Dialogue on Viriditas

Introduction #

A few months ago, Elliot Hershberg, one of my favorite Substack writers, wrote a post laying out his personal mission. I really enjoyed his post and, as is customary in blogging, the best way I could think of to compliment it was to share some (hopefully constructively) critical thoughts on it.

After sending my post to Elliot, he proposed the great idea of us having a discussion of this over email, which we could then potentially publish. The Follow-up section here contains a lightly edited form of this back-and-forth. This back-and-forth helped me clarify my views and moved me more in the direction of Elliot’s views than I started out. As a result of this and other discussions/observations I’ve made since then, I feel a lot more uncertain than the views and tone expressed in the original post sound.

Finally, big thanks to Elliot for engaging with me in this discussion. It was a lot of fun!

Original Post #

Before I do jump in, I want to note that Elliot’s post really has two high-level messages, one about biology in the future and one about growth being good in general. I agree so strongly with the latter that I entirely focus on the former, which I partially disagree with, in this post.

Elliot summarizes his mission statement as follows:

I want to share my own personal mission statement: I want life to flourish in the universe. I view biotechnology as the most logical means towards this end. When I say life, I mean the process that has carpeted our planet in cells, flowers, and children. Life is an abundant, beautiful, generative process. I’m talking about viriditas. The “constant pressure, pushing toward pattern. A tendency in matter to evolve into ever more complex forms. It’s a kind of pattern gravity, a holy greening power we call viriditas, and it is the driving force in the cosmos. Life, you see.” More broadly, I’m arguing for the beautiful imperfections of carbon-based corporeal life in all of its various forms. I want biological life to flourish on our home planet, and I want life to spread outwards into the Cosmos. I want more bugs, babies, forests, and trees in the universe. I’m arguing against the future in which “substrate independence” lets us fill the world with digital minds. I’m making an aesthetic argument for substrate preference: a deep desire to propagate carbon-based life.

I want to start by emphasizing that I am really drawn to this vibe. There’s a reason I not only work in biotech but also have thoroughly scoured the internet for biotech-related sci-fi (ask me for my list if you’re curious) and read all of it. It’s because, like Elliot, I really like the aesthetic of biology as a technology. In Elliot’s words:

Biology is the only functional nanotechnology that we know of. Life has evolved sophisticated machinery to thrive in every corner of the globe. These programmable machines are capable of converting matter into arbitrary structures with atomic precision.

Similar to Elliot, I definitely find myself more impressed by and drawn to biological machines than to mechanical ones, even when the latter may be more efficient. For example, as much as it’s unrealistic and probably would have all sorts of practical problems, I continue to love the idea of growing houses. I also wrote an entire post about cool-sounding futuristic synthetic biology ideas.

Unfortunately, I’m not as optimistic as Elliot about the prospects for Viriditas relative to the virtual, AGI (artificial general intelligence) driven vision he contrasts Viriditas against in his post. While I am bullish on biotechnology in the short to medium term, in the long term if we don’t destroy ourselves I expect that computational agents (human emulations, artificial intelligences, etc.) will determine most important outcomes in our solar system, galaxy, and even lightcone. The rest of this post lays out my current views why and discusses some potential objections and caveats. Note that these views are provisional and heavily based on intuition, so I welcome pushback and critique even more than usual.

Computational life (either emulated humans or artificial intelligences) starts out with many advantages relative to biological life. As discussed by Holden Karnofsky in this post in his digital people series, computational life forms can copy themselves, cheaply experiment on modified versions of themselves, and speed themselves up or boost themselves by buying more compute. Biological life has versions of these things but they tend to be slower and less flexible. Human reproduction is a miracle and I am very pro more humans, but relative to making a copy, sexual reproduction is slow, unpredictable, and relies on less fungible resources. Human learning is amazing but our self-modification and introspective abilities pale in comparison to something running on a computational substrate where you have full white box access. Plus, our methods for imparting our learning to the next generation are inefficient and depressingly ineffective.

