We have models that accurate classify things, e.g. whether or not an email is spam. There isn’t a fundamental limitation into building something like a truth classifier into a generative model so that it optimized for outputting “true” statements. The hardest part is probably identifying what is truth and what is falsehood. That’s a fundamental problem with humanity, not neural networks.
Well, we could quibble about what "fundamental" means but my point is that the way they train large language models doesn't work for this. Something different needs to happen.
Truth has nothing to do with humanity unless you mean the specific way humans construct belief systems.
Anyway I already told you the answer. The AI will need a series of trainable belief systems to verify whether statements are internally consistent. The strange part about this is that the AI would need to have a way to obtain validation and each prompt would have to derive a new belief system which you must use in the next prompt.
In other words, the model must be able to learn continuously. That is something that these single shot AI models are not capable of.
> There isn’t a fundamental limitation into building something like a truth classifier into a generative model so that it optimized for outputting “true” statements.
You're equivocating on "solved." Solved as in performing as well as humans, not solved in the mathematical sense which is both 1) not necessarily possible, and 2) nothing anybody has ever named as a test for AI.
No, that's correct. Checkers is solved; there is an algorithmic solution. Chess and Go have computer systems that exceed human performance, but are not solved.
"Solved" means having a solution to a problem. In context, we're talking bout whether or not neural networks can detect truth better than "decades of work by experts to reach consensus." So, in this case, solving would be detecting truth better than the status quo, not detecting truth 100% of the time. In the example of Go, the problem was "playing Go better than the best humans." So in that sense, the problem was solved. Adding your own, unfavorable definition of "solved" to the discussion is unwarranted.
“Hmm how should I get to work tomorrow? Normally I’d take the car, but after adopting a stance of distractive pedantism I realized that a car isn’t an acceptable solution to my transportation problem.”
Like please explain what definition of solved you are using. It’s not one most people would be familiar with.
Note that we are not talking about neural networks in general, but specifically the sort of generative autoregressive language model that Galactica is. What reason do we have to think that such a model is more likely to produce a true statement than a false one? - especially as just one misplaced truth-valued function or operator is likely to turn a true proposition into a false one. Truthfulness (not to be confused with truthiness) of their productions does not seem to be something we should expect from how they work, and the empirical evidence from Galactica supports this view.
Training a large model to guess when it doesn't know the answer results in fiction. They need to do something else to get nonfiction.
By contrast, for Go the model was trained not to make illegal moves, because checking for that as part of the training is easy and cheap.