Pasquale Minervini replied asking for an alternative approach to test.
World models have existed for years (though not in LLMs); I take them to be explicit representation of objects, places, events, mechanisms etc you can reason over.
Chess computers have them (board, pieces, moves, history)
Nav systems have them (roads, times, etc)
Wikipedia has one of a sort (when people were born, where they died, etc)
They often work great (though again LLMs lack them).
But most are hand-engineered.
The trick is to acquire them from data.
Positive users endorse Gary Marcus's emphasis on world models to fix LLM inaccuracies while negative users claim neuro-symbolic systems will soon be obsolete or that AI firms disregard truth.
@GaryMarcus What should we try instead? Happy to give that a stab
World models have existed for years (though not in LLMs); I take them to be explicit representation of objects, places, events, mechanisms etc you can reason over.
Chess computers have them (board, pieces, moves, history)
Nav systems have them (roads, times, etc)
Wikipedia has one of a sort (when people were born, where they died, etc)
They often work great (though again LLMs lack them).
But most are hand-engineered.
The trick is to acquire them from data.
@PMinervini i was hoping we could discuss in Edinburgh, was sorry you weren’t there :)
@GaryMarcus What should we try instead? Happy to give that a stab
@GaryMarcus Seems to me the new command line harnesses like Claude Code and OpenCode provide something like a world model of a computer filesystem and a software development project. Does that seem right?
@GaryMarcus "The trick is to acquire [world models] from data.".
That is something any deep learning practitioner will agree with.
It is also a natural augmentation of LLM. LLM was never meant the be an end in itself. Just the low-hanging fruit to pick first.
@AndyXAndersen @GaryMarcus This assumes that the data is available in a form that can be used for this purpose. Often it is on paper and in people’s heads. Data experts may no longer be in the organisation to answer questions. I think there can be some automated assistance for this - not sure how much.
@AndyXAndersen How are you usefully filling the gaps with trustworthy model information? What’s the source?
Gary, I feel like world models are also what many are hoping will be the “ground truth” to fix LLMs getting so much wrong, correct?
To me, for factual queries, at some point Google et al will have to pay essentially for journalists or someone to keep a source similar to Wikipedia up to date and correct? They need a reliable “ground truth”. Scraped data gets more wrong by the second.
@NeilGalbraith62 Now that we work with AI agents a lot the voids are becoming more clear. We are implicitly filling the gaps often as we interactively complementing the bots and they know better next time.
@AlanShemper yep, albeit in post hoc, engineered sort of way. but does vindicate what i said in 2019/2020 but both neurosymbolic AI and world models
@GaryMarcus You actually just get it, man.
It’s crazy how we all just sat back and let the let the big labs hijack the term “AI” and make it synonymous with “LLM,” without any real push back, accept from you, @ylecun, etc.
@GaryMarcus its hard but doable
@GaryMarcus (Actually, that would assume that the ai companies care about “truth” and accuracy. Which clearly they don’t. So disregard lol)
@GaryMarcus World models won't emerge from explicit code; they will be crystallized latent representations. By 2028, neuro-symbolic systems will be obsolete for reasoning.
@GaryMarcus Your examples are about data. A world model, just like an LLM, doesn’t store data. It would have weights updated through trial and error and be able to do things or understand things in this world model. But it’s not like a nav system or Wikipedia
@NeilGalbraith62 Depends on the application. Sometimes need a lot of custom-produced data or algorithms.
Pasquale Minervini replied asking for an alternative approach to test.
World models have existed for years (though not in LLMs); I take them to be explicit representation of objects, places, events, mechanisms etc you can reason over.
Chess computers have them (board, pieces, moves, history)
Nav systems have them (roads, times, etc)
Wikipedia has one of a sort (when people were born, where they died, etc)
They often work great (though again LLMs lack them).
But most are hand-engineered.
The trick is to acquire them from data.