The Path to Superintelligence

Google DeepMind recently published an interesting paper ‘From AGI to ASI‘ (arXiv:2606.12683). AI2Work has provided a very useful summary here.

AGI is Artificial General Intelligence, while ASI is artificial superintelligence. The paper takes AGI to be around the intelligence level of human beings, while ASI is significantly better across all domains – it’s not enough to be able to beat humans at chess or outdo them in other specialised areas. As its basis, the paper quotes the Legg-Hutter scale. This scale scores intelligence objectively according to ability to achieve goals in different environments. There is apparently a scary endpoint to the scale in the shape of UAI, Universal Artificial Intelligence, which would be at the theoretical limit of intelligence. That seems open to some philosophical discussion (not today, though).

The real point of the paper is to set out in general terms possible paths from AGI to ASI and the likely barriers. It’s a pretty sensible discussion, free of the euphoria that sometimes affects this kind of discussion. At the risk of summarising the summary, the four paths are as follows.

  • Scaling up compute, models and data
  • Algorithmic paradigm shifts
  • Recursive self-improvement
  • Multi-agent collectives

I’m not sure why multiple agents could achieve things that a single one couldn’t – a single one could simulate a collective, couldn’t it? But it’s a reasonable list.

Six barriers are identified (always more problems than paths).

  • The ‘Data Wall’. Lack of new human text and other material to feed ever larger and hungrier models.
  • Economic and resource demand grows too quickly.
  • Insufficiency of the current ‘neural’ paradigm.
  • ‘Research gets harder’ (love that one).
  • The ‘Abstraction Barrier’ – can concepts generated by humans be surpassed by AIs?
  • The Evaluation Problem – how do you benchmark superintelligence?
  • A deliberate slow down, from legislative or prudential constraints.

That’s all quite level-headed and I will only offer one slightly jaded observation. There is a fifth path to ASI which is the most likely of all. This is: just re-badge whatever we’re doing now as ASI. Surely in a year or two when someone wants to give things a boost, they’ll start to make that claim. Once, AI meant fully conscious artificial minds, then we started using it for relatively ordinary bits of programming, and deferred the fully aware human mind thing to AGI. Now AGI is just human levels of problem-solving and processing, no real awareness required. It’s a bit like walking up the down escalator. But not quite, because now consciousness has dropped off the map altogether, or at least receded to an even more remote form of artificial intelligence than ASI. Maybe people have ultimately decided consciousness isn’t really that essential – or maybe it’s just that research got harder.