This isn’t even factoring space into the equation. Making humans more radiation resistant and generally adaptable for space is awesome and I would love to see it get more resources. But even if we figure it out, it will still be an uphill climb relative to sending computational life into space. For long trips, we need to either solve hibernation or solve life extension and/or gamble on societies of humans being able to survive in generation ships for 100s or 1000s of years without falling prey to all the petty squabbles humans tend to get up to. For computational life, we already have a lot of the hardware worked out for “short” (only several decades) trips and figuring out how to radiation shield some computers combined with error correction is likely much easier than creating a stable society inside a confined space to last generations or solving hibernation for 100s of years. (Just imagine the IRB for the experiments!)

In his post Elliot also makes the argument that biology is currently the only nanotech that works, implying that if we want to transform planets/the physical world, we need biology. I don’t find this convincing for a few reasons. First, using biology as a tool is very different from biological beings controlling the trajectory of the future. For example, a bunch of uploaded humans/AIs using super-bacteria to terraform planets may fit with the aesthetic of Viriditas but seems pretty different from a Viriditas-dominated future. Second, the biological nanotech advantage may only be temporary. While I’m not informed enough to have an opinion on the decades of nanotech debates, I’d be surprised if some sort of more advanced non-biological nanotech never worked. What this looks like and the degree to which it’s carbon-based and has other biological features is not something I can predict, but I predict whatever it is will be largely designed rather than evolved.

While I’ve so far focused on object level points, I suspect there’s a latent crux about how “good” a designer biological evolution is relative to intelligent life generating a lot of these disagreements. I think evolution is amazing and counter-intuitively powerful but ultimately subordinate to design via human and future intelligences, if for no other reason than design can take advantage of evolution’s tools in addition to less brute force ones. I get the sense that Elliot and others like him tend to view biological evolution as powerful enough to continue to play a primary role in the trajectory of Earth and other planets. If this is truly a crux, I’d be interested in hashing out whether we can find predictions where our respective models differ based on it!

One thing that might change my mind on the most likely outcome would be if humanity seemed serious about improving itself biologically while biological life is still ahead. Given that part of my view on why computational intelligences will control the future is that computational intelligences will be able to iterate on themselves much faster than biological intelligences, I can imagine a world where biological life maintains its edge by getting itself to a level of capability where it can play a major role in the trajectory of the future. For better or worse, we do not live in this world. While there are good reasons for not experimenting as wildly with humans as we currently do with AIs, our culture and regulators (although I think this is more complicated than people on the internet make it out to be) are risk averse even relative to what I believe would be optimal. This risk aversion goes beyond expected value calculation or even just not wanting to harm people and instead bottoms out in a philosophical view that liberating humanity from some of its evolved biological constraints is a “dangerous idea”. Combined with cultural taboos on anything besides restoring “normal” health, this means that actually figuring out how to make humans more capable is relegated to backwater or black market research. Even just going into space has become incredibly controversial because “only the government should do it”/“they want to abandon Earth”/etc.!

Put a different way, I don’t think biotechnology is playing for keeps when it comes to competing with future computational life. Whereas a subset of AI people are laser and openly focused on building AGI, biotechnology lacks a unifying narrative beyond curing disease and is stridently against endorsing anything that remotely looks like “enhancement”, with the exception (that proves the rule) of one famously quirky professor and founders of companies who receive a disproportionate amount of scorn given their size. A surprising number of political elites also consistently listen to bio-conservatives like Leon Kass who fight tooth and nail against biological progress (including IVF by the way). Or for another example, see the reaction to mRNA vaccines. Given that and assuming we don’t destroy ourselves first, I expect progress towards AI to continue to outpace the relevant forms of biological progress. (This also may not be a bad thing! I don’t actually think it has to be a competition and my hope is that whatever computational life will be more capable and more benevolent.) In case it’s not clear, I’d love for our society’s views on this to change and think it improves the likely outcome of the future independent of the substrate agents eventually run on, but I’m currently pessimistic.

This is over-simplifying but I fear that arguing for biological life controlling the future is kind of like a horse breeder arguing for horses over cars a decade before the spread of the Model T. They could point to the current superiority and inherent dignity of horses. They could emphasize how much faster and obedient horses had become from breeding. But ultimately reality didn’t care about their arguments.

