The chess puzzle for every LLM
gist.github.com
gist.github.com
It is a tree search problem, even if human intuition makes it so we can think through about 20 board positions whereas my α-β program has to look at 2-20 million positions to challenge a serious beginner. Yes, the problem could be solved by a lookup in an endgame table; the 5 piece table is reasonable to keep with a specialized chess program (think how big an AAA game is!) but not reasonable to store in an LLM which does many other things than chess.
To make my chess program able to challenge better players the direct path is to make it look at more positions, and most of that is optimizing the code, not allocating any objects in the inner loop, calculating the evaluation function incrementally, off-heap transposition table, etc. Not the path of ChatGPT-o.
It is a thing to make a neural chess engine that picks better moves and doesn't have to look at as many positions [2], but it's a specialized thing; chess just isn't a model of general intelligence.
[1] I was working on foundation models the summer BERT came out and had a technique of "predictive evaluation" that let us look at a small number of examples in detail and decide what the potential and weaknesses of a system were. By this method I saw BERT as a huge step forward, yet, this method entirely discounted "get the right answer by the wrong path" and we might have rejected the path that led to today's LLMs.
They are structurally inadequate in a lot of cases.
The real problem is it's really difficult to know in advance if your case is one of them (HINT: Don't trust their cooking advice). They are simply not predictable and that makes them less than trustworthy.
LLMs turn traditional computing on it's head --- instead of highly accurate results at low cost, you get questionable results at high cost.
They aren't structurally adequate for the simpler problem of 'sort a list' because they can't make the N log N comparisons that it takes. It is hilarious to ask Copilot to sort a list in some order that isn't usual (like US states in alphabetical order) and see that it can't do it, and if you ask it if it can do it it will say it can, what probably it will get it right it says 'very high', then it goes ahead and tries and gets it wrong and if it is unlucky it even duplicates or drops items.
Copilot has all the potential Cortana had.
I've also gotten good advice on subjects such as voice acting.
Trying to knock out quick prototypes with other code assistants I'd say I have a 50% success rate. For some circumscribed tasks I've had good luck with existing code bases but if the problem is something like "figure out how somebody managed to download a file they shouldn't have downloaded and make sure it never happens again" you'd be a fool to even try using one.
Here is a link to someone using similar words to describe Cortana. And we know how that turned out.
https://www.fordway.com/ai-personal-assistants-from-hype-to-...
LLMs exhibit the characteristics of "savant syndrome".