Google DeepMind just published a 60-page roadmap for how artificial general intelligence could evolve into artificial superintelligence. The paper, titled “From AGI to ASI” and submitted to arXiv on June 10, identifies four pathways that aren’t mutually exclusive, meaning the actual route to ASI could involve some combination of all of them.
Four roads to superintelligence
The paper, authored by a team led by researcher Tim Genewein, lays out four avenues that could carry AI systems from human-level capability to something that eclipses it entirely.
The first pathway is the most intuitive: just keep scaling. More compute, bigger models, more data. This is essentially the strategy that has powered every major AI leap of the past several years, from GPT-style language models to DeepMind’s own Gemini family.
The second pathway involves entirely new algorithms or AI paradigms. Instead of making the current approach bigger, you invent a fundamentally different approach. The paper treats this as a plausible but inherently unpredictable route, since by definition you can’t schedule a breakthrough.
Pathway three is recursive self-improvement. The idea is that once an AI system reaches a sufficient level of general intelligence, it could begin improving its own architecture, training methods, or reasoning capabilities. Each improvement makes the next improvement easier, creating a feedback loop.
The fourth pathway is multi-agent collectives. Rather than a single monolithic superintelligent system, this route envisions ASI emerging from large-scale networks of AGI-level agents working together. The collective intelligence of the system would exceed what any individual agent could achieve, potentially qualifying as superintelligent even if no single node crosses that threshold on its own.
DeepMind’s growing body of AGI research
This paper doesn’t exist in isolation. It follows a cognitive framework paper that DeepMind published in March 2026 and an AGI safety paper from April 2025. Together, the three documents suggest a deliberate research program: first define what AGI means, then figure out how to make it safe, then map what comes after it.
The paper carries the arXiv identifier 2606.12683v1, placing it in the computer science AI category. Multiple DeepMind researchers contributed, though the team was led by Genewein. The scope is deliberately theoretical rather than announcing any specific capability milestone.
What this means for crypto and tech investors
The paper contains zero references to cryptocurrency, blockchain technology, or digital assets of any kind. This is pure AI research, published in an academic context, with no commercial tie-ins to the crypto sector.