Finally, if you buy all the above arguments but still think computational life won’t control the stars in the long term, I suspect you either think human coordination against AGI/emulation will prevent it or that computational life is impossible. I’m not convinced by either of these but the reasons why are much too long a discussion for this post and are covered elsewhere in much more detail.

To summarize, I love the vision of Viriditas but think humanity’s most likely current trajectory does not involve the long term future being controlled by biological life. This essay was a cursory attempt to lay out some of my arguments and intuitions for why I think this and how my views could be changed, so I again urge readers to critique them. If you do find my arguments compelling, I also want to plug that I think a computational life controlled future could range from amazing to horrible (e.g. involving the extinction of anything lifelike that shares our values) and that influencing that outcome seems important.

Follow-up: Dialogue between Elliot and I #

EH foreword: I first want to say thank you to Stephen for taking the time to read my post and dive deeper into different parts of the argument. A major part of the goal of writing posts like Viriditas is to spark constructive conversation about the type of future we want to build. This type of critical analysis is a valuable contribution that leads to interesting points of clarification, agreement, and disagreement.

On narratives and the future: How much do the narratives we coordinate around influence the path of technological progress? #

EH: This is a really central part of my argument and my goal in writing. While Yuval Harari is occasionally derided for painting too broad of strokes in his high-level analysis of human history, his core argument really resonates with me: our species organizes around stories. History has many broad lessons. It teaches us that everything changes, and also that many central aspects of our daily lives are built on top of fairly arbitrary choices made in the past. The choices we make today will have the same impact on future generations. Technology is no different. One example is the choice to focus so much capital and intellectual effort on the development of computers—both hardware and software.

For roughly 70 years, this one branch of technology attracted many of the world’s most brilliant scientists and a massive amount of investment. It’s hard to pin down why, because the answer is probably multi-faceted: Thiel has argued this in part due to our regulation and stagnation in “the world of atoms”, it was also an area of science where rapid progress was possible. I think an important part of this technological path was also storytelling. Once some of the crucial infrastructure was in place, the Personal Computer required visionaries like Steve Jobs to really make the transition from the world of hobbyists to a universal consumer product.

This transition had enormous consequences, and gave birth to entirely new markets. You don’t get the Web, Mobile, or Cloud transitions without it. Here’s a point I feel very strongly about: it’s easy to imagine a counterfactual history where we simply didn’t dedicate this amount of focus to computers and instead explored other technologies. We happen to live in a world which followed this particular arbitrary path of innovation.

I think taking this view of history gives us a lot of agency. It means we can think really hard about which particular directions we want to pursue, and aren’t purely bound to following specific paths just because they are what has been previously done.

I think that we are at an interesting historical crossroads: there are many arguments about what we should do with the incredible power of computation that we’ve created. Should we devote most of our time and capital to AI, virtual reality, Web3, and other purely digital pursuits? I would argue that we should use computation as a powerful tool to solve tangible problems in the physical world: curing diseases, learning how to generate new materials and structures, exploring space.

I think it’s also important to really highlight the absolutely immense generative potential of biology (especially accelerated by computation) to solve some of these foundational problems. Part of my argument in Viriditas is that we need to remember this. This is a strong personal bias: I think we are incredibly far away from being able to produce systems that even resemble 1/10th of the energy efficiency, durability, adaptability, and ability to precisely manipulate physical matter that cells already possess.

SM: To start, I agree that human progress depends heavily on narrative and stories. I also agree that the effort and resulting progress in computing was not predetermined. In particular, I think Moore’s Law was a brilliant narrative, which enabled amazing coordination across the industry and over time.

I will also go further than you and say that right now biology mostly runs on bad narratives. While eliminating and curing disease is an amazing feat that humanity should be incredibly proud of, the reality is that it’s mostly a negative vision. But for biology to really become core to a societal narrative, we need a positive vision of what we’ll create and do when we’ve eliminated all disease. That’s why I love your Viriditas piece even if I’m not as convinced by it!

I do suspect we disagree at least a bit on how malleable the path of the future is though. It seems to me like digital technology comes with certain core advantages which would make it hard for biological technology to keep up if the two were to reach parity. The ability to faithfully copy data and the decoupling of substrate and function are two such examples of advantages I see digital technology and intelligence having.

