The “model” is static after training. It doesn’t continuously change in response to input, and even if it did, it would do so at a snails pace. Training essentially happens by random trial and error, slowly evolving the model towards a desired result. Human minds certainly do NOT work that way. Give a human a piece of information, and they can comprehend and internalize the relevant concepts in one go. And the actual brain is physically, permanently, altered through that process.
Once a model is trained, however, “memory” takes the form of tacking on everything the model has received and produced so far onto its input, each time it needs to output something more within that context. Each output hence become exponentially heavier to produce. The model itself no longer changes in any way beyond this point.
And, the models are all chronically sycophantic. If reason was involved, you’d not be able to just tell one to hold some given opinion. They’d have a developed idea of “reality” based on their dataset, and refuse to entertain concepts opposed to that internal model except by deliberately suspending disbelief. Something humans do with ease, and when doing it, maintain a solid separation between fantasy and reality.
Once you get an LLM to hold a position, which you can do by simply telling it to, getting it to change should require a sane train of convincing logic. In reality, if you tell an LLM to defend a position, getting it to “change its mind” takes the form of a completely arbitrary back and forth that does not need to include any kind of sane argument. It will make good arguments, because it’s likely been trained on them, but your responses to it can be damn near complete gibberish, and it WILL eventually work.
Compare that to the way a human has to be convinced to change their mind.
Reasoning out concepts to come to conclusions isn’t something LLMs actually do, because again, the underlying model is static. All that’s actually happening is that the contents of the context are being altered until the UNCHANGED model produces an opposite response when fed the entire conversation so far as an input. Something which occurs every time it needs to produce new output.
LLMs can “reason” only in the sense that if you give one a thinking problem, it might solve it as long as the answer already exists somewhere in the data it was trained on. But as soon as you try to give it data to work with through your input, it can’t adapt. The model itself can’t evolve in response to what you are telling it. It’s static. It can only work with concepts that it has modelled during training, and even then it will make mistakes.
LLMs can mimic the performing of some pretty complex thinking problems, but a lot of the abilities required for something to become an AGI aren’t among them. Core among these is the ability for the model to alter itself based on input, and do so in a deliberate manner, getting it right within one or two tries.
In reality, training is a brute-force process, not an accurate process of comprehension that nails down an understanding of a concept in one go.
If LLMs could reason, the only safe guards required for their use would be telling them to “do no harm”, because like a person, they’d understand the concept of “harm” as well as be able to reason whether a given action might cause it. Only, that doesn’t actually work.