Tech • IA • Crypto
Google Just Revealed What Comes After AGI—And It’s Shocking
9.1/10TL;DR
A new Google DeepMind paper argues that artificial general intelligence (AGI) may be only a starting point, outlining multiple pathways—and major constraints—toward artificial superintelligence (ASI).
KEY POINTS
AGI Redefined as Baseline
The paper defines AGI as systems performing at roughly median human level across most cognitive tasks, including reasoning, planning, communication, and adaptation. This framing rejects the idea of AGI as elite intelligence, instead positioning it as a broadly capable but average-level benchmark.
ASI Sets an Extreme Bar
ASI is described as intelligence capable of outperforming tens of thousands of top experts working in coordination for a decade on complex problems. This shifts the comparison from individuals to entire institutions or research fields, emphasizing scale and collective capability.
AIXI as Theoretical Ceiling
The work references AIXI, a mathematically defined but uncomputable model of optimal intelligence. It serves as an upper bound, suggesting that while intelligence can improve, there are theoretical limits analogous to physical constants.
Pathway 1: Scaling and Collective Intelligence
Continued growth in compute, data, and efficiency could multiply AGI instances rapidly. A scenario outlined shows expansion from 1,000 to 100 million systems in five years under sustained growth. These systems could share knowledge instantly, effectively forming a coordinated “digital civilization” whose collective output may qualify as ASI.
Pathway 2: Algorithmic Breakthroughs
Progress may depend on paradigm shifts beyond current transformer-based models. Key gaps include long-term planning, persistent memory, and robust world modeling. New architectures or hardware, such as neuromorphic computing, could rapidly invalidate forecasts based solely on scaling.
Pathway 3: Recursive Self-Improvement
AI could accelerate its own development by improving algorithms, designing chips, generating data, and optimizing research workflows. This feedback loop may resemble an “intelligence explosion,” though physical constraints—like manufacturing timelines and experimental limits—could slow it.
Pathway 4: Multi-Agent Systems
Rather than a single superintelligence, ASI may emerge from large-scale AI collectives. These systems could coordinate at high speed, run parallel experiments, and dynamically reorganize. Compared to human institutions, such collectives would avoid bottlenecks like slow communication and bureaucratic friction.
Key Bottlenecks Identified
The paper outlines six major “frictions”: the data wall (limited high-quality training data), resource constraints (energy, chips, infrastructure), potential limits of current neural network paradigms, increasing difficulty of research progress, reliance on human-derived abstractions, and regulatory or societal slowdowns.
Limits of Superintelligence
Even ASI would remain bound by physics, computation costs, and uncertainty. Constraints such as energy use, time delays, and complexity theory mean that superintelligence would not equate to omnipotence or instant solutions to all problems.
Shift in Strategic Focus
The central implication is a reframing: AGI is not an endpoint but a transition. Once achieved, its ability to scale, replicate, and integrate into systems could transform intelligence into an industrial process, accelerating innovation beyond human limitations.
CONCLUSION
The analysis highlights profound uncertainty about how—and how fast—ASI could emerge, while making clear that the transition from AGI may redefine the pace and structure of technological progress.