At cheeky face value, a town isn’t enough, since we have done that before and we got a bomb instead. But if they are better than any geniuses we’ve had so far, maybe?
Anthropic’s Dario Amodei has been one of the few tech CEOs that has done well with his predictions – from his essay Machines of Loving Grace, the progress from 2024-2026 lines up well with what has been achieved for agentic coding and assisting various other tasks. In his followup from a few months ago, The Adolescence of Technology, he currently predicts 2 years (~1-4 years) after which we should have AGI and geniuses with country population-level availability. The AI companies and government will have the country of geniuses, but it’s feasible you could have access small town of geniuses as an individual if this turns out to be the case.
Another AI CEO that I’m watching carefully is Demis Hassabis, of Isomorphic Labs and DeepMind. Hassabis is working on solving drug discovery via molecular simulation, and said in an interview I can’t find yet that we would have full cell simulation in 5 years. Since we are just now able to simulate how single molecules are shaped and interact with, which was a huge leap, most people are reasonably skeptical that we’ll be able to handle the entire cell, which has interactions between structures of these molecules and proteins (which number in the high millions or billions per cell). Elliot Hershberg’s writeup on virtual cells and the problems to overcome is a nice writeup on the topic. But maybe AI will have a breakthrough here. For example it could discover some structural understanding of the problem that really does let us abstract the problem to be small enough to be manageable without loss of simulation quality.
- Longevity biotech
For Longevity Biotech, where we the goal to discover integer or order of magnitude improvements for living healthily for longer and longer, the drug discovery route from small molecules has been devastatingly slow. It’s not just because of the often lamented regulatory pipeline, which is necessary to some extent for safety. We still don’t have much better actual results than Rapamycin, which is a drug we found in the ground by accident in the soil/sand in the 70s, of which the human effects will at best probably add single digit years to healthspan. This looks similar to what GLP-1s are going to do for a large part of the population via a different mechanism. Neither drug came directly out of the search for longevity drugs (Rapamycin was first looked at for antifungal and immunosuppresant properties, GLP-1s for diabetes).
The root causes of aging initially seemed like a great way to start the search. The diseases of aging like cancer, heart disease, brain dysfunction are merely symptoms of some totally different root causes. If we could just categorize these problems, maybe they are easy to solve? But it turns out the 7, 9, or 11 so called hallmarks of aging are also beasts of their own that are difficult to target with precision. Genomic instability or proteostasis for example, are rather broad and complicated and may have their own root causes.
2. AI
Today, there are at least three alternatives to solving the root cause of aging. Replacement (of organs, body parts, cells, bodies) is at the top of the list as far as Overton window shifting goes. Cryonics is at the bottom in terms of timeline, but over the past two years it has gained traction in the field. ‘AI’ has been rising as the progress in coding and engineering picks up. ‘AI’ has always been the black sheep, with no concrete path to longevity other than the belief that AI supergeniuses would be able to produce a path. Now we start to overlap with Dario’s Country of Geniuses idea.
This is brushing over the various uses of ML/AI that have been used in biotech/pharma labs for years, and the recent success with AlphaFold and protein docking we discussed earlier. But this wider ‘AI’ is really not about any current pathway. It’s a superset beyond what Isomorphic set out to do with cell simulation, and could also go down the replacement or cryonics pathway. I think the hope is that it will discover totally novel pathways to longevity.
The less concrete things are, the more skeptical we should should become. On the one hand, if scaling laws hold up, throwing more compute and physical resources may allow us to build many supergeniuses, which should in theory be able to come up with better solutions, including roadmaps to longevity interventions.
On the other hand, there are other bottlenecks to worry about besides geniuses. How will these geniuses experiment? In silico experimentation is used to some extent today, but only on tiny problems, and only because of the innumerable in vitro and in vivo experiments in wet labs and clinics. The CRO wetlab-as-a-service structure seems like it might actually help here, but is still a potential bottleneck and will definitely regulate the speed of iterations between experiments that depend on each other.
