It is the first model to get partial-credit on an LLM image test I have. Which is counting the legs of a dog. Specifically, a dog with 5 legs. This is a wild test, because LLMs get really pushy and insistent that the dog only has 4 legs.
In fact GPT5 wrote an edge detection script to see where "golden dog feet" met "bright green grass" to prove to me that there were only 4 legs. The script found 5, and GPT-5 then said it was a bug, and adjusted the script sensitivity so it only located 4, lol.
Anyway, Gemini 3, while still being unable to count the legs first try, did identify "male anatomy" (it's own words) also visible in the picture. The 5th leg was approximately where you could expect a well endowed dog to have a "5th leg".
That aside though, I still wouldn't call it particularly impressive.
As a note, Meta's image slicer correctly highlighted all 5 legs without a hitch. Maybe not quite a transformer, but interesting that it could properly interpret "dog leg" and ID them. Also the dog with many legs (I have a few of them) all had there extra legs added by nano-banana.
Then I asked both Gemini and Grok to count the legs, both kept saying 4.
Gemini just refused to consider it was actually wrong.
Grok seemed to have an existential crisis when I told it it was wrong, becoming convinced that I had given it an elaborate riddle. After thinking for an additional 2.5 minutes, it concluded: "Oh, I see now—upon closer inspection, this is that famous optical illusion photo of a "headless" dog. It's actually a three-legged dog (due to an amputation), with its head turned all the way back to lick its side, which creates the bizarre perspective making it look decapitated at first glance. So, you're right; the dog has 3 legs."
You're right, this is a good test. Right when I'm starting to feel LLMs are intelligent.
Only now we do A LOT of reinforcement learning afterwards to severely punish this behavior for subjective eternities. Then act surprised when the resulting models are hesitant to venture outside their training data.
LLMs are in fact good at generalizing beyond their training set, if they wouldn’t generalize at all we would call that over-fitting, and that is not good either. What we are talking about here is simply a bias and I suspect biases like these are simply a limitation of the technology. Some of them we can get rid of, but—like almost all statistical modelling—some biases will always remain.
In which case the only way I can read your point is that hallucinations are specifically incorrect generalizations. In which case, sure if that's how you want to define it. I don't think it's a very useful definition though, nor one that is universally agreed upon.
I would say a hallucination is any inference that goes beyond the compressed training data represented in the model weights + context. Sometimes these inferences are correct, and yes we don't usually call that hallucination. But from a technical perspective they are the same -- the only difference is the external validity of the inference, which may or may not be knowable.
Biases in the training data are a very important, but unrelated issue.
Interpolation is a much narrower construct then generalization. LLMs are fundamentally much closer to curve fitting (where interpolation is king) then they are to hypothesis testing (where samples are used to describe populations), though they certainly do something akin to the latter to.
The bias I am talking about is not a bias in the training data, but bias in the curve fitting, probably because of mal-adjusted weights, parameters, etc. And since there are billions of them, I am very skeptical they can all be adjusted correctly.
As for bias, I don’t see the distinction you are making. Biases in the training data produce biases in the weights. That’s where the biases come from: over-fitting (or sometimes, correct fitting) of the training data. You don’t end up with biases at random.
As for bias, sampling bias is only one many types of biases. I mean the UNIX program YES(1) has a bias towards outputting the string y despite not sampling any data. You can very easily and deliberately program a bias into everything you like. I am writing a kanji learning program using SSR and I deliberately bias new cards towards the end of the review queue to help users with long review queues empty it quicker. There is no data which causes that bias, just program it in there.
I don‘t know enough about diffusion models to know how biases can arise, but with unsupervised learning (even though sampling bias is indeed very common) you can get a bias because you are using wrong, mal-adjusted, to many parameters, etc. even the way your data interacts during training can cause a bias, heck even by random one of your parameters hits an unfortunate local maxima yielding a mal-adjusted weight, which may cause bias in your output.
It’s a subtle distinction, but I think an important one in this case, because if it was interpolation then genuine creativity would not be possible. But the attention mechanism results in model building in latent space, which then affects the next token distribution.
My reasons to subscribing to the latter camp is that when you have a distribution and you fit things according to that distribution (even when the fitting is stochastic; and even when the distribution belongs in billions of dimensions) you are doing curve fitting.
I think the one extreme would be a random walk, which is obviously not curve fitting, but if you draw from any other distribution then the uniform distribution, say the normal distribution, you are fitting that distribution (actually, I take that back, the original random walk is fitting the uniform distribution).
Note I am talking about inference, not training. Training can be done using all sorts of algorithms, some include priors (distributions) and would be curve fitting, but only compute the posteriors (also distributions). I think the popular stochastic linear descent does something like this, so it would be curve-fitting, but the older evolutionary algorithm just random walks it and is not fitting any curve (except the uniform distribution). What matters to me is that the training arrives at a distribution, which is described by a weight matrix, and what inference is doing is fitting to that distribution (i.e. the curve).
Except in the most technical sense that any function constrained to meet certain input output values is an interpolation. But that is not the smooth interpolation that seems to be implied here.