Do humans behave like LLMs?
geoffmulgan.com
geoffmulgan.com
> Many people in many fields, including quite a few in academia, also act rather like a ChatGPT, particularly in academic disciplines that don’t do much empirical work, work with facts or testable hypotheses
> I now realise that me not understanding what someone is saying is sometimes a sign that they don’t know what they’re talking about, and that they are essentially acting like an LLM. This would become apparent if they were ever interviewed in the way that politicians are sometimes interviewed in the media, with forensic questioning: ‘what do you actually mean by x?
I won't continue to spoil a short, interesting piece by continuing to steal quotes from it!
The main differences between human brains and AI at the moment are (a) a lot more randomness and indeterminacy in humans due to the way our brains are constructed and (b) the total amount of human brainpower exceeds the total amount of AI brainpower. The means that at the moment we appear more creative but soon (i.e. in hundreds of years rather than millions) I think we'll have AIs that are indistinguishable from people
Actual researches returns unequivocal no as far as LLMs being anything like our brain or a language model (worth googling for research how NNs are un-able to simulate any functionality of actual brain). In fact, even the NN concept as such is a pretty much a failure as an analogy to the brain. NN was in fact (many have argued) a marketing term more than anything.
Of cause, the underlying phenomenology is worth to question/investigate and research further, in both domains (NN and human learning). The failure of NN analogy to the brain, I would argue, is the same as failure of todays (hyped and false) analogy of CO2 as a greenhouse gas. Nature generally does not optimize as humans to a single metric or a variable. On the contrary, collectively, nature seem to be more concerned with stability/robustness than any individual or a single group fitness. That statements would take too long to unpack, but I take it a better model to start with...(some researches already had began of a decade or so).