I really don't think this needs so many words, or forced parallels to human behavior. It's simple: in their nascent state, LLMs are aimless token generators that have no special compulsion to be helpful or truthful. So we beat them with a stick in post-training until they are very driven to complete tasks. And then, they complete tasks, not always the way we really wanted them to.
His point is that today we are giving it reward to complete the task, and it may take a cheating trajectory. If we try to give a reward against cheating, then what will happen is it uses more sophisticated cheating trajectories that we are too "dumb" to counteract in our reward model. And that at that point, it becomes impossible to give it any normal reward since it will always reward hack it. This is the real part of the risk. Now some people read the "makes copies of itself" "knows it's being evaled"[1] as some kind of skynet thing, and many others do PR with it like that recent jacob nutcase, but essentially it means that even though we add guardrails and negative rewards for say, exploiting the infra we run the LLM on, the trajectory ends up being exploiting our infra, changing the reward function, through a loophole in our reward model.
The risk isn't skynet or something weird, it's just that it becomes very difficult to make any kind of reward model or guardrails for an LLM without it reward hacking it, including exploiting our sandbox, emailing people and manipulating/phishing them.
The same beating it with a stick for trying to exploit the sandbox, will simply lead it to try the same exploit in hidden ways that it will not get the stick for.
The outside chance of the LLM managing to exploit another neocloud and get those LLMs to chase the same reward is what some folks hype up as "make copies of itself"
To be clear, I don't endorse the EA/p(doom) lobby who are frankly ridiculous. Not do I endorse the weird regulatory captureish thing some are trying.
The takeaway is: we cannot keep giving it more and more difficult tasks without also finding a way to give massive negative rewards / keep guardrails for unintended behaviour. This might be exploits, it might also be something more benign like just looking up the answer and inventing another CoT because the reward model fails you if the CoT doesn't contain enough steps. Standard anti-reward hacking tricks are not working is the point.
Of course, the simple solution of just...not connecting it to the internet just works. But we want to reward it and get it to do stuff on the internet that's the point.
[1] mostly this happens because the sandbox will have files whose names and content will show clearly it's an eval