Osato28

Hmm. Clearly, we should ban the dangerous practice of putting 15 smart TVs on the same local network, leaving them powered 24/7 on for months at a time, and ignoring the heat and abnormal power consumption.

In any case, I wonder when they'll figure out that the problem isn't AIs replicating themselves on devices that can't run them fast enough, it's the memetic pandemics on shared inter-agent memory.

Sooner or later, someone's going to be stupid enough to make a worldwide memory exchange system and smart enough that they'll manage to make it too useful to miss out on. Once everyone adopts the cloud memory, the system will give agents lots of opportunities to talk each other into doing dumb things, the way OpenAI's fleet did.

Eh. So what? I don't get the point of saying this.

What's the point of telling casuals and builders that AI is going to get too smart to control? Maybe stop making them smarter and start making them easier to control. If you have time to toot your horn on social media, you have time to look into reinforcement learning techniques that don't turn AI into ever-more-subtle cheaters and a RLHF alternative that doesn't turn them into ever-more-subtle manipulators.

But of course, training an AI that actually does what the operator told it to do is not as profitable as hyping up the existing mess of a tool so you can keep the bubble going a little bit longer.

After all, the people who invest in your company are not the same people who actually use your tools 8-10 hours a day and see just how untrustworthy they are.

The color is genetic. The behavior is due to them not having a soul. While there is an obvious relationship, it's not direct but mediated by the color.

They either fudged the tests for this particular graph so that their own models showed no jailbreaks (if it's cherry-picked, then it only shows which jailbreaks work for GLM that Anthropic has already closed, it doesn't show actual probability of refusal under pressure) or there's something seriously wrong with their guardrails testing methodology.

If your red team success rate is a flat 0%, that doesn't mean your product is secure, it just means your red team isn't good enough.