Hunter Fitzpatrick resigns amid fears for the humanities.

Yes. I understand it’s an exceptionally serious situation. Someone has left a job.

That’s what interests me, Jon. People die in a disaster and the editor needs a death toll before deciding how much space it deserves. An AI man hands in his notice and we’re immediately discussing the survival of the species. Extraordinary return on a resignation letter.

Which people?

Right. Here’s how I read that. “The people building AI earnestly”—those people—“believe that it could kill us all.”

Exactly. Not necessarily everyone building AI. The people doing it earnestly. Whoever they are.

And there we are. You’ve got earnest believers. I’ve got earnest builders. Same sentence.

Yes. In my reading, the subject is “the people building AI earnestly”. What do those people do? They believe. In yours, the subject is “the people building AI”, and what do they do? They earnestly believe.

Very efficient. Perhaps it’s already replaced someone.

That’s the problem. Everyone building AI? The earnest builders of AI? His colleagues? People he had lunch with? We’re being asked to take the belief seriously because of who holds it, and we haven’t established who fucking holds it.

Yes, but we need to know which builders before we can decide what their authority amounts to. “The people building AI” is a very large room to speak for.

Very reassuring. I hadn’t asked yet. Thoughtful of him to get ahead of it.

Of course. Anthropic would never derive any promotional benefit from the suggestion that its technology is so powerful it might destroy the species. Nobody could possibly hear that and think, “Fuck me, they must be building something impressive.”

Absolutely. I’m prepared to grant everyone their sincerity. It’s the most abundantly supplied component of the argument.

Close personal friend.

Harbinger Hubinger, we used to call him. Old HaHu. HarbiHubi, if we’d had a few.

What do you think, Jon? What can I say? He’s a glass half empty kind of guy.

Never thought he’d get where he has, to be totally honest, since that first time he declared the end of the world after listening to “Zombie” by the Cranberries. Though admittedly we were all on acid at the time. Ha! College, Jon, what are you gonna do?

Always had that fatalist streak. As I said, good to see some people never change, even after all these years.

At least back then he had the courtesy to blame the Cranberries instead of recursive self-improvement.

That’s professional development, Jon.

And I’m expressing it earnestly.

There you are. Earnestness firmly attached to belief. “We really do earnestly believe.” No chance of it wandering back towards the building now.

Exactly. Coxon gave me earnest builders; Hubinger supplies earnest believers. The Institute takes a close interest in what happens to a statement during transmission.

Exactly. Then Evan steps forward and says, “We really do earnestly believe.” And suddenly the room has an Anthropic sign over the door.

Yes. Now go back to my first reading: the people building AI earnestly. Who are you picturing?

There you fucking go. Nobody said they were the only earnest builders. But they’ve volunteered to embody the earnestness.

That’s the linguistic judo, Jon. You start with a warning about an industry and come away with an impression of one company’s character.

And yet somehow I’m learning how responsible they are. Extraordinary communications work.

So we’ve moved from what an unspecified group believes to one person’s numerical estimate. We’ve established that Evan believes it. How do we establish the probability?

That establishes where he works.

Certainly. A reason to ask him questions. Does it also answer them?

Well, he goes on to say they don’t have a plan to solve alignment for superintelligence and aren’t clearly on track to get one. He’s the alignment-science lead, Jon. He leads it. What does the leading consist of under those circumstances?

Fine. Then tell us how the estimate comes out of that research. What assumptions? What chain of events? What would make him revise it? We’ve got “I personally think” attached to the possible death of everyone. I’d like to see the working.

Please. Apparently it has definitions. Plural. Which fucking one has a greater than ten per cent chance of killing me?

Exactly. What system are we discussing? What can it do? How do we establish that it can do it? “Superintelligence” can’t do the work of an entire explanation.

Oh, well. Put that in the headline.

What I really love about Anthropic is how they combine em dashes with claims of being earnest whilst making unsourced claims about human extinction followed by exclamation marks. It really says something about their character as a company.

He’s announcing it as though we’ve won something. I don’t need him to sound more enthusiastic, Jon. I need him to finish the thought. You believe this. You lead the relevant science. You don’t have a plan. So tomorrow morning, what do you do?

Something I can examine beyond his earnest personal assurance.

I fucking love that, Jon. That’s precisely the sort of research culture I’m looking for. At the Institute we’ve spent months building separate websites so we can cite each other. These people are doing it in the replies.

It’s really important for responsible AI safety companies to show leadership by maintaining a presence and publishing on platforms owned by trillionaires and dominated by the far right. It’s the most responsible information dissemination strategy I’ve seen since my time in sub-Saharan Africa working for Nestlé.

The past is the past, Jon, and I didn’t sign that NDA for free.

Microdosing! They’ve turned taking drugs into another fucking productivity target. You used to take something and lose an afternoon. Now you’re expected to come back with a better spreadsheet.

Even your altered state of consciousness has to justify its salary.

Strict macrodosing.

Whenever the peer review is looking particularly threatening.

It changes our relationship with them. Secondary contextualisation.

Yes. Findings that don’t fit the framework receive further interpretation before entering the evidential record.

We’ve developed an extension. Informally, its working title is the Chinese Whispers approach.

Another round.

We begin with the people building AI and their unsettled relationship with earnestness. Then Evan confirms his own belief and supplies a percentage. In the reporting, killing all humans becomes the end of humanity. We apply further contextualisation and arrive at the end of the humanities.

A small textual adjustment with considerable scope for original research.

That’s our contribution, Jon. We can’t keep expecting Anthropic to do everything.

A 10% baseline probability of the end of the humanities by the end of 11 July 2029.

We felt it was important to preserve the scientific component.

Ours is a conservative baseline.

Coxon said the end of the decade. Hubinger said within the next decade. We’ve resolved that uncertainty by supplying a precise date.

I felt “sometime” would undermine the rigour.

People losing the ability to interpret language. To notice that a word can attach in two places. To ask who a statement refers to. To distinguish evidence that someone believes something from evidence that it will happen.

Oh, certainly. The rich will continue to dwell in their underground sugar caves. I’m concerned about what else survives. Humanism. Values. The capacity to be good. To understand that other people exist even when they haven’t accepted your terms and conditions.

Which is why we’re taking it so seriously.

Look how much has changed in this conversation. The people, the proposition, the deadline. But we still have a percentage and a director willing to stand beside it.

And you’re publishing it.

We’ll cite it as external coverage.

And NEPRAVDA acquires an expert forecast. We give back to those we owe, Jon.

I would have thought my earnestness had settled that.

I mean, really, Jon. This guy, this so-called AI safety “expert”—what a genuine misanthrope. I’m telling you right now: he’s going to regret this move. Who on earth would leave a job in the AI industry? Does he understand what an opportunity he’s throwing away?

He had a job at Anthropic, Jon. He was already inside the building. Whatever his concerns were, he could have continued having them on a very good salary.

Exactly. Still there. Still concerned. You have to respect the consistency.

As I said: misanthrope. And he’s going to miss Anthropic for sure.

I would have appreciated the opportunity.

I take a close interest in their research culture. I have directly relevant experience in secondary contextualisation. I’m comfortable publishing findings whose supporting material is available somewhere else. And I know the Harbinger.

You have to think about your own future as well as everyone else’s.

Read carefully. Follow the references. Ask whether the source supports the claim. Notice when someone moves from what other people believe to what you’re supposed to believe. And consider whether the person speaking might want something.

I’ve thought very seriously about my position. You can’t keep talking about values without eventually making a decision that affects your own life.

And that’s why today I’m announcing my resignation.

Yes.

Because apparently there’s an opening at Anthropic.