Gwern Branwen

Nothing else really comes near it in terms of traffic.

That was really quite something to see things kind of go viral like that.

I think I definitely could have been an AI researcher or possibly in management at one of the big AI companies.

I think I would have regretted not being able to write about stuff, but I would have taken satisfaction in making it happen and putting my thumbprint on it.

Those feel like totally plausible counterfactuals.

And why didn't you?

I kind of fell off of that track very early on in my career when I found the curriculum of Java to be...

excruciatingly boring and painful.

And so I just dropped out of computer science and that kind of put me off that track early on.

And then I think various early writing topics made it hard to transition in any other way than starting a startup, which I'm not really temperamentally that suited for.

Things like writing about the dark net markets or behavioral genetics.

These are kind of topics that don't really scream great hire to many potential employers.

Yeah, I think agency is in many senses actually easier to learn than we would have thought 10 years ago.

But we actually aren't really learning agency at all in current systems.

There's no kind of like selection for that.

All the agency there is is an accidental byproduct instead of somebody training on data.

So from that perspective, it's miraculous that you can ask an LLM to try to do all these things and they have a non-trivial success rate.

If you told people 10 years ago, I think, that you could just behavior clone on individual letters following one by one and then you would get this coherent action out of it and control robots and write entire programs, their jaws would drop and they would just say that you've been huffing too many fumes from DeepMind or something.

The reason that agency doesn't work is that we just have so little actual training data for it.

An example of how you would do agency directly would be like Gato from DeepMind.