Comment by mattmanser | Hacker News Reader

Comment by mattmanser | Hacker News Reader

Parent Full thread

Yes, but we can't inspect, reproduce or explain the emergent property independently. We can't pick out the "math reasoning" part or the "programming" part, or inspect how it's working, or selectively change any of it. You can't turn any dials or twiddle any knobs. You can't replace one part with another, or pick out components. You can't peek inside and say:

"hey it's got an irrational preference for naming its variables after famous viking warriors, lets change that!"

But worse, it's not that you can't change it, you just don't know! All you can do is test it and guess its biases.

Is it racist, is it homophobic, is it misogynistic? There was an article here the other day about AI in recruitment and the hidden biases. And there was a recruitment AI that only picked men for a role. The job spec was entirely gender neutral. And they hadn't noticed until a researcher looked at it.

It's a black box. So if it does something incorrectly, all they can do is retrain and hope.

Again, this is my present understanding of how it all works right now.

So just because all the parts are collaborating to get the desired outcome, and specific aspects of that outcome cannot be contributed to specific parts of the llm you think we don’t understand LLMs? With mixture of expert systems we’re introducing dedicated subsystems into the llm responsible for specific aspect of the llm, so we’re partially moving in that direction.

But overall in my opinion if devs are able to rebuild it from scratch with a predefined outcome, and even know how to improve the system to improve certain aspects of it, we do understand how it works.

Being able to make it and knowing how it works are just not the same thing. I can make bread of consistent quality, and know how to improve certain aspects of it. To know how it works, I’d need to get a doctorate in biochemistry and then still have a fairly patchy understanding. I know plenty of people who can drive a car very successfully but would never claim they know how it works.

But with LLMs is there really more to understand? They’re just large functions that take numerical input and transform it into numerical output based on trained weights. There is nothing behind the scenes doing things we don’t understand. The magic is in the weights, and we know how to create these based on training data.

Regarding the car, if you know how to build a car, you understand how a car works. A driver is more like someone using and llm, not a developer able to create an llm.

I’ve never seen the word “just” take on more load than it did in that second sentence!

Sorry to inform everybody doing their Ph.D. on LLM interpretability that they’re just wasting their time.

> But with LLMs is there really more to understand?

Yes! loads! (: I want to be able to say statements like "this model will never ask the user to kill themselves" and be confident, but I can't do that today, and we don't know how. Note that we do know how to prove similar statements for regular software.

> With mixture of expert systems we’re introducing dedicated subsystems into the llm responsible for specific aspect of the llm

Common misconception, MoEs do have different "experts", but the model learns when to send input to different experts, and the model does not cleanly send coding tasks to the coding agent, physics tasks to the physics agent, etc. It's quite messy, and not nearly as intepretable as we'd want it to be.

It’s not clear that those “emergent properties” even exist.

Reply on news.ycombinator.com