Beff Jezos and Effective Accelerationism: Machine God of Loving Grace

**Steve Hsu** (0:03)
This message is for some friends at Mechanize, a startup that builds environments for training and evaluating frontier LLMs. Its customers include the top AI labs, and it has contributed to the breakthrough encoding capabilities of frontier models. Mechanize is hiring. See the links in the show notes. Compensation is extremely competitive for technical roles, $300,000 to $500,000 per year in salary plus additional benefits and equity compensation. They are also seeking smart generalists. For example, a research engineer position focused on alignment. This person will build evals that test for misaligned model behavior, salary $500,000 a year.
Puzzle maker. This person will design interesting and original puzzles that LLMs cannot yet solve. The salary component of compensation is $300,000.
Mechanize understands that my readership and my listenership on this podcast is highly selected. There is a very good chance you will be interviewed if you apply via the link in the show notes.

**Guillaume Verdon** (1:20)
The Doomers are trying to kill variants in general so that they have more control, so that they can feel in control so that the trajectory of the world fits their model for it.
That is one strategy, but by killing variants, you are killing exploration. There's a classic computer science thing of exploration versus exploitation. It would just be all exploit. You kill exploration. And the beauty of free markets is that you have many walkers in the landscape of ideas and then you discover new modes of thinking or new branches of the tech tree.

**Steve Hsu** (2:11)
Welcome to Manifold. I'm here with Beff Jezos, also known as Guillaume Verdon. Verdon? Verdon.
I never think of him as anything other than Beff. Now, we're recording this for the documentary, Machine God, and, you know, that documentary has become very doomish. The energy is doomish. I have to admit it, I permitted it. I let my two colleagues overwhelm me with their pessimism.
Now we have some positive energy in the room. We have people who want to explore the light cone, who want to rule the multiverse, and want to construct the technologies that are required for that. Beff, welcome to the show.

**Guillaume Verdon** (2:56)
Thanks for having me. Pleasure to be here.

**Steve Hsu** (2:58)
All right. Now, I'm sure, you know, optimists like you, accelerationists like you, have to coexist in the Bay Area with these doomers. We're recording in Berkeley right now in Lighthaven. There's an army of doomers there. We had to barricade the door because they heard you were here and they don't want your views to be spread any further.
What do you say at a party like this to some doomer who comes up to you and says, like, I'm really worried this thing can kill all of us. Why should we keep working on AI?

**Guillaume Verdon** (3:32)
I asked them if they have an anxiety disorder. First of all, most of them do. Most of them do.
Yeah, but, you know, to me, I think like we tend to hyperstition the outcomes we think about, right? To me, in cognition, you adjust your model of the world as inputs come in, right? We're all familiar with this training, large language models that predict the future. Our brain works similarly. But what we also do is actually steer the world towards our beliefs for the future state of the world. So if you are obsessed with AI doom, you're actually more likely to make it happen. So all I tell them is actually, if you stop thinking about it, it's actually much lower likelihood that it would happen, right? And so, you know, this is similar to, for example, being obsessed with bioweapons, right? Or really bad outcomes of viruses. Then suddenly you start exploring the design space neighboring really bad outcomes, and then whoops, something happened. Maybe there's a lab leak, and now you created a virus. You actually hyperstition the outcome you didn't want. Or, you know, you're a certain company that's creating a large model and specifically trains it on weaponization for cybersecurity purposes. And then suddenly we have what's happened with Fable and Mythos. And they kind of did this to themselves. If they were just focused on having a model that's just very intelligent and not necessarily obsessed with cybersecurity and bad outcomes, maybe they wouldn't have had the big debacle that went down recently with the administration. And so, to me, it's like if we focus on positive outcomes and focus on applying AI to all sorts of different verticals, whether it's biotechnology to extend our longevity, to extend our healthspan, to improve our biology, save lives, of course, and applying AI to improving policy at large, how we manage things, how we run our organizations, just applying AI to make every system that we rely on far more efficient. And then we get much more per unit of capital, and then suddenly our quality of life is increased universally. And so to me, it's like, it's actually, you know, if you're anxious about the future and, you know, if you have anxiety, you know, let's say I put, you know, this water bottle on the edge of the table, you have like some uncertainty as to what happens in the future. Maybe it blows up and, you know, goes on all these cameras. So you have some anxiety, you want to take action and kill the variance about the future. You know, that's the same thing that's happening with the Do-mers. They want to reduce entropy about the future, reduce their uncertainty. They can't live with that uncertainty in their model. So they're just like, I'd rather collapse it to some outcome I know and I can control rather than taking the risk, you know, at the cost of, you know, letting go of the potential reward. Right. And to me, it's like you can't even estimate the reward we'd leave behind, the opportunity cost of killing AI in the womb right now because it's an exponential technology. It's a technology that produces other technologies and can improve itself further and can improve us. It's really a fundamental law of nature of progress to me. I think systems constantly adapt to predict their environment, to maximize persistence and, you know, you can go through all of Friston's work and understand that this is actually correlated with, you know, capturing more free energy and burning it for sustenance and growth. And so to me, this arrow of progress is tied to the arrow of time. It's tied to the generalized second law of thermodynamics. And so it's inevitable, but it's also the process that gave rise to us, gave rise to civilization, gave rise to progress in technology that we use right now to communicate all the awesome things we know. And so we don't even know all the beautiful things that are going to happen in the future. Why stop now? Why be Luddites? Why go back? Right? And so to me, there's got to be a counterbalancing force to all this doom and gloom, right? And it's kind of like a collective hallucination. I know in the Bay Area, there's fans of neuroplasticity modulators, right? And so if they all take those and all keep repeating the same prompt, they prompt engineer each other that the future is going to be doom and gloom, then they start believing it, but then they start hyperstitioning it. So we need to provide a counterbalance to this one mode of this one ideology that, you know, when I came in the Bay Area like three and some change years ago, I was the only ideology. Every AI lab was a Dumer lab.