Grant Harvey

But fundamentally, if you go into your phone and you go into WhatsApp and you type a word and you keep pressing next word, that's predictive text too.

It's just not as good because it's based on a much simpler model.

Like all AI is trying to do is find the next character or the next pixel or whatever.

It has no concept of truth.

It has no concept of anything.

And so these are just various different mechanisms to make it better at finding the next token.

And that could be because you've given it a load of label data, which is making it be rewarded for certain behaviors, or you've given it loads of information, which is making it be punished for certain behaviors.

So red teaming is exactly that.

You're trying to find, oh, the model has used...

racist language, or the model has exhibited some bias, or the model has used a false source.

For simple things, a model is going to be looking in its underlying dataset for reference data.

And there was some stat that came out that said 40 something percent of all the references the model is using is Reddit.

And like at a certain point, they're going to have like, is Reddit the best sort of truth?

Someone like Perplexity is going to have a very high weighting for certain sites it's using for citation because they're the ones that it trusts.

And that is going to come from a set of decisions that someone is making around which sites to trust.

And that will be because there'll be a set of hyperparameters which say, oh, Bloomberg, Reuters, like focus on those above Reddit or whatever random forum someone's saying something.

I'm fairly sure that the GPT-3 model was not that much better than the GPT-2 model.

The difference was the longer context window and the application that allowed it to be conversational.

That's what made it magic.

That's what made your grandmother go, oh, I'm talking to the computer.