The Final Frontier For Knowledge

[towforce]

I have been following the progress of AI for over 40 years: I did a six-month course on the subject in 1984 (at that time, recognising things from images would have been regarded as virtually impossible if it weren't for the example of the human eye - and translation was so poor that it took as long to translate the translation as it did to translate from the foreign language directly). Vincent Lejeune's thread about Deepmind's latest project (link - thank you Vin!) has alerted me to what is an absolutely key moment in the field: Google's latest project has changed the emphasis.

What they have done is to target the deep underlying patterns of chess, rather than the surface (tactical) features, which are basically what current NNs are finding.

They haven't done this yet, but going forward, having found these deep underlying patterns, they then want to extract them as algorithms. If they can do this well, these algorithms will be small, fast, and relatively accurate. Thus small fast algorithms will be able to do better than the current big slow algorithms.

I have discussed extensively in the past why I believe that chess (and other bodies of knowledge) can be encoded in much smaller algorithms than most people today think possible: I am willing to go over the reasons again if anyone has forgotten (or missed the first time I went over them).

My own ideas (which I have discussed with Matthew Lai from DeepMind, e.g. in this thread - link) would be more compact and accurate than even what Google have done - but Google have actually done it! I am absolutely NOT claiming any credit for this stunning breakthrough, but it's so similar to what I have discussed that I have a feeling that I might have been a small influence in this research direction - even if not at the conscious level!

Summary:

(1) NN training will target deep underlying patterns rather than the shallow surface patterns they're currently finding

(2) Algorithms will be extracted from these new NNs

(3) This will enable knowledge encoding to be smaller, faster, and more accurate

(4) Point 3 above is the final frontier for knowledge

[towforce]

A quick note: "knowledge" is not the whole of human intelligence or human behaviour - but it is an important part of it: if I were to sit opposite a master for a game of chess, I'd lose - and the main reason would be that the master's knowledge of chess would be better than mine.

[smatovic]

???

I see a LLM that finally plays decent chess and can not outperform dedicated purpose neural networks.

And I agree with other posters, that Lc0 pretty much encodes "deep-chess-patterns" in NN, "grandmaster level with ply 1 search".

--
Srdja

[smatovic]

...maybe to mention, these days people talk a lot about AGI and ASI, the topic of AI was divided from "strong AI" into the sub-topics of "weak AI", you can view computer chess as one weak AI sub-topic, we have different, dedicated AIs to solve specific problems, the next step would be to combine them all into one big, single AI system, the AGI, artificial general intelligence, then next step further, the ASI, artificial super intelligence.

--
Srdja

[towforce]

...maybe to mention, these days people talk a lot about AGI and ASI, the topic of AI was divided from "strong AI" into the sub-topics of "weak AI", you can view computer chess as one weak AI sub-topic, we have different, dedicated AIs to solve specific problems, the next step would be to combine them all into one big, single AI system, the AGI, artificial general intelligence, then next step further, the ASI, artificial super intelligence.

Definitely. Today's LLMs are remarkably useful - but they're weak in specialist areas. It's amazing that LLMs can play chess and teach you a new language - but they cannot do those things well. So - how is this going to be overcome?

(1) Just keeps making the NNs bigger and train them with ever larger amounts of knowledge

(2) Have the system first understand what knowledge would be most useful in the provided context, then ask for it using an API to another system

Surely option 2 is the better choice right now.

I wrote an article in Selective Search, a computer chess magazine, in 1997 suggesting something similar for chess: basically:

(1) a database of evaluation functions (EF)

(2) select the best match for the current position

(3) use that EF to score the position

If that had been done, we could have entered the knowledge era of chess a lot sooner.

Having said all that, it actually seems likely to me that the best LLMs today are more intelligent than the smartest humans overall: no human has the breadth of knowledge that these systems have, and no human can give good responses as quickly.

[smatovic]

I still see some room for the expert-systems approach in future, currently they work to combine LLMs with expert-systems, point is that neural networks like LLMs are what I call probabilistic and expert-systems are deterministic, in some distant future, when the hype flattens, we might see expert-systems rising, or they figure a way how to fix the hallucinations in LLMs, Idk.

--
Srdja

[towforce]

I still see some room for the expert-systems approach in future, currently they work to combine LLMs with expert-systems, point is that neural networks like LLMs are what I call probabilistic and expert-systems are deterministic, in some distant future, when the hype flattens, we might see expert-systems rising, or they figure a way how to fix the hallucinations in LLMs, Idk.

My idea for selecting the evaluation function that best matches the position has it's root in case-based reasoning (CBR), not expert systems. Done well, CBR makes a really good AI helper, with a lot less of every kind of resource than you'd expect to need for an AI project. In particular, people who build one for the first time are usually amazed to discover that they require far fewer cases to make it work well than one would intuitively expect.