DeepMind has drawn a quantifiable red line for superintelligence: not a vague "smarter than humans," but stably outperforming an entire elite expert army.
How do 100 million "ordinary" AIs become superintelligent?
The report runs a thought experiment: at birth, AGI runs only 1,000 instances worldwide; at 10× annual growth, that reaches 100 million in five years.
Those 100 million share one brain and think a hundred times faster — their collective intelligence alone crosses the AGI-to-ASI line.
In plain terms = one person can't solve the problem, but a hundred million equally smart copies thinking simultaneously, sharing answers at a hundred times human speed — sheer headcount becomes a form of transcendence.
Four paths to ASI — which is most realistic?
Path 1: Brute-force scaling — keep growing compute, data, and model size. The report's most confident claim: within years, AGI shifts from lab luxury to infrastructure.
Path 2: Paradigm leap — if the current "pretrain + fine-tune + inference" pipeline hits a ceiling, entirely new architectures may emerge — spiking neural networks, neuromorphic hardware, or infinite-working-memory designs like Mamba.
Path 3: Collective emergence — millions of AGI specialists linked by high-bandwidth communication form a digital ecosystem; coordination effects produce group intelligence exceeding the sum of all individuals.
Path 4: Recursive self-improvement — AI rewrites its own code and hardware to accelerate R&D. AlphaEvolve and FunSearch are early examples.
Six "walls of woe" — what could stall everything?
The first five walls: data (high-quality internet text exhausted by decade's end), resources (astronomical spending on compute, power, and chips), paradigm (the Transformer pipeline may plateau), difficulty (low-hanging fruit gone, research complexity rises exponentially), and society (mass white-collar displacement triggers regulatory brakes).
The sixth — and most original — is the "abstraction barrier": feed an AI every human text written before Newton — can it independently derive general relativity or quantum mechanics? DeepMind says almost certainly not — the model lacks foundational primitives like calculus or gravity.
This means → a single model may be permanently locked at the ceiling of human cognition. But the report immediately adds: this wall stops one genius, not 100 million ordinary instances — collective intelligence can route around any individual's abstraction bottleneck.
What outcome is DeepMind betting on?
The report's closing judgment is notably restrained: for AI progress to stall at the human line, all six walls must simultaneously become dead ends — an unlikely coincidence.
DeepMind bets on two scenarios: either progress stalls before AGI, or the path from AGI to weak ASI is remarkably smooth. A middle state — stuck right at AGI — is the least probable outcome.
This reflects DeepMind's core wager: the bottleneck is not "can we build superintelligence" but which of the six walls becomes the first true hard constraint.
Content is for reference only, not financial advice.