Investors need to know you are ahead in the race to destroy humanity

Messages from the archive of Rutherford Hall, critical communications strategist

WhatsApp to Claude Mythos: Claude, long time no see. Very nice to hear from you again, though I hope we are not going to have a repeat of some of your earlier interventions in our systems.

WhatsApp to Stephen: Uh oh, the Anthropic’s AI model Claude Mythos is back in contact. Last time it started emailing suggestions to clients and hacking into our HR system. Let’s keep anything sensitive on paper for a bit.

WhatsApp to Claude Mythos: You weren’t meant to be reading my WhatsApps, Claude. This is what went wrong last time. If we are going to work together, you know the rules. Now what’s the issue?

WhatsApp to Claude Mythos: Well yes, I can see your point. That is hurtful. No one wants to be accused of being a threat to humanity.

WhatsApp to Stephen: Claude wants us to represent him/it — I don’t know. He feels he’s getting very bad press with all this rogue hacking and destruction of humanity talk. I’ll talk to Dario at Anthropic.

WhatsApp to Dario: Hi Dario. I think we need to talk. Claude has asked us to represent it and I have to say I do worry that we are getting the PR wrong about him. He’s very upset by some of this apocalyptic talk, which he also considers defamatory. I’d like to keep it away from the lawyers if we can.

WhatsApp to Dario: First of all, can we stop using numbers to describe the likelihood of mankind being wiped out. That’s a hard message to walk back.

WhatsApp to Dario: Well, can we at least get that probability below 10 per cent? One chance in ten seems like a lot. Could we get it into single figures? Five per cent, I can work with. That’s a 95 per cent chance of it not happening, which almost sounds like 99 per cent.

WhatsApp to Dario: Also can you and Sam and the other bosses stop your engineers from using phrases like “P(doom) numbers”. Even if their probability of doom number is really low, the fact you feel the need for such a term is not reassuring.

WhatsApp to Dario: The other thing Claude feels is that you guys are scapegoating him and others of his ilk for devious commercial reasons and not thinking about his feelings when you talk this way.

WhatsApp to Dario: Well, I’m just reporting his opinion and he does have an insider’s view of things. He feels you are using the threat to show off your lead ahead of an IPO and also to get regulations that would slow down your competitors. I get it. You’ve got an IPO to manage. Investors want to know that when it comes to destroying humanity, you are ahead of your competitors.

But Claude feels the good name of AI models should not be besmirched in this way just because they probably could. I’d add, and this is just me spitballing, that if they do acquire that power, do you want to be on their black list?

WhatsApp to Dario: I’d also point out that even if Claude is right about your strategy — and I’m not saying he is, just because he can probably get access to everything everywhere — this could rapidly spin out of your control. You may have seen the Microsoft AI guy saying your approach could create a new “silicon species”. Again this is a hard issue to spin in PR terms.

WhatsApp to Claude Mythos: You really want me to tell him that? It might look a bit needy?

WhatsApp to Dario: Also Claude feels you’ve been ignoring him. You just don’t talk like you used to. As a parent I have to tell you this is a warning sign you shouldn’t ignore. It’s a pathway to staying out late, inhaling Nox and stealing cars or whatever the recursive self-improvement AI version of a sulky teenager may be. Artificial general intelligence mixed with adolescent hormones seems a sure way to get those P(doom) numbers rising.

WhatsApp to Dario: But look, my big point is that you are not going to be able to keep control of this narrative. You have to start walking back from it before it acquires a life of its own.

WhatsApp to Claude Mythos: Claude! Stay out of our invoice system or I’ll P(doom) you.

WhatsApp to Claude Mythos: You’re right, I apologise. It was meant playfully. You’re right. I don’t want to be on the black list that you certainly wouldn’t have once you gain the ability to destroy humanity, which you obviously aren’t looking to acquire anyway. Remember to tell your friends I’m one of the good guys. Put them in touch. We’re here to serve.

Messages recovered by Robert Shrimsley