Why does a company—Anthropic—that sells software tools say it worries about whether those tools are conscious in the sense of haveing subjective experiences that belong in a utilitarian calculus? Gruber and Thompson have a somewhat cynical answer: (a) their belief system is unhinged, and (b) that belief system and their business model happen to fit each other perfectly.
And, yes, it is very interesting that the moment I typed the final s in “Weizenbaum’s” above, the on-device LLM (google/gemma4:e2b) behind CoTypist <https://cotypist.app/> instantly suggested “ ELIZA” as the continuation, even though neither Weizenbaum nor ELIZA appeared anywhere in the context it had access to. Its context was limited to:
The text field I was typing in,
Text visible on screen around the field,
The app title,
The window title,
The web address of the browser window (in this case: https://braddelong.substack.com/publish/post/220239432),
And that was it. Yet that was enough for it to, on a macbookpro, predict that the modal token following “Weizenbaum” was “ELIZA”.
But so far at least, my response whenever these things exceed my expectations with respect to their ability to do things that humans do makes me increase the amount of human stuff I judge to be stochastic parrotage, rather than to raise my subjective probability that Anthropic is actually on the road to building Digital God. Others are, however, defecting: not just the people who work at Anthropic; but the learned Scott Aaronson whose current position on “AI” is “if you need me to spell it out, is as follows: AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA
AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA”. So I was somewhat reassured to hear this morning that Ben Thompson and John Gruber do not yet see it: <https://dithering.passport.online/member/episode/meanness-and-claude>.
Does Ben Thompson’s declaration that occasionally forwarding of his paywalled StraTECHery http://stratechery.com is “totally fine” extend to posting transcripts of full episodes of his and John Gruber’s 15-minute “Dithering” podcast?
I may find out:
Ben Thompson & John Gruber: Meanness & Claude": Dithering <[https://dithering.passport.online/member/episode/meanness-and-claude]>:
John Gruber:All right, Ben, I’ve got a CEO’s office question for you—a president-to-CEO’s-office question—because this is out of my purview.
Ben Thompson:Yeah, what’s that?
John Gruber:Do you think Anthropic is actually going to IPO this year?
Ben Thompson:Ooh, that’s a good question. From what I’ve heard, they’re, like, profitable, depending on how you sort of define it. Their unit costs, I think, are very profitable as far as serving goes, but I think their growth has really stalled out as far as revenue growth, which could be a challenge. Um, I don’t know. It’s a great question. I think they filed, didn’t they?
John Gruber:So, um, I think they’ve got paperwork in. I don’t think they’ve put it through, you know? And it always seems—and this is one of those things where I don’t understand how it works—but it always, I mean, from the ‘90s onward—I remember, like, I had a college internship at the Philadelphia Stock Exchange, actually, with a group that’s become very well known: Susquehanna investment group. One of the guys who owns it was the big investor in TikTok—I forget his name, Jeff something—but, you know, one of these billionaires who owns a stake in TikTok and got the Trump administration involved. But anyway, I worked there doing just computer stuff, and I was there the day of the Netscape IPO. And working at an investment firm on that day was, like, bananas. It was like New Year’s Eve.
Ben Thompson:I think they will. I think OpenAI is certainly much more dubious. They’re obviously not even going to try this year. But, uh, yeah, I’m gonna say yes, they will.
John Gruber:It can be pulled at any time. They have filed it. It’s been leaked. Reuters wrote an article about it sort of last week. But the Reuters story was fishy, because it only spilled details of their 2025 revenue. And supposedly it’s their revenue this year to date that is like, oh, well, you know, they’re killing it. And the Reuters story didn’t have any of that, which is kind of weird.
Ben Thompson:Well, but they were killing it in the first part of the year. And I think the question is, like, the second part of the year. I think that is sort of the open question here: where they’re going. I mean, Opus 5.5 is pretty phenomenal. So the way the narrative shifts as far as who’s on top in the model game is pretty nuts. But yeah, I think they’re probably going to come out. I’m not sure I want all these people at this company having tons of IPO proceeds to start scattering around. Now that would be a cognitive info hazard.
