Geoffrey Hinton said that the breakthrough the AlphaGo team had was getting it to play against itself and improve in that means, since it could then go beyond the human training data it had learned on. He said that an equivalent form of self-training for generalized information would let a superintelligence take off (this is from my memory, not an exact quote).

The TechCrunch article doesn't specify how/what kind of data a recursive general AI could use to achieve such a thing. If it is possible that's exciting. Seems like a real philosophical question to answer- How could a general AI self-train?

▲

tim333 2 days ago | parent | next [-]

Silver has a couple of recent papers that probably give an idea of what they are up to:

>...Here we show that it is possible for machines to discover a state-of-the-art RL rule that outperforms manually designed rules. This was achieved by meta-learning from the cumulative experiences of a population of agents across a large number of complex environments... https://www.nature.com/articles/s41586-025-09761-x

and

>A new generation of agents will acquire superhuman capabilities by learning predominantly from experience. This note explores the key characteristics that will define this upcoming era. https://storage.googleapis.com/deepmind-media/Era-of-Experie...

▲

quicklywilliam 2 days ago | parent | prev | next [-]

Same question. Given the people involved I am inclined to take it seriously, given the money and hype involved I am somewhat less inclined.

Their website doesn’t even have a hint of what the approach is.

▲

vrighter 9 hours ago | parent | prev | next [-]

but we already know it's not. self training leads to model collapse

▲

AntiUSAbah 2 days ago | parent | prev | next [-]

Yesterday i watched a video stating that evolutionary algoritm become more relevant in machine learning again.

But if you think about our brain, if you learn something new, you play with it and recall it and challange the new information. Perhaps we can build something similiar. A model adjusting itself until its perfect.

▲

colonelpopcorn 3 days ago | parent | prev | next [-]

You'd probably have to embody it.

	▲	dwd 2 days ago \| parent [-]
		Or simulate embodiment. Human learning relies on movement. We learn as we navigate the world and interact with objects as noted by Jeff Hawkins in "A Thousand Brains". It may be an AI will also require this ability, either via simulation or a robotic interface. Will an AI then also lose its train of thought when it walks through a doorway and context (reference frame) switches?

▲

physicsm4n 3 days ago | parent | prev | next [-]

[dead]

▲

andsoitis 2 days ago | parent | prev [-]

> If it is possible that's exciting.

Would it be exciting though? I mean it would certainly excite some things, but I don’t know that it would be something to rejoice.

	▲	bko 2 days ago \| parent [-]
		Why wouldn't it be exciting? It would learn through logic and reason as opposed to our faulty human artifacts. And it wouldn't be limited to what we currently know. A good test would be if it could rediscover mathematics or relativity.