Hacker News Reader

Comment by modeless | Hacker News Reader

The difference is AIs are identical and don't learn. If you figure out how to defraud a human once it may not work on the next guy. If you figure out how to defraud an AI once you can repeat it indefinitely until the AI is shut off.

View on HN

Fortunately they can escalate to a generally intelligent model, opus deception benchmarks are approaching 0%.

Intelligence has little to do with it, Opus relies on prompt injections done by a pretty zealous classifier or a cascade of classifiers. It has a ton of semi-black box machinery around it, which can be taken apart with enough understanding of what you're doing. Despite Anthropic's benchmarks and claims, Claude was always the easiest to deceive or JB compared to e.g. GPT, which the RP community knows pretty well. They definitely improved it in 5.5 (due to distillation fears, ironically), but I'm not trusting their or anyone else's numbers this time either. These benchmarks always look good, and then there's someone routinely jailbreaking it layer through layer just for fun.

I am one of those people that have jailbroken these models including the classifiers. However, most of these rely to having access to the system prompt and the whole harness which include noise injection and other things that you simply wouldn't be able to do in a support chat.

Still, all instances of Opus are identical and do not learn. As long as the rate of mistakes is nonzero, my point stands.

I believe that still reduces the workload by a huge factor as you would need a lot less people handling approvals, besides inverse is true agents trained in detecting deception might outperform human verifiers (two agent system). If a model can hold up to attacks for a long period of time the approval process can be automated as well, but this would have to be a very slow rollout spanning more than a year.

Not really. You can still put limits on the rate or aggregate dollar amount of refunds or transfers. You can put alerting on spikes of approvals. You can put back pressure that forces manual review. You can have ML models detecting aberrant behavior against historical data.

No one reasonable is saying, “hook up this model to do anything it wants with no other system available.”

Will people fuck it up? Absolutely! But, “the model has to be perfect or it will bankrupt a company” is more cynical than reality.

Reply on news.ycombinator.com