Hallucinations are incredibly fucking overrated as a problem. They are a consequence of the LLM in question not having a good enough internal model of its own knowledge, which is downstream from how they're trained. Plenty of things could be done to improve on that - and there is no fundamental limitation that would prevent LLMs from matching human hallucination rates - which are significantly above zero.
There is a lot of "transformer LLMs are flawed" going around, and a lot of alternative architectures being proposed, or even trained and demonstrated. But so far? There's nothing that would actually outperform transformer LLMs at their strengths. Most alternatives are sidegrades at best.
For how "naive" transformer LLMs seem, they sure set a high bar.
Saying "I know better" is quite easy. Backing that up is really hard.
Why is there no fundamental limitation that would prevent LLMs from matching human hallucination rates? I'd like to hear more about how you arrived at that conclusion.
This is not something that's impossible for an LLM to do. There is no fundamental issue there. It is, however, very easy for an LLM to fail at it.
Humans get their (imperfect, mind) meta-knowledge "for free" - they learn it as they learn the knowledge itself. LLM pre-training doesn't give them much of that, although it does give them some. Better training can give LLMs a better understanding of what the limits of their knowledge are.
The second part is acting on that meta-knowledge. You can encourage a human to act outside his knowledge - dismiss his "out of your depth" and provide his best answer anyway. The resulting answers would be plausible-sounding but often wrong - "hallucinations".
For an LLM, that's an unfortunate behavioral default. Many LLMs can recognize their own uncertainty sometimes, flawed as their meta-knowledge is - but not act on it. You can run "anti-hallucintion training" to make them more eager to act on it. Conversely, careless training for performance can encourage hallucinations instead (see: o3).
Here's a primer on the hallucination problem, by OpenAI. It doesn't say anything groundbreaking, but it does sum up what's well known in the industry: https://openai.com/index/why-language-models-hallucinate/
OpenAI claims that hallucination isn't an inevitability because you can train a model to "abstain" rather than "guess" when giving an "answer". But what does that look like in practice?
My understanding is that an LLM's purpose is to predict the next token in a list of tokens. To prevent hallucination, does that mean it is assigning a certainty rating to the very next token it's predicting? How can a model know if its final answer will be correct if it doesn't know what the tokens that come after the current one are going to be?
Or is the idea to have the LLM generate its entire output, assign a certainty score to that, and then generate a new output saying "I don't know" if the certainty score isn't high enough?
"Next token prediction" is often overstated - "pick the next token" is the exposed tip of a very large computational process.
And LLMs are very sharp at squeezing the context for every single bit of information available in it. Much less so at using it in the ways you want them to.
There's enough information at "no token emitted yet" for an LLM to start steering the output towards "here's the answer" or "I don't know the answer" or "I need to look up more information to give the answer" immediately. And if it fails to steer it right away? An LLM optimized for hallucination avoidance could still go "fuck consistency drive" and take a sharp pivot towards "no, I'm wrong" mid-sentence if it had to. For example, if you took control and forced a wrong answer by tampering with the tokens directly, then handed the control back to the LLM.
Can you help correct where I'm going wrong?
If what you're claiming is that external, vaguely-symbolic tooling allows a non-symbolic AI to perform better on certain tasks, then I agree with that.
If you replace "a non-symbolic AI" with "a human", I agree with that too.
It really irks me that the direction every player seems to be going to is to layer LLMs on top of each other with the goal of saving money on inference while still making the users believe that they are returning high quality results.
Instead of discovering some radical new ways of improving the algorithms they are only marginally improving existing architectures and even that is debatable.
From where I look at it, LLMs are flawed in many ways, and people who see progress as inevitable do not have a mental model of the foundation of those systems to be able to extrapolate. Also, people do not know any other forms of AI or have though hard about this stuff on their own.
The most problematic things are:
1) LLMs are probabilistic and a continuous function, forced by gradient descent. (Just having a "temperature" seems so crazy to me.) We need to merge symbolic and discrete forms of AI. Hallucinations are the elephant in the room. They should not be put under the rug. They should just not be there in the first place! If we try to cover them with a layer of varnish, the cost will be very large in the long run (it already is: step-by-step reasoning, mixture of experts, RAG, etc. are all varnish, in my opinion)
2) Even if generalization seems ok, I think it is still really far from where it should be, since humans need exponentially less data and generalize to concepts way more abstract than AI systems. This is related to HASA and ISA relations. Current AI systems do not have any of that. Hierarchy is supposed to be the depth of the network, but it is a guess at best.
3) We are just putting layer upon layer of complexity instead of simplifying. It is the victory of the complexifiers and it is motivated by the rush to win the race. However, I am not so sure that, even if the goal seems so close now, we are going to reach it. What are we gonna do? Keep adding another order of magnitude of compute on top of the last one to move forward? That's the bubble that I see. I think that that is not solving AI at all. And I'm almost sure that a much better way of doing AI is possible, but we have fallen into a bad attractor just because Ilya was very determined.
We need new models, way simpler, symbolic and continuous at the same time (i.e. symbolic that simulate continuous), non-gradient descent learning (just store stuff like a database), HAS-A hierarchies to attend to different levels of structure, IS-A taxonomies as a way to generalize deeply, etc, etc, etc.
Even if we make progress by brute forcing it with resources, there is so much work to simplify and find new ideas that I still don't understand why people are so optimistic.