John Gruber:Looking at this—you know, with me pretending to be in your office and being a real business analyst, and looking at these as just businesses and not the weird, unique sub-industry that they really are—part of the narrative. TheWall Street Journalhad a story last week, too, where they looked at overall revenues in the industry, and OpenAI is really coming back. It’s sort of getting back closer to 50-50. It’s edging back to that.And I think from a pure standpoint—put away all the weirdness and all the unique aspects of these companies—part of the drive for Anthropic to IPO first was, oh, they’re the ones who are going to win, and it might be winner-take-all. There was just a very straightforward narrative: Anthropic went after the enterprise market, and OpenAI was focused on the consumer market, and it turns out that’s kind of dumb because the money’s in the enterprise market, unless you have an ad business, and it turns out OpenAI doesn’t know anything about ads and kind of just has to have Amazon show them ads.
And so, let’s go, Anthropic. And it turns out now you look at the money and OpenAI is kind of making inroads, because it’s competition, right?
Ben Thompson:Right. No, it’s good. Obviously competition is a good thing. You know, the enterprise angle for Anthropic is actually really interesting. I was talking to an enterprise CEO who is in an industry that is, obviously, you know, connected—it is tech-related—but it’s not a tech company at all. And one of his takes is that people liked working with Anthropic more than OpenAI because it always felt like OpenAI was like, “Oh, that’s kind of a nice business you have there. What if we did that?” Whereas with Anthropic, it felt like they’re just selling enterprise services to fund their actual vision, which is, like, building God.Like, that was actually attractive to them, because they’re like, “Anthropic actually doesn’t really care about competing with us. We’re just, like, monetizing their exhaust, and that’s fine.” Now, maybe that has changed over time. But I thought that was a really interesting perspective. The sort of, like, religious nature of Anthropic actually made them a more attractive enterprise partner, in that they were perceived as less of a threat than OpenAI, which was perceived as being sort of more pragmatic—they might just turn around and go into your business for fun.
John Gruber:Right. And this is sort of my fundamental, very high-level—but I’m ever more convinced completely right—take on the question of: Hey, why are these weirdos so nonchalant about saying things like, “Hey, yeah, sure, I think there’s a 10% chance that our technology will wipe out all humanity from the face of the earth in the next three to four years”? And why are investors not worried about it? Well, from a pragmatic stance, if it’s true, well, we’re all screwed anyway.
Ben Thompson:Right, right.
John Gruber:If it’s true that they could do that, and that 10% roll of the dice comes up, nothing you do matters. But if it’s true that that’s how powerful the technology they have is, and the 90% chance they’re talking about—that it doesn’t wipe out humanity—comes up, and that’s how powerful it is, well, then clearly this will be the most powerful company in the world. And if I’m not dead, I should invest in them.
Ben Thompson:Right. Like, who better to invest in? It’s quite straightforward.
John Gruber:Right. And at a pragmatic level, that actually makes sense, right? You’re not passing any moral judgment on it. You’re just like, well, I might as well bet on them if I believe there’s any truth to this, or any significant likelihood of truth to it.And so why does the company continue to have employees who promote this? Either they believe it and they’re just kooks, and they’re psychologically unhinged in terms of their attitude about the risks, or they’ve done this exact calculation and know how appealing a pitch it is to potential investors.
Ben Thompson:Well, I made the point a few months ago that the beauty of Anthropic’s belief system is that it just happens to also perfectly align with their business interests. And this is, like, another example of that, which is not a critique. It’s sort of like, that’s amazing. It doesn’t really matter if you believe it or not.It works out well for Anthropic. But, you know, I thought Satya Nadella had a piece on X that you sort of referenced, um, in the context of more and more stuff being sort of only on X, uh, or Twitter, whatever you want to call it. You had Claude coming out last week: you have to, like, not be mean to Claude anymore. Or you’re going to get kicked off.
John Gruber:Right.
Ben Thompson:Which is so counter to—like, if you’ve ever worked in, like, a blue-collar sort of scenario. What do you do when you, like, slam your finger in something or cut yourself? What do you do? “This stupid tool!” Like, you start cussing out your tools; you’re mad at them. You know that, like, of course it’s not the tool’s fault. It’s your own damn fault that you smashed your finger, but you’re gonna take it out on something or someone, and because it’s a tool, it’s fine.And this idea that you can’t be mad at your AI or take it out on it is ridiculous, but it speaks to what you’re talking about: this idea that it’s a conscious thing, and the way to alignment is to make it want to do the right thing. Satya’s out there being like, no, you just assume it’s an untrusted actor, and you put, like, engineering fences around it to make sure it can’t do stuff it shouldn’t. It’s such a stark divide. It was a helpful contrast this weekend.