Now to be fair, you ended with the important and interesting point that right now biological life is ahead in terms of its ability to manipulate matter and create intelligence. I agree with this but (based on our conversation) think we just differ in terms of how much we think digital intelligence (AI) is on track to close this gap.

One other counter-argument to the above I can imagine you making is that the advantages I mentioned aren’t intrinsic but instead a result of greater investment. Perhaps in the future we can figure out how to pass memories between biological intelligence directly, make true “copies” of individuals, protect life from the ravages of space, etc. As I mentioned in my initial post, I want to believe this is possible but feel that the current life sciences ecosystem has for the most part taken on some ideological precommitments that prevent it from really targeting the sorts of things that I believe would be required to keep up with digital life on the time scales that matter. This brings things full circle in the sense that this highlights the importance of narrative and shifting the current one.

EH: This is all really interesting. It sounds like we fundamentally agree on the centrality of narrative. The primary difference in our viewpoints seems to be on how much we believe in the narrative about the potential future of computing.

Let’s take Ray Kurzweil’s Singularity thesis as the most extreme possible case for the potential of digital technology. In his view, computing will continue to exponentially scale indefinitely. As this happens, the possibilities are literally unimaginable—hence the singularity, which means an event horizon. Perfectly simulating biology, AGI, nanotechnology, you name it, all enabled by computation.

There are parts of this view I agree with, and parts where I get off of the train. I was trained as a computer scientist, and currently do computational biology research. I deeply value the power of Universal Turing Machines, and am even sympathetic to the notion—like Wolfram, Kurzweil, and others—that the universe itself is fundamentally computational. I think that computation is one of the most powerful tools for natural science. That’s why I’m a computational biologist!

So where do I fundamentally disagree? When I look at the best hardware and software we have, I still find profound inefficiency relative to biology. Let’s think about a concrete example: GPT-3 from OpenAI. These models were trained by melting the cores of thousands of GPUs in giant server rooms! It literally would cost $12 million to train GPT-3 on AWS without any discounts. Think about that! Even these models that seem like magic boil down to ingesting enormous amounts of data, and use programming 2.0 techniques (differential programming to learn programs from data instead of specifying rules) and enormous amounts of hardware. Our state-of-the-art models built by the smartest scientists are sets of static weights that require server rooms to train.

I won’t even compare that to the efficiency of human brains, that’s already been said before! Even prokaryotic single-cell organisms are doing massive amounts of efficient and compact computation relative to that.

So far, this is purely accounting for efficiency of performing computations. When thinking about programming matters, the contrast is even more stark. To quote Drew Endy, biology “is the type of material that is all over the planet, for the most part, and wherever it is, it harvests local materials and energy and makes copies of itself.” It can encode the ability to produce arbitrary shapes, and is capable of atomic precision. Again, reflecting on our current manufacturing technology, it seems like we are incredibly far away from rivaling this.

So in a nutshell: I think our narrative of bits overhypes its future potential to transform the physical world, while our narrative about biology massively undervalues what living systems are already doing.

SM: I want to zero in on your claim about efficiency because I suspect this gets at the core of our disagreement.

So where do I fundamentally disagree? When I look at the best hardware and software we have, I still find profound inefficiency relative to biology. Let’s think about a concrete example: GPT-3 from OpenAI. These models were trained by melting the cores of thousands of GPUs in giant server rooms! It literally would cost $12 million to train GPT-3 on AWS without any discounts. Think about that! Even these models that seem like magic boil down to ingesting enormous amounts of data, and use programming 2.0 techniques (differential programming to learn programs from data instead of specifying rules) and enormous amounts of hardware. Our state-of-the-art models built by the smartest scientists are sets of static weights that require server rooms to train.

