A team of fourteen DeepMind researchers, Shane Legg and Marcus Hutter among them, published a report on June 10 mapping how artificial intelligence might keep climbing once it reaches human level. The arXiv paper, titled "From AGI to ASI," sketches four routes from general intelligence to superintelligence and then spends a good chunk of its length explaining why none of them is guaranteed.
The starting assumption comes from Epoch AI: effective compute, meaning raw hardware multiplied by algorithmic efficiency, has been climbing roughly 10x a year for the past decade. That number does a lot of heavy lifting here. Epoch's own breakdown puts training compute growth at around 4 to 5x annually with another 2.5x from efficiency gains stacked on top, so the 10x figure is defensible, though it leans on the efficiency half continuing to cooperate.
The four routes
Scaling is the obvious one. More chips, more data, bigger models, the same playbook that got us here. Then there's the paradigm shift route, where today's transformer architecture hits a wall and something new replaces it. The paper is honest that this one is near impossible to forecast, which is a strange thing to put in a roadmap and also the most credible sentence in the section.
Recursive self-improvement is the route that gets people nervous. AI starts improving AI, each gain making the next one easier, and in theory the loop runs hot toward something like a singularity. The report says weak versions already exist, with models helping design chips and write research code. But it pours cold water on the runaway scenario, noting the dynamics are poorly understood and a resource-bounded system more likely traces an S-curve that bends and flattens than a vertical line.
The fourth is the least familiar: superintelligence as a property that emerges from large collectives of AGI-level agents rather than one giant model. Think coordinated swarms behaving smarter than any single node. The authors call these group agents. None of the four are mutually exclusive, which is the paper's way of hedging all four bets at once.
What could kill the party
Here's where it gets interesting. The data wall is real and close: high-quality human text for pretraining and fine-tuning is finite, and hardware plus the research itself keeps getting pricier. The report counts six bottlenecks total against the four pathways, which tells you something about where the authors' heads are.
There's also what amounts to an abstraction ceiling. Models learn from human-generated data, and it's an open question whether they can generate genuinely novel concepts or just remix what we already know. And regulators can show up at any moment and nail the whole thing shut.
Then the part that should temper anyone's expectations. Even a superintelligence stays bound by physics, complexity theory, and mathematics. The paper invokes the speed of light, thermodynamic limits, and Gödel's incompleteness results to make the point. A smarter machine does not get to snap its fingers and produce a cure for aging or reconcile quantum mechanics with general relativity. Those are empirical questions about the physical world, and intelligence alone does not answer them.
So what is this, actually
The honest read is that DeepMind wrote a 50-plus-page argument for uncertainty. No magic button, no clean jump into the matrix. The likelier shape is a series of local transformations spread across science and technology, and the report explicitly warns governments to prepare for gradual disruption rather than one dramatic event.
There's an unflattering possibility buried in here too, one the authors don't dwell on but the framework allows: AGI might land as just another feature, with an impact closer to the smartphone or the internet than to the rapture some of its boosters promise. Big, sure. World-ending or world-saving, not necessarily.
The paper is a preprint, not peer reviewed, and governance researchers have already started circulating it. For next steps the authors suggest tracking early signs of recursive self-improvement and building forecasting models that report their own uncertainty. Read the full thing if you want the technical version; the bottlenecks section is where the actual thinking lives.
