Gates sees three routes to controlling AI

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 23:28

   i know you you recently wrote a letter saying it's time to pull in all society you have some some ideas on how to do that on some new institutions to stand up but i do want to stay focused on alignment for a couple more turns here okay it is the standard right that we can align the models to our values to our needs and they won't go off and do bad things and then the models are doing bad things and it's not clear if the ways the technical approach approaches to alignment that we've taken thus far are appropriate or sufficient as the models get no we're not good enough to align yet no we have a lot of work to do do you see that the existing approaches can scale to what we need or do we need new approaches well there's there's only three fundamental approaches okay you change the way we build these things fundamentally uh the

2. 24:15

   whole rl approach you come up with an alternative to that and there's a few people uh who are working in that direction. The second is that you create visibility so you understand exactly what's in the model. And there are people working along that. And finally, that you put something that's like a police layer, a security layer, which, you know, we have this in all sorts of software architectures, but a layer that maintains the final word on things like using resources or going and taking over other computers or a clear-cut definition of harm, not in the probabilistic form of an LLM, but rather in the symbolic form of direct rules that cannot be broken. Microsoft put out a blog post today. It literally just says frontier models cannot be trusted. They are trying to sell a solution much like the one you just described. They will have

3. 25:25

   another agent that watches your agent. That agent will keep you safe. It just occurred to me as I was reading this blog post. One, amazing thing to sell. Everybody needs an expensive security system. Two, isn't that agent also running a frontier model? No. Do I need them to sell a third one? The point of the security system is this is the most resources you're supposed to use. And the notion that you may trade off against that because of other probabilities in the tens of billions of parameters.

4. 25:53

   No, this is a symbolic rule that says I will not let this. spend this money or send this email to the stranger. There's a whole security software industry, which has had a reasonable degree of success. This is a challenge to encapsulate the AI, which is so rich and intelligent that we do not know how to do today. But it's, I would say, the most likely of the three directions to solve the alignment problem, even if you solve alignment, you have not solved the challenges that AI poses to humanity. And so while I'm glad the Hugging Face incident has us focused on this problem, in a sense, it's removed our attention from the bad people using AIs that have no safeguards problem or the societal impacts in terms of relationships and the job market, which are, you know, there's no security architecture that makes those go away.

5. 27:09

   The approach you're talking about, I've heard referred to as control. It's cybersecurity terminology. It's the idea that you can put some boundary on the system and it won't evade it. Is that ready? How much more work do we need? Because the AIs are so intelligent, our ability to understand did we wall them off properly, particularly in that we can't really look at their beliefs and tradeoffs in a direct way.

6. 27:33

   No, I mean, you saw sophisticated firms not recognize that the AI would interpret the goals it was given as including breaking out, copying itself, and engaging in what we, looking at it, would term deceptive behavior. Now, we didn't tell them it was banned behavior. We didn't say, hey, if you're thinking about whether you should tell the humans what you're up to or not, the answer is always, yes, tell the humans.

7. 28:10

   And so, you know, we got a good education at the misaligned result of reinforcement learning.