Ask ChatGPT to estimate the carbs in your lunch. Now ask it again. And again. Five hundred times. You’d expect the same answer each time. It’s the same photo, the same model, the same question. But you won’t get the same answer. Not even close — and the differences are large enough to cause a
The embedding layer post tokenization is not just a probability machine the way you’re suggesting it. You can argue that it is probabilistic with inferred sentiment, but too many people think it works like how text prediction on your phone does and that is just factually inaccurate.
Verify output of course, but saying “it doesn’t understand anything” and “probability machine” is a borderline erroneous short sell. At the level of tokens it “understands” relationships, and those relationships are not probabilistic, though they are fundamentally approximated based on a training corpus.
The fact that it uses a non-trivial neural network.
If it was simply a rate count of based on a corpus of how much time each word is followed by each it wouldn’t be stronger than keyboard word predictions. To make accurate suggestions requires emergence of primitive reasoning on the semantics of the tokens, LLM neural networks (transformers) can be analyzed to find subnetworks dedicated to modeling reality. It is still probability, but saying it’s just probability is not faithful
It’s still just predicting the next token, it’s just using more past data points than your keyboard. The rest of the phenomena are emergent from that. I think it’s important to keep that in mind given how much they can imitate human reasoning.
The embedding layer post tokenization is not just a probability machine the way you’re suggesting it. You can argue that it is probabilistic with inferred sentiment, but too many people think it works like how text prediction on your phone does and that is just factually inaccurate.
Verify output of course, but saying “it doesn’t understand anything” and “probability machine” is a borderline erroneous short sell. At the level of tokens it “understands” relationships, and those relationships are not probabilistic, though they are fundamentally approximated based on a training corpus.
Can you explain how it’s more than probability? It’s using a neural network to guess the most likely next token, isn’t it?
The fact that it uses a non-trivial neural network. If it was simply a rate count of based on a corpus of how much time each word is followed by each it wouldn’t be stronger than keyboard word predictions. To make accurate suggestions requires emergence of primitive reasoning on the semantics of the tokens, LLM neural networks (transformers) can be analyzed to find subnetworks dedicated to modeling reality. It is still probability, but saying it’s just probability is not faithful
It’s still just predicting the next token, it’s just using more past data points than your keyboard. The rest of the phenomena are emergent from that. I think it’s important to keep that in mind given how much they can imitate human reasoning.
If you recognize there’s emergence, the “it’s just probability” take is misleading