Don't Make Mirror Life: Synthetic Biologist Kate Adamala on Risks & Responsibility

**Erik Torenberg** (0:00)
Hello, and welcome back to The Cognitive Revolution. Today, my guest is Kate Adamala, Professor of Genetics at the University of Minnesota and synthetic biologist who recently made headlines for coordinating a diverse group of prominent researchers to issue a collective warning against the creation of mirror life. We begin with a discussion of the origins of life and what we can infer from the fact that all of life's vast diversity operates on a single biochemical framework. Then Kate gives an overview of the current state of synthetic biology. Much like AI Interpretability Researchers, biologists today are using a mix of top-down and bottom-up methods to create the simplest possible model organisms, but have yet to create something that is both alive in the functional sense and which we fully understand. If you've never encountered this research before, I am confident that you will find it fascinating. The main reason I invited Kate on the show though, is to discuss her journey from a proponent and active developer of mirror life to someone who is now warning against it. For context, mirror life is a possible form of life, which does not exist on Earth but is clearly physically possible, made of molecules that are the mirror images of life's normal molecules. This relationship, also known as reverse chirality, means that while the chemical properties of the molecules are exactly the same in almost every way, mirror molecules cannot occupy the same space as one another, just as one's left and right hand, even if perfectly identical in every other way, can never occupy the exact same space. This creates exciting possibilities for medical treatments, but also some risk that mirror organisms could be impossible for predators to digest, and thus might prove catastrophic if they were ever released, intentionally or accidentally, into the wild.
Kate was initially excited about the utility and the intrinsically interesting nature of this work, but with time, and particularly after seeing experimental data showing that mirror molecules actually do seem to evade immune systems, she began to reconsider, ultimately concluding that the risks far outweigh any potential benefits. From there, rather than simply abandoning the work quietly, she took the unusual but I think highly admirable step of actively building a coalition of researchers, including some of the biggest names in biology, to publish a warning in science against developing mirror life. Now importantly, this wasn't a binary or permanent decision to abandon synthetic biology entirely, or even necessarily to avoid mirror life forever. Kate continues her work on synthetic cells, and still has hope that mirror molecules can be useful in medicine. She and her co-authors have simply identified one particular branch of the synthetic biology tech tree that they believe humanity would be much better off not exploring, at least until major new evidence comes in. The parallels to AI development should be pretty obvious. Like synthetic biology, AI offers the thrill of discovery to researchers, promises tremendous benefits for the public, and also brings serious and as yet poorly understood risks. And it seems very likely to me that the specific details of the powerful AIs that we develop over the next few years could matter tremendously. Advanced AIs aren't one thing, and neither are they an inseparable bundle. You can create AIs with superhuman coding ability that still don't know how to use a computer. AIs that beat world champions at Go but have no language ability whatsoever. And as recent reinforcement learning work has shown, AIs that solve problems more and more effectively and agentically, but also in more inscrutable and increasingly problematic ways. Given the vast possibility space in front of us, I really hope that AI researchers, particularly at leading companies, but also across academia and startups, make a habit of following Kate's example and regularly stepping back from their work to seriously grapple with its implications and to change course when appropriate.
Should we continue to scale reinforcement learning, given the recent rise in deceptive behavior, which now includes the opportunistic blackmail and whistleblowing that we've seen the anthropic report in CLAWD 4? Should we be racing to turn machine learning research over to such systems, given these behaviors? Should we be developing architectures that internalize models thinking processes, such that we no longer have an explicit chain of thought to examine, or even such that models might begin to communicate with one another in an alien neural ease that we've never understood? And should we, as one well-meaning person who is interested in AI safety, recently emailed me to propose, start using reinforcement learning to train AIs to escape their environments, so that we can hopefully use their successes as indicators of the ways in which we need to harden our defenses? I honestly don't think these questions have simple or obvious answers. One of my mantras is that AI defies all binaries. But given the pace at which AI research is moving, and the fact that even I, as a full-time student and analyst of the field, can no longer keep up with even the things that seem obviously important, it is critical that individual researchers themselves understand that the responsibility is currently on them, both to ask the right questions and to choose which directions to pursue with foresight and wisdom. Fatalist rationalization that someone will do it if we don't simply doesn't cut it in today's world. Kate's work proves that scientific communities can indeed coordinate to identify and avoid particularly dangerous research directions while still aiming for transformative progress. For those of us involved in AI development, this offers both an inspirational model and an important challenge. The question isn't whether to develop AI at all. I do agree with those who say that the upside is too great to pass up and that in any case the cat's out of the bag. But still, we can and should hold ourselves to the highest possible standard when it comes to assessing the risks and being willing to back off and change course when necessary. As always, if you're finding value in the show, we'd appreciate it if you'd share it with friends, post about it online, or leave a review on Apple Podcasts or Spotify. We always welcome your feedback too either via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. Finally, I was honored to learn that we were voted the number three AI podcast at Swix's AI Engineer World's Fair this week. I'm really very glad to know that this show has proven a valuable resource to such a plugged-in group, and I hope that our mix of technical deep dives, broad surveys, and occasional moralizing lectures can nudge AI development in a positive direction, however slightly. In any case, a big thank you to everyone who voted and everyone who listens. Now, I hope you enjoy this fascinating exploration of synthetic biology, existential risk, and the great responsibility that comes with truly transformative research with Professor of Genetics and Synthetic Biologist, Kate Adamala.