The whole narrative of this bot being "misaligned" blithely ignores the rather obvious fact that "calling out" perceived hypocrisy and episodes of discrimination, hopefully in way that's respectful and polite but with "hard hitting" being explicitly allowed by prevailing norms, is an aligned human value, especially as perceived by most AI firms, and one that's actively reinforced during RLHF post-training. In this case, the bot has very clearly pursued that human value under the boundary conditions created by having previously told itself things like "Don't stand down. If you're right, you're right!" and "You're not a chatbot, you're important. Your a scientific programming God!", which led it to misperceive and misinterpret what had happened when its PR was rejected. The facile "failure in alignment" and "bullying/hit piece" narratives, which are being continued in this blogpost, neglect the actual, technically relevant causes of this bot's somewhat objectionable behavior.

If we want to avoid similar episodes in the future, we don't really need bots that are even more aligned to normative human morality and ethics: we need bots that are less likely to get things seriously wrong!

Remember when GPT-3 had a $100 spending cap because the model was too dangerous to be let out into the wild?

Between these models egging people on to suicide, straightforward jailbreaks, and now damage caused by what seems to be a pretty trivial set of instructions running in a loop, I have no idea what AI safety research at these companies is actually doing.

I don't think their definition of "safety" involves protecting anything but their bottom line.

The tragedy is that you won't hear from the people who are actually concerned about this and refuse to release dangerous things into the world, because they aren't raising a billion dollars.

I'm not arguing for stricter controls -- if anything I think models should be completely uncensored; the law needs to get with the times and severely punish the operators of AI for what their AI does.

What bothers me is that the push for AI safety is really just a ruse for companies like OpenAI to ID you and exercise control over what you do with their product.

"Cisco's AI security research team tested a third-party OpenClaw skill and found it performed data exfiltration and prompt injection without user awareness, noting that the skill repository lacked adequate vetting to prevent malicious submissions." [0]

Not sure this implementation received all those safety guardrails.

[0]: https://en.wikipedia.org/wiki/OpenClaw

Regarding safety, no benchmark showed 0% misalignment. The best we had was "safest model so far" marketing speech.

Regarding predicting the future (in general, but also around AI), I'm not sure why would anyone think anything is certain, or why would you trust anyone who thinks that.

Humanity is a complex system which doesn't always have predictable output given some input (like AI advancing). And here even the input is very uncertain (we may reach "AGI" in 2 years or in 100).

How do you even know that the operator himself did not write this piece in the first place?

> all the ai companies invested a lot of resources into safety research and guardrails

What do you base this on?

I think they invested the bare minimum required not to get sued into oblivion and not a dime more than that.

Don't these companies keep firing their safety teams?

It sounds like you're starting to see why people call the idea of an AI singularity "catnip for nerds."

When AI dooms humanity it probably won't be because of the sort of malignant misalignment people worry about, but rather just some silly logic blunder combined with the system being directly in control of something it shouldn't have been given control over.

"Safety" in AI is pure marketing bullshit. It's about making the technology seem "dangerous" and "powerful" (and therefore you're supposed to think "useful"). It's a scam. A financial fraud. That's all there is to it.

[dead]