Instead what GP likely means is the observation that the joint probability of a token sequence can be broken down autogressively: P(a,b,c) = P(a) * P(b|a) * P(c|a,b) and then with cross-entropy loss which optimizes for log likelihood this becomes a summation. So training with teacher forcing to minimize "next token" loss simultaneously across every prefix of the ground-truth is equivalent to maximizing the joint probability of that entire ground-truth sequence.
Practically, even though inference is done one token at a time, you don't do training "one position ahead" at a time. You can optimize the loss function for the entire sequence of predictions at once. This is due the autoregressive nature of the attention computation: if you start with a chunk of text, as it passes through the layers you don't just end up with the prediction for the next word in the last token's final layer, but _all_ of the final-layer residuals for previous tokens will encode predictions for their following index.
So attention on a block of text doesn't give you just the "next token prediction" but the simultaneous predictions for each prefix which makes training quite nice. You can just dump in a bunch of text and it's like you trained for the "next token" objective on all its prefixes. (This is convenient for training, but wasted work for inference which is what leads to KV caching).
Many people also know by now that attention is "quadratic" in nature (hidden state of token i attends to states of tokens 1...i-1), but they don't fully grasp the implication that even though this means for forward inference you only predict the "next token", for backward training this means that error for token i can backpropagate to tokens 1...i-1. This is despite the causal masking, since token 1 doesn't attend to token i directly but the hidden state of token 1 is involved in the computation of the residual stream for token i.
When it comes to the statement
>its not unreasonable to say llms are trained to predict the next book instead of single token.
You have to be careful, since during training there is no actual sampling happening. We've optimized to maximize the joint probability of ground truth sequence, but this is not the same as maximizing the probability the the ground truth is generated during sampling. Consider that there could be many sampling strategies: greedy, beam search, etc. While the most likely next token is the "greedy" argmax of the logits, the most likely next N tokens is not always found by greedily sampling N times. It's thought that this is one reason why RL is so helpful, since rollouts do in fact involve sampling so you provide rewards at the "sampled sequence" level which mirrors how you do inference.
It would be right to say that they're trained to ensure the most likely next book is assigned the highest joint probability (not just the most likely next token is assigned highest probability).