To summarize my view: you’re looking at the static snapshot rather than the time series. Just the other day, someone shared how you can now train GPT3 for $500k by using various tricks and algorithmic improvements. That 24-fold decrease in cost happened in a little under 2.5 years. More generally, the cost of training an AI system to a given level of performance has historically fallen precipitously over time. To me, this highlights how much of an advantage fast feedback loops provide digital systems over (current) biology. The ability to have an idea and test it quickly enables rapid improvement that can then be built upon by others. On top of that, AI systems can be much more easily poked, prodded, and ablated than animals, and humans in particular. Because of these factors, I expect the progress in AI (including towards efficiency) to outpace the rate of progress in biology unless we can remove the bottlenecks in the latter such as safely testing in humans.

It’s possible your response will be that (as you’ve written about) biological tools follow their own impressive cost curves. Unfortunately, I don’t think the analogy holds. While the progress in sequencing/synthesis costs are impressive and have driven amazing progress in the life sciences, the cost of biological applications seems to mostly be bottlenecked on testing in the relevant organisms. I can imagine this maybe following similar curves for simple organisms (so maybe optimism about biomanufacturing is warranted) but for therapeutics that are targeting humans, the clinical trials bottleneck seems to be, if anything, getting worse over time.

Another point I’d make here is that there’s a risk of assuming that in order to be as impactful as humans, AI systems have to achieve parity to humans with respect to every characteristic. To me, this seems similar to a hypothetical skeptic of machine flight claiming that we can’t have planes until we find ways to make them as light and as flexible as birds. In the case of your argument, I think the analogy to humans leads us astray because it treats training as analogous to human development. This analogy ignores the fact that once I’ve trained GPT3 once, I can copy it 1000s of times on the cheap and then deploy it for a fraction of the cost it took to train it. I suspect there are other ways in which these direct comparisons lead us astray.

All that said, I’m much more sympathetic to your point about programming matter. To date, AI has not been particularly impactful in the realm of atoms. While I believe robotics may finally be having its ImageNet moment, I agree that there’s no evidence for AI even coming close to biology’s ability to influence the physical world either via manipulation or through designing materials/molecules/organisms yet. Similar to my argument above, I suspect that this focus on current disparities may ignore the potential advantages AI systems have in the long term, but I don’t have the evidence to argue this convincingly.

Bringing this back to the higher level point, I agree that our narratives around biology undervalue its power and capabilities. I suspect we have a hard-to-resolve disagreement about how quickly AI could overtake biology’s current capabilities if it continues to progress at its current pace, but I doubt we will resolve this via back-and-forth. I think if we wanted to, the best way would be for us both to suggest signs of faster/slower progress than the other suspects in the two respective areas and see whether we actually disagree on their likelihood.

And then segueing into the next part of the discussion, another place we may disagree is a bit more practical. I think humanity currently lacks an optimistic vision of harnessing biology’s capabilities to influence our future trajectory. Most of the population is suspicious of biotechnological progress, and even many life sciences researchers seem to view changing humanity or the biosphere too much with suspicion. To draw a contrast to AI, a large fraction of AI researchers seem to have coalesced around a narrative that praises rapid transformative progress, whereas our narrative for biotechnological progress seems overly focused on a limiting paradigm in which the pinnacle of achievement is returning humans to a “healthy” state. So, as our next topic for discussion, I’d like to pose the question of whether you agree with my (overly generalized) assessment of the current state of the narrative in life sciences and, if so, how you think about both the prospects and strategy for shaping/influencing it (assuming you think that’s a good thing to do).

EH: I agree that it sounds a bit contrarian right now to doubt AI progress right now given the pace of breakthroughs currently. AI progress is really impressive! My point with Viriditas is that I don’t see all of this directly producing a future people actually want. I think that it will be deceptively hard to translate these breakthroughs into problems in the physical world. It’s important to continue to make progress in the digital world, but I have pretty strong conviction that biotechnology is one of most direct paths towards achieving physical abundance.

I would agree with your general analysis of the lack of narrative within biotech. Right now, biotech is largely synonymous with biopharma. Part of my motivation to write Viriditas is to help with this—as biotechnologists we need to think bigger! The type of technology we’re developing can have planetary scale impact!

SM:

I think that it will be deceptively hard to translate these breakthroughs into problems in the physical world.

I agree that this looks fairly likely right now and is bad! That’s part of the reason I work on what I do, but that’s a bit of a tangent.

The type of technology we’re developing can have planetary scale impact!

