If I could give it a novel task outside of its explicit training and see it actually improve just through accreting context, I'd be convinced it was thinking.
The opposite happens in practice. I test new models with two tasks: iteratively generating SVGs based on a text description with rendered rasters for feedback; and generating "Before and After" clues like on Jeopardy, where the response has two overlapping phrases such that the last word of the first phrase must be identical to the first word of the last phrase. I have yet to find a model that is consistently good at either. And actually they tend to exhibit context rot with these tasks, where they seem drunk or stoned and the quality degrades.
They're extremely good pattern filters, and that includes some level of logical reasoning. But they aren't reflective or adaptable. Just last night, for instance, I was teaching my son about rounding to the nearest millions. It became clear that he didn't know the place values of large numbers, so we reviewed that till he was consistently correct, and then he was consistently great at rounding to the nearest millions or ten millions or hundred billions or whatever. He's thinking. LLMs are not.
Now that you've spoiled it, models will be trained on these benchmarks, especially the second one that's very specific.
The second one has caused hours of entertainment over the past year or so. My kids find the LLM's failed clues and profuse apologies for getting these wrong hilarious, so it's become a family activity with me performing dramatic readings of the chat transcript with them. The LLM's apologies also seem to get more exaggerated as the context increases and the LLM seems to get more deranged.
I'd prefer the models to get better at SVG. I really hate working with the rasters that diffusion models generate, but the vector outputs are just really bad even when tokenizable like SVG. I've done some experimentation with trying to make these work better with some newer techniques with some success. But I also think the SVG Paths mini-language may be a bit too concise and unforgiving for LLMs to consistently get them right without specialized training.
see sibling comment,
> see it actually improve just through accreting context
this actually happens and has been tested.
> see sibling comment,
> > see it actually improve just through accreting context
> this actually happens and has been tested.
I specifically said a novel task outside of the explicit training. And I already agreed that the so-called thinking models do some level of logical reasoning. But being able to engage in some level of reasoning because it has learned logical inference rules doesn't mean it's actually thinking, regardless of what the researchers wish to call it.
Also, why does each model always fail at the two tests I give it? The models not only fail to improve, but they start to degrade after many subsequent iterations. Someone who can think would at least not get worse.
LLMs are filters or tuners for extremely subtle patterns, patterns that humans frankly are not great at finding. That's what the attention mechanism does: attend to the other tokens that are most related in a given context, even if that related context is distant in the token stream. Some patterns they fail to detect because they haven't been sufficiently trained or post-trained, and so the LLM just attends to noise (or at least that's what appears to be happening).
A lot of intelligence can be effectively mimicked through this pattern synthesis by transformer architecture alone. That's surprising. But I have yet to see them think.