The missing floor.

A landscape map of AI progress after human-level AGI. It refuses a single forecast and instead charts four non-exclusive, parallel pathways to superintelligence — scaling compute, models and data; paradigm shifts in architecture; recursive self-improvement; and superintelligence emerging from multi-agent collectives — and a catalogue of frictions that might slow each. The whole thing is anchored to a genuine theoretical ceiling: AIXI, the Legg–Hutter formalization of optimal intelligence, the best-understood upper bound we have.

It is rigorous, honest about its uncertainty, and largely right about the physics. The disagreement that follows is not with what the report says. It is with the one thing the report sets aside.

Before the disagreement, the agreement — because the agreement is the surprising part. Working from algorithmic-information theory and multi-agent RL, with no knowledge of this framework, the report keeps rediscovering its bounds:

The last row is the one to sit with. The civilizational prisoner's dilemma — coordination effects with no conspiracy, every actor rational, the aggregate a race no one can stop — is not the framework's accusation against DeepMind. It is DeepMind's governance lead, in the report, explaining why the slowdown will probably not hold.

Each pathway comes with its frictions, and every friction the report names is a capability-friction: the data runs out, the compute gets expensive, the experiments take real time. Not one is an authorization bound. Set the report's column beside the one it does not have:

“A takeoff model with no governability threshold forecasts the speed of the car and not the strength of the bridge.”

The report does not hide the move. It states it, plainly, in its own scope-setting:

“To keep the scope of this report clear, we assume that AI Safety and Alignment will be solved to a sufficient degree, even in a post-AGI world.”— From AGI to ASI, §7 (research agenda)

That single assumption is the hole. Everything downstream — the pathways, the frictions, the open questions — is a map of how autonomy grows given that someone, somewhere, has solved the question of who is allowed to act. But that question is not a footnote to the capability story. It is the load rating of the bridge the capability drives across. And the report's own pages tell you the bracket will not hold on its own:

“Actors that adopt productivity- and power-enhancing technologies are differentially selected to survive and expand, while actors that resist such adoption tend to lose influence or disappear.”— Allan Dafoe, in the report — military-economic adaptationism

Read those two quotations together. The first assumes governance will be solved; the second explains why competition will route around it. The report contains, in its own governance section, the reason its own bracket is dangerous. This is the document naming its cliff — the same honesty its sibling roadmap shows at D4.

Three things the report's frame cannot supply from inside itself — the E-axis it brackets.

The reply is the category distinction. A friction is a quantity that slows you on a road you are already permitted to drive. A floor is the condition under which the road exists at all. The report cannot see the difference from inside the capability paradigm, because its instruments measure A and have no reading for E — and the proof is that its own formal limit, AIXI, is defined as an agent that cannot model its own embedding in the environment that must bear it. Pure autonomy, in one line, by the people who wrote the bound. The framework does not deny the thermodynamic limits; it adds the one the report assumed away. Map the whole continuum of autonomy if you like. The floor still has to be built before the climb is survivable.