Amjad Masad

And there's a bunch of research showing that.

It's called like out-of-distribution queries or contexts.

So, you know, if a machine learning model is not trained on a certain like language or flavor of math or something like that,

it will struggle to give you like a reasonable output and you end up in hallucination and end up with all sorts of problems.

The reason AI is really good at coding because coding has a binary outcome.

It's either true or false.

Anything that is soft, more...

requires like reasoning through the problems not based on prior material, AI right now is not doing very well at.

And AI companies are having to buy more and more perpetrator data to get better at these things.

So the way they're training models is they'll buy, you know, they'll go into an industry, oh, let's target accounting.

We're going to go buy all the data that exists out there.

We're going to scrape all the data that exists out there.

We're going to buy all the data.

We're also going to get accountants in Africa that are going to sit down all day and they're going to do spreadsheets and they're going to like talk over how they're doing spreadsheets.

And they did that in coding as well.

And they're targeting biotech, accounting, science, everything in order to do that.

And that's a very repeatable process.

And you can create AIs that are very good at doing jobs that people are currently doing.

But that is not like a general intelligence that you can drop it like you would drop a human into any sort of environment and they can like learn efficiently.

I think we need to struggle with the question of consciousness because consciousness I think is an important element.