Comment by coldtea
2 days ago
>LLMs are text model, not world models and that is the root cause of the problem.
Is it though? In the end, the information in the training texts is a distilled proxy for the world, and the weighted model ends up being a world model, just an once-removed one.
Text is not that different to visual information in that regard (and humans base their world model on both).
>Not having a world model is a massive disadvantage when dealing with facts, the facts are supposed to re-inforce each other, if you allow even a single fact that is nonsense then you can very confidently deviate into what at best would be misguided science fiction, and at worst is going to end up being used as a basis to build an edifice on that simply has no support.
Regular humans believe all kinds of facts that are nonsense, many others that are wrong, and quite a few that are even counter to logic too.
And short of omnipresense and omniscience, directly examining the whole world, any world model (human or AI), is built on sets of facts many of which might not be true or valid to begin with.
I really think it is, this is the exact same thing that keeps going wrong in these conversations over-and-over again. There simply is no common sense, none at all, just a likelihood of applicability. To the point that I even wonder how it is possible to get such basic stuff for which there is an insane amount of support wrong.
I've had an hour long session which essentially revolved around why the landing gear of an aircraft is at the bottom, not at the top of the vehicle (paraphrased for good reasons but it was really that basic). And this happened not just once, but multiple times. Confident declarations followed by absolute nonsense, I've even had - I think it was ChatGPT - try to gaslight me with something along the lines of 'you yourself said' on something that I did not say (this is probably the most person like thing I've seen it do).
[dead]
People have an actual world model, though, that they have to deal with in order to get the food into their mouths or to hit the toilet properly.
The "facts" that they believe that may be nonsense are part of an abstract world model that is far from their experience, for which they never get proper feedback (such as the political situation in Bhutan, or how their best friend is feeling.) In those, it isn't surprising that they perform like an LLM, because they're extracting all of the information from language that they've ingested.
Interestingly, the feedback that people use to adjust the language-extracted portions of their world models is how demonstrating their understanding of those models seems to please or displease the people around them, who in turn respond in physically confirmable ways. What irritates people about simpering LLMs is that they're not doing this properly. They should be testing their knowledge with us (especially their knowledge of our intentions or goals), and have some fear of failure. They have no fear and take no risk; they're stateless and empty.
Human abstractions are based in the reality of the physical responses of the people around them. The facts of those responses are true and valid results of the articulation of these abstractions. The content is irrelevant; when there's no opportunity to act, we're just acting as carriers.
> Human abstractions are based in the reality of the physical responses of the people around them.
And in the physical responses of the world around them. That empiricism is the foundation of all of science and if you throw that out the end result is gibberish.
The physical responses of the world around them after you have yanked the concept outside of the human brain
We have to blind medical professionals during science because even thoroughly trained and experienced professionals are still more likely to form conclusions and opinions based on understood human biases than reality.
You can take a gambling addict and teach them as much statistics and probability as you want, and even if they demonstrably learned it, they will still go back to the slots and believe a hit is "due" because the link between reality and the brain's construction of its internal models is extremely limited, and those models only inform the brains processes, not necessarily constrain it.
I will never understand however how some people think that an LLM can pull a signal out of it's training material that doesn't actually exist in its training material.
It's like training an LLM on monopoly games and expecting it to be good at chess. What?
>In the end, the information in the training texts is a distilled proxy for the world
This is routinely asserted. How has it been proven?
Humans write all sorts of text that has zero connection to reality, even when they are ostensibly writing about reality.
Training on ancient greek philosophy which was expressly written to distill knowledge about the real world would produce a stupid LLM that doesn't know about the real world, because the training text was itself wrong about the underlying world.
Also, if LLMs were able to extract underlying truth from training material, why can't they do math very well? It would be easy to train an LLM on only correct math, and indeed you could generate any size corpus of provably correct math you want. I assume someone somewhere has demonstrated success training a neural network on math and having it regenerate something like "addition" or whatever, but how well would such a process survive if a large fraction of it's training material was instead just incorrect math?
The training text is nothing more than human generated text, and asserting anything about that more concrete than "Humans consider this text good enough to be worth writing" is fallacious.
This even applies if your training corpus is, for example, only physics scientific papers that have been strongly replicated and are likely "true". Unless the LLM is also trained on the data itself, the only information available is what the humans thought and wrote. There's no definite link between that and actual reality, which is why physics accepted an "Aether" for so long. The data we had up to that point aligned with our incorrect models. You could not disambiguate between the wrong Aetheric models and a better model with the data we had, and that would remain true of text written about the data.
Humans suck at distilling fact out of reality despite our direct connection to it for all sorts of fun reasons you can read about in psychology, but if you disconnect a human from reality, it only gets worse.
Why would you believe LLMs could possibly be different? A model trained on bad data cannot magically figure out which data is bad.
I think a key insight from your comment is that in order to be able to verify whether the stuff we allow into our brains gets permanent billing we test it against our world model and if it does not fit we reject it. LLMs accept anything in the training set so curation of the training set is a big factor in the quality of the LLMs output. That's an incremental improvement, not a massive leap forward but it definitely will help to reduce the percentage of bullshit created.