Yet another trend is the cell cultivation field, for stem, and other cell types, by multiplying cells or simple structures in a substrate on a petri dish. If optimizing the substrate or environment for cell quality, efficiency, or numbers, the loop should be pretty manageable with a bit of robotics. This is where I would bet to see early results in the bio AI space. Of course, having unlimited stem cells is not going to solve the problem of longevity. You could squint a little and see that working out for companies that test drugs on organisms with some AI and robotics driven rapid testing platform like Olden Labs (mice) or Ora Bio (c. elegans).
It’s also plausible we have AI-based chemical biologists that are able to just look at enough structures, proteins, and somehow become able to hypothesize a miraculous de novo small molecule or some other solution for each problem or root cause of aging with high specificity and low toxicity with minimal wet-lab requirements. This is close to sci-fi in that they would appear to be like magic to us. Let’s leave aside the problem of delivery for now. Such an agent would have to go far beyond today’s chemical biology and really understand the off-target effects and how the molecule would interact with every other part of the body. This approach really relies on understanding far beyond what we have today, and defers most of the work onto the AI. This makes it the most fantastical.
Replacement of bodies, limbs, or organs would allow us to solve longevity without understanding how the underlying unit works as a whole, just enough to know how to insert or connect it in a safe and timely manner to the other parts. Jean Hebert’s vision that he laid out in his book Replacing Aging was powerful before the AI boom, and now he’s a director at ARPA-H (MIT Technology Review summary), so I expect to see major progress there. Robotics is one of the things being elevated by the new AI capabilities, and this should be a boon for replacement-based approaches with advances and rapid iteration for surgical robotics. The other side of replacement, which is the production of the thing you are trying to replace, has less clear of a surface that AI could approach. There was a somewhat skeptical piece last month about this in the MIT Technology Review. It’s harsh, but serves as a good reminder of the optics, economics, and social issues that need careful consideration for those brave pioneers in this area.
3. Refrain
It definitely feels like everything is a nail right now for our AI hammers. But I am excited that the AI slide deck on more companies might actually represent a real advantage. Of the AI-combining approaches above, I’m sad to say that the one that has the most appeal is the magical de novo approach, followed by the replacement approach, because these are the only two that are the kind of true paradigm shifts needed to bring non-incremental results quickly. I really want to put replacement first, followed by cell-simulation. To be clear, I fully expect there to be AI-driven improvements in all the other approaches as well (at a minimum, we should see the improvements from the ease of coding and ML models), but the other approaches are slow and incremental. The skeptic in me thinks they will be able to make incremental improvements that we have seen so far, just a bit faster.
What would make me change my mind: if Isomorphic produces ‘full’ or reasonably complex cell simulation at the molecular level to the extent that they can accurately predict the molecular results of new drugs in vivo (not the efficacy results or proteomics, but actual molecular states), then I think it won’t be long before they can demonstrate full organism in c.elegans or other simple organisms. After that true in silico should yield the immortal worm, and probably the more complex life forms to follow. I think that will be hard to get to in 5 years with the current transistor scaling paradigm, but I’m not willing to discount it.
I’m a long time pessimist on quantum compute, but it does have the promise of doing powerful parallel computing that would in theory be useful for molecular bio simulations. If we actually see improvements and adoption to the level we saw with AI in 2022, this would change my mind too. On the replacement side, if we see rapid progress in the production side of the problem because of AI and the robotics part is working out, I will be very willing to change my mind.
Lastly, this area will always have pessimists and naysayers that are just gut-reacting to something new. Being skeptical is an excellent quality only if it is paired with the ability to update on new information. I applaud knowledgeable skeptics that seek our current information like Derek Lowe and his column in Science that calls out AI hype in drug discovery (here’s a recent piece looking at an open source model similar to AlphaFold). These are the people to watch for those slightly outside the field, because their minds can change. In this sense, the open minded skeptic messages contain much more information in the Shannon sense than the startup founders, who are always bullish. Both are worth listening to.