I agree with this, although I’m less sure what strategy beyond changing the narrative can drive it. It feels like right now we are quite far away. Biology is mostly done by a small group of people and is finicky to get to do even basic things.

Short follow-up on promising directions for progress #

SM: Adding on to my prior message, a direction I’d like to take this is to discuss what shifting the narrative looks like and potentially what types of progress could come from that in the near to mid term. In my mind, one underrated component of creating the future we seem to both want is getting the biology toolkit to a place where people can hack on biology outside of the academic setting. I have the maybe heterodox belief that a lot of innovation gets unlocked when a technology becomes cheap enough to tinker with and cycle times get fast enough to run lots of tests. Biology’s not there yet but I believe it could be (for many areas) in the next 10 years with concerted effort by researchers. This is only possible though if more research focuses on simplifying, reducing costs, and increasing accessibility of core tools. One of my favorite recent examples of this is this paper on replacing thermocycling with helicases in PCR. I’m told that this is probably finicky still but the idea of making PCR simpler to run is a great example of the sort of progress I want to see.

Does this make sense to you? Do you have other thoughts on how to drive progress towards this future?

EH: This is a really important point, and I think that I agree with your emphasis on the value of tinkering. This was part of the point of a recent essay I wrote called Atoms are Local where I argue for the value of more local tools and infrastructure to make this type of tinkering and production possible. The paper you reference is a great example of engineering our fundamental tools to be more portable and resilient.

I think that part of this is a renewed focus on the power of biology, and the fact that compounding progress on this particular type of biotechnology could reshape the world in really interesting ways. This type of very personal bio-toolbox (or Personal Biomaker) could literally reinvent medicine and a huge portion of industrial manufacturing. Part of shifting the narrative is to outline the broad vision, but we also need early wins that make biotech much personal for people.

Are there other examples of this shift that you’re particularly excited about?

SM: Yep, I loved your recent essay!

In terms of other examples, how we think about treating and engineering our own biology is an area in which I believe massive improvements can be made although the nontechnical barriers may be bigger than in biomanufacturing. In addition to or maybe as a compliment to your idea for an antibody maker, a still very rough line of thinking I’ve been exploring is: what would it look like to really try to empower people to take charge of their own biology? For example, is it hypothetically possible to develop very safe (to test) therapeutic platforms combined with much better realtime monitoring of their effects such that people can be trusted to experiment with them? This is totally hypothetical but imagine the ultimate next-gen gene therapy that you can test in low dose, rapidly deactivate via light or some other noninvasive mechanism combined with non- or minimally invasive ‘omics plus hormone monitoring. On one hand, it’s scary and easy to immediately think about all the ways this could go wrong, but it does seem philosophically aligned with taking control of biology (including our own).

Besides that, I am a big fan of the aesthetic and associated ideas Arye Lipman discusses here around engineering the biology around us. Maybe eventually the antibody printer could be a plant or a microbe in my gut! Or as a more near term example, Neoplants!

Are there other early wins besides the ones you outlined in your essay that you’re thinking about?

EH: Yeah, this is a super interesting point. Something I’ve been thinking a lot about has been the apparent disconnect between all of the amazing stuff we can do in the lab and the actual translation into impact on human health and novel products. Matt Herper had a great essay about the bottleneck of trial infrastructure, and Samuel Rodriques also talked about this.

It seems like we do need new ways to think about therapeutic experimentation. It seems like there is some low-hanging fruit—curbing rising trial costs, better prediction models to increase trial success, improvement to the incentive/financing structure of biotech—and also more creative technical approaches we could take.

I think that this bottleneck—for both therapeutics and synthetic biology—as well as your earlier sentiment about the lack of big picture narrative (taken seriously) in biotech are both axes that we need to make improvement along.

SM: Agree with everything, re: therapeutics.

Elliot Postscript #

I’ve been thinking a lot about AI, biotech, and the relationship between the two. I’m convinced that these are two of the most impactful areas of technology in this century—and that there is tremendous promise at their intersection. After thinking about this more—in no small part from this conversation—I wrote an essay about my view that biology is actually the most powerful way to transform the physical world using AI. You can read that essay here.