Raw LLM Responses

1:55 There are two issues with this statement: 1. The original GPT-1 was trained as merely a text predictor, OpenAI developed from this starting point to train for other things too - its initial training is as a text predictor, but it is also trained for accuracy, honesty etc (with issues obviously), plus a stage with human feedback learning, where humans vet answers and tell it what good responses look like and what poorer ones do. Performing these tasks does require understanding. Because what's the best way to accurately predict text correctly, factually, and helpfully? To understand it and the information you are providing. There's a word for describing when AIs don't actually understand what they're doing: overfitting. Does it understand everything it says or that is said to it? No. Neither do I. But it has some understanding of some things. 2. The way the AI is designed to run is to *output* one word at a time. But that does not mean it can't plan ahead. *Inside* its neural network it may well be thinking ahead. Internally, it might work out what its entire next sentence is going to be before outputting the first word. Saying "it only predicts one word at a time, it has no idea what it's going to say next" would be equally valid if I said it about a human talking to me - they only say one word at a time, they have no idea where their sentences are actually going. But of course the *output alone* is not all that's important - the human is thinking *inside* their mind. The AI may be doing the same, just in a different way. Of course everyone intrinsically understands this to not be true for humans because we are all human (except Zuckerberg) and we know from our own experience of our thoughts that humans can plan sentences ahead of time. But we just assume the AI doesn't for some reason?