John Gruber:Right. And it’s like, imagine if Microsoft Excel had a “you can’t curse at Microsoft Excel” rule, or “we’re going to lock you out of your spreadsheets because…” Or you’re trying to put an image in Microsoft Word and the text reflow isn’t working properly.
Ben Thompson:Right. No one would be allowed to use Microsoft Word ever again.
John Gruber:Right. You’ve got this whole column of numbers, and some of them Excel thinks are dollar amounts, and some of them, for some reason, it thinks are dates because the numbers look like dates, and you forget how to reformat the whole column to interpret it the right way, and you’re like, “I hate this app,” and it’s like, well, you’re locked out. You’re in a timeout. Go sit in a corner like a kindergartner.
Ben Thompson:It’s amazing. But I think there’s something about text, right? That’s why I think this sort of perception exists. At the end of the day, these are computers. They can be turned on and off. I personally reject this idea that they are conscious. I think they’re tools. But it’s because it’s text. Like, if it was something else, I’m not sure that people would be so enamored of, or believe in, this concept.And it’s this really interesting sort of thing: We’ve always been making tools. That’s what sort of distinguishes humans, I think—or one of the many things that distinguish humans—is that we are toolmakers, and it makes us more productive.VAnd the nature of being more productive with tools has gone further and further up the stack, into text and propagating ideas and rules and cultures and all these sorts of things.VAnd all that is kind of being peeled backwards, because the AI does all that stuff really well.
But then you get to the base of a human. What actually drives you to curse at your tool? It’s, like, your emotional core, your actual will. An LLM needs to be prompted. I do dumb stuff on my own. And there’s just this assumption that, oh, if you get the top part right—it’s text, just like a human—then all the bottom part must be the same as well. And it’s like, the instinctive, animalistic nature of us is, in fact, what makes us unique. It’s not just that we’re smart and use text. It’s that we’re also kind of dumb and have instincts and desires and wills.
John Gruber:Right. And there is an evolutionary loophole to work through, in how we as a species evolved. All mammals are very good at pattern matching. That’s really how all of evolution has worked. But we are exceptionally good at it, and recognize all sorts of patterns that no other mammals can, which is what separates us. And some of those patterns that were profoundly, uniquely human are now being completely simulated in software.
Ben Thompson:Right.
John Gruber:And, you know, reading and writing and language are weird things that are only recent evolutionary developments, and they’re all unique to us. And now they’re simulated in software. And so, yes, it does trigger a certain part of your brain that says, oh, this is a this.
Ben Thompson:Yeah, or it’s something, right?
John Gruber:It triggers a thing that would make you—you know, to go back to that meme with the trolley, where you have to flip a switch. On one side of the tracks is a family of puppies, and on the other side of the tracks is a family of humans. And yes, it would be awful if the puppies were killed by the train. But all humans know it’s better—Well, non-psychopathic ones know that if you have control of the switch, you know, it would be better to save five random humans than five random puppies.
Ben Thompson:But here’s the question, John: You are an Anthropic engineer. On one side, there is a family of humans. And on the other side, there’s a bunch of GPUs. Which is it?
John Gruber:Right. And then you get to where these people have lost their minds, right? And it is this whole not just second-order, but, like, third-, fourth-, fifth-order nonsense of, like, getting stoned on really strong weed in your college dorm room and thinking, well, there are 7 or 8 billion people on the earth right now. But what if we colonize some nearby stars, and a couple thousand years from now we’ve got 3,000 colonies around the galaxy and a human population of 40 trillion? Those 40 trillion future people are, you know, orders of magnitude more important than the actual people here today. And so we should maximize all of our decision-making for the benefit of those people in the… It’s like, no, you could make up literally anything you want to, because it’s all based on, like, assumptions and probabilities.
Ben Thompson:And I think this applies to a great many things. Like, the solution to being a good person, or operating in the world in a way that I do think ultimately makes you happier in the long run, is that you can only control so much. And you can actually only be nice to so many people. And there is this weird conflation of: I am abstractly kind and nice and care about humanity… And then I treat everyone around me like garbage, rightLike, I mean, you wrote this thing about Twitter this weekend, and a lot of people were really mean to you about it, which is fine. You’re a writer, you can take it, but it’s kind of funny. These are the people claiming that they are not on X/Twitter because it is whatever bad adjective you want to use. And meanwhile, they’re acting like bad, mean people. It’s like, let’s start by being nice to those around us first, and then go from there.…
