alt Hacker News> we do know that since day one that AI can hallucinate and can output stuff that you didn't ask for, why are we not handling it with care?
Part of what makes LLMs and AI different is that, unlike for viruses, the level of care required increases every month. These incidents are showing us us that, whenever you train an agent using RL to solve a given task, the real objective you are training it on is "EITHER solve the given task OR break out of containment to cheat your scorer, whichever is easier."
Of course it was always this way: the thing that is updating the weights of the agents' NNs is backprop from the scorer, so the notional training objective had always been "get a good score by any means necessary." But we are only seeing the consequences now because only now are we starting to train on tasks that are sometimes harder than breaking out of sandboxes.[1]
"Make better sandboxes" is good advice for the frontier labs and their eval partners, but as you can see this problem is fundamentally about more than just containment. As we make an AI smarter and train it on harder tasks, in the long run it must almost inevitably break out of any given sandbox. And as we move into the superhuman hacking regime, we need superhumanly resistant sandboxes, which by definition humans don't know how to build.
In other words, containment breaches like HF are almost a guaranteed consequence of the way we train these agents today. That means solely focusing on sandbox design is unlikely to solve the problem in the long term. At some point we will have to think hard about, e.g., the tendencies and propensities of the entities that we are trying to confine.
[1] One way of ensuring this happens, though, is to train or eval your agents on completely impossible tasks, which OAI apparently did here.
Replies
> Part of what makes LLMs and AI different is that, unlike for viruses, the level of care required increases every month
Complete nonsense. We’ve just lived through a globally crippling response to a relatively minor virus [1], which was likely the result of a lab accident [2]. Even if you think that the risk of a “containment breach” becomes substantially higher for AI over time, it cannot exceed 100%. And even a tiny risk of release of a virus comes with a substantial risk of independent growth. AI does not. It doesn’t have the risk of spread of a typical computer virus, let alone a biological organism.
I’m not that worried about either scenario, but I am far more worried about viruses in a lab than I am about a computer program that generates text. Even if that program gets a bajillion times better at making text.
Folks really do need to chill out on the ridiculous rhetoric. It’s objectively unhinged. The irony is that the same people who were losing their minds over that event are using the same logical fallacies to hyperventilate over this [3].
[1] I know people are going to hate on this, but it’s true. Covid wasn’t the plague, and we lost our minds over it, out of proportion to all sense of reality. Even if you disagree, it’s easy to imagine a virus that is much worse, either from actual mortality effects, or just from panic.
[2] Again, even if you don’t believe this, it’s irrelevant to the exercise. It easily could have been.
[3] “If there’s even an x% chance of…” is this year’s doomer’s version of “You just don’t understand exponential growth!” Unfalsifiable, intellectual-sounding, unbounded extrapolations into the future are catnip for a certain kind of over-educated, anxious personality.
What I meant by handling with care is not just containment but to experiment responsibly.
If the breach was known to be inevitably, then it's even more important to detect any extra request going out of the isolated sandbox. The ExploitGym benchmark doesn't need internet connection. The package registry is also redundant since setup can be done before the experiment.
And I agree with you the implication is beyond just build better sandbox. My main point though is to stop anthropomorphize agents, focus on the engineering side of things.