From AGI to ASI

Artificial superintelligence is AI that is superhuman across virtually all tasks, exceeding not one expert but assembled teams of the best. The report does not predict when it arrives. Instead it maps four parallel pathways that could carry capability from AGI to ASI, grounds the target in a formal theory of intelligence, and catalogues six bottlenecks that might be temporary friction or hard walls. The throughline is uncertainty: forecast with many models and keep measuring.

Most forecasting stops at AGI because human level is a familiar anchor. The report's wager is that the AGI-to-ASI leg could be short and is badly under-studied - compute, capital, and the number of (increasingly automated) researchers are all scaling at once, and they compound. The point is not to call a date but to name the levers and watch them.

Intelligence here is a spectrum, not a switch. Walk it from narrow systems through human-parity AGI to superhuman ASI, and finally to the theoretical ceiling. Each step changes what generalises and how you would measure it.

The decisive gap is between AGI and ASI. AGI is pinned to a single competent human; ASI is defined relative to the best human collectives working in parallel. So the move from one to the other is not a tweak but a change of regime - and the same digital substrate that reaches AGI can keep going well past it, which is why the two are worth separating.

Why might a digital intelligence overshoot human ability so far? Six structural advantages. But the same report is blunt about six limits no amount of intelligence escapes. Flip the panel to see both halves of the ledger.

The advantages all trace to one fact: a mind made of code and weights, not neurons. Code can be copied perfectly, run faster by adding hardware, and shared across instances - none of which biology allows. That is the lever that lets digital intelligence keep climbing past human level instead of plateauing there. The walls are the reminder that "past human" still means "finite".

"Effective compute" is the product of three yearly multipliers: cheaper hardware, growing investment in the largest training run, and better algorithms per FLOP. Drag each and watch them compound. The report puts the combined figure near 10x every year.

They multiply rather than add because each pulls a different lever: a cheaper FLOP, more FLOPs bought, and more capability squeezed from each FLOP. Stack three modest yearly gains and the product is an order of magnitude. Run that for a few years and the total is staggering - which is the whole case for scaling reaching ASI on its own, and also why a single stalled multiplier (the data wall) matters so much.

The formal ceiling is AIXI: a single agent that defines optimal behaviour in any computable world by solving three problems at once. It is incomputable, but it tells you what intelligence is converging toward. Pick a problem to see how AIXI dissolves it.

Why care about a model you can never run? Because it turns "intelligence" from a vibe into a definition you can approximate. AIXI says an optimal agent is just compression plus reward-seeking over every possible world. That gives the Legg-Hutter score its meaning - and, as the next note argues, hints that today's pretraining is already crawling up the same curve from below.

These are not exclusive - they likely run together and feed each other. Select a path to see its thesis, the mechanism behind it, and the friction it has to clear.

Read them as a rough ladder of confidence. Scaling is the best-understood and most measurable; paradigm shifts are real but unschedulable; recursive self-improvement is the highest-leverage and least-understood; multi-agent coordination sidesteps single-model limits by making intelligence a property of the group. The interesting bets are where they combine - scaled models doing the research that finds the next paradigm, run as a coordinated collective.

Path three deserves its own look. AI can improve itself along the same four axes humans evolved on - but on radically compressed timescales. Pick an axis to compare the human analogue with the AI version.

The reason RSI is the wildcard: it is the one path with a feedback loop into itself. A better model designs better data, chips, and successors, which build a still-better model. Where the other paths add capability, RSI multiplies the rate at which capability is added - so small differences in loop strength produce wildly different futures.

ASI need not be one mind. It can emerge from many AGI agents coordinating - the way human institutions exceed any individual. The open design question is how they organise. Toggle the three forms.

This path is attractive because it routes around single-model limits. One agent has a fixed context window and training; a collective splits the work, specialises, and pools results, so capability scales with population and bandwidth instead of architecture. The open question is how much real synergy similar agents achieve - and how you steer a superintelligent group without becoming its bottleneck.

Every pathway can stall. The report names six bottlenecks and, crucially, does not resolve whether each is temporary friction or a fundamental wall. Click one to see its counter and where the report leans.

That friction-versus-fundamental call is the whole game. Friction slows the climb but yields to more compute, money, or a workaround - annoying, not fatal. A fundamental limit caps it until something genuinely new appears. The honest position in the report is that we mostly do not yet know which is which, and that turning these unknowns into measurements is the real research agenda.

The report ends not with a date but with a research agenda. Four habits for reasoning about a transition no single model can forecast.

The closing posture is deliberately humble: a single transformative step-change is the less likely story; a series of cascading breakthroughs is more plausible. So the work is not to pick the one true scenario but to instrument the levers - effective compute, automated research, recursive-improvement dynamics, resource limits - and update continuously as the numbers come in.