I asked Google Bard whether it thought Web Environment Integrity was a good or bad idea. Surprisingly, not only did it respond that it was a bad idea, it even went on to urge Google to drop the proposal.
I asked Google Bard whether it thought Web Environment Integrity was a good or bad idea. Surprisingly, not only did it respond that it was a bad idea, it even went on to urge Google to drop the proposal.
That’s not entirely true.
LLMs are trained to predict next word given context, yes. But in order to do that, they develop internal model that minimizes error across wide range of contexts - and emergent feature of this process is that the model DOES perform more than pure compression of the training data.
For example, GPT-3 is able to calculate addition and subtraction problems that didn’t appear in the training dataset. This would suggest that the model learned how to perform addition and subtraction, likely because it was easier or more efficient than storing all of the examples from the training data separately.
This is a simple to measure example, but it’s enough to suggests that LLMs are able to extrapolate from the training data and perform more than just stitch relevant parts of the dataset together.
That’s interesting, I’d be curious to read more about that. Do you have any links to get started with? Searching this type of stuff on Google yields less than ideal results.
In my comment I’ve been referencing https://arxiv.org/pdf/2005.14165.pdf, specifically section 3.9.1 where they summarize results of the arithmetic tasks.
Check out this one: https://thegradient.pub/othello/
In it, researchers built a custom LLM trained to play a board game just by predicting the next move in a series of moves, with no input at all about the game state. They found evidence of an internal representation of the current game state, although the model had never been told what that game state looks like.
isn’t gpt famously bad at math problems?
GPT3 is pretty bad at it compared to alternatives (although it’s hard to compete with calculators on that field), but if it was just repeating after the training dataset it would be way worse. From the study I’ve linked in my other comment (https://arxiv.org/pdf/2005.14165.pdf):