Approaching the AI Event Horizon? Part 1, w/ James Zou, Sam Hammond, Shoshannah Tekofsky, @8teAPi

**Nathan Labenz** (0:00)
Hello, and welcome back to The Cognitive Revolution. You're about to hear part one of what turned out to be a four-hour live show that I co-hosted with my friend Prakash, also known as Adapai on Twitter, on the topics of AI for science, geopolitical competition, and recursive self-improvement. With everything moving so quickly in the AI space, I am actively looking for ways to shorten my own personal productivity timelines and to deliver high-quality analysis in more timely and time-efficient ways. And talking to six top-notch guests over the course of four hours is one attempt to do that. In this part one, which we are publishing as a standalone episode, we talk to Professor James Zou of Stanford about his work on AI for science, which ranges from applying interpretability techniques to protein models to building virtual labs of AI agents. To Sam Hammond about how the current US administration is doing on AI policy, what the US is really getting out of its deals with Gulf countries, and why he believes that current AIs are at least as likely as not to be conscious. And finally, to Shoshannah Tekofsky about the many fascinating observations she's made and the lessons she's learned from a deep study of AI agent performance and behavior in the open-ended setting of the AI Village. In part two, which we'll release tomorrow, we talk to Abhimahajan, also known as Owlposting, about AI for Biology and Medicine, Helen Toner about a recent report on automated AI R&D within Frontier Model Developers, and Jeremy Harris about the twin security dilemmas at the heart of the strategic AI landscape. As you'll hear, the challenges of making sense of massive disagreement among leading experts and simply keeping up to date with AI developments broadly come up repeatedly in these conversations. And to be honest, it seems to me that nobody has great solutions. One that I can recommend though is using large language models to help identify your blind spots. And for that purpose, I am really enjoying the blind spot finder recipe that I recently created on Granola. Granola works at the operating system level of your computer, so it can capture all of the audio in and out, including if you wish, the contents of this episode. And its recipe feature can work across sessions to identify trends, opportunities, or blind spots that only become apparent with that zoomed-out view. Obviously, this is a tool that grows in value over time. But if you want to try it, I suggest downloading the app, starting a session while you play this episode, and then asking it to identify blind spots based on this conversation. What is so cool about this feature, at least for active Granola users, is that the blind spots it identifies will be different for you than the ones that it identifies for me. With that said, this episode was a lot of fun, but because it is a new format, I would love your feedback. Do you feel you got as much value from this more time-efficient approach as you usually do from our deep dive episodes? Or did we miss the mark in some way? Please let me know in the comments, or if you prefer by reaching out privately via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. And now I give you The Cognitive Revolution Live from February 11th, co-hosted with Adapai.

**Prakash** (3:19)
We have our first guest, James Zou, adding to the stage.

**Nathan Labenz** (3:26)
Hello, sir. Great to see you.

**Prakash** (3:30)
Thanks for joining us this morning.

**Nathan Labenz** (3:32)
So, quick introduction. We did a full episode not too long ago, and at that time, I was, and I've continued to be, super impressed by your range and productivity in the AI for science domain. When I say range, we're talking all the way from low-level interpretability stuff, which folks can go back and hear about inter-PLM and the work you guys did there to understand what it is that a protein language model is learning. And then on the high end, the virtual lab, a high-level agent framework that was able to do meaningful scientific work and even generate new candidate nanobodies to address new strains of COVID. You've got a bunch of new stuff since then, but maybe just a quick check-in on those previous two projects, both of which I thought were really fascinating. What's happened with them since, if any news? One thing people sometimes worry about is like, well, we thought we maybe understood something based on the interpretability of this, but with time, we maybe realized it wasn't so clear-cut or the agents came up with nanobodies, but did the nanobodies actually work? Are there any new updates or reflections on those previous projects before we get into the latest and greatest?