Avi Marcus

@swe_acc @shae_mcl Directly obviously not, but you could try to bootstrap over 2-4 aa peptides and start from there…. If I had infinite time I would probably go down that rabbit hole, unfortunately I don’t

@Avi_Marcus @shae_mcl Solving folding by QM is, uh, going to take a bit more than that.

It’s estimated that the Protein Data Bank (PDB) cost around $13B to create. Alphafold was only possible because of it. If we want ML to solve biology, we should be funding the creation of databases and the development of new assay technologies. ML is nothing without data.

@beffjezos Yud as a service

Yud did this

@ATinyGreenCell @_inc0_ I have to admit that I am more intrigued by an enzyme being able to absorb 9000 photons a second productively. How doesn’t it denature???

@_inc0_ yeah absolutely we need this to be true at some point. this is why im so deeply critical of people who make claims like this with no backup since it wastes everyones time and dwindles hope for cynical misanthropic engagement farming

"That which is asserted without evidence can be dismissed without evidence." -Hitchens' Razor

@Gaurab R0 is the most important metric and we don’t know it yet

I’d only get scared of a pandemic if hanta can mutate to be non-lethal and transmit significantly from respiratory transmission in humans. Its lethality and inability to spread via respiration limits the chances of a pandemic. Can’t transmit the virus if you’re dead. 150 ppl on that cruise ship, but only 7 cases. The barrier for these mutations to evolve is massive. You’d need to infect a lot of humans to see that kind of selection. Paging doctor @balajis .

I love that people are working on cures for viruses. I don’t even have a problem with gain of function research, we need it to learn what to protect against. My problem is that all of this research is done by humans in labs where a mistake has consequences. I just want these things to happen in autonomous labs, where only data is passed to the outside

@Avi_Marcus Lots of scientists are afraid to work on cures for viruses because they know they’ll be accused of evil things

I wonder how many scientists tried to study hantavirus in the last 14 years but couldn’t find funding

@SynBio1 I am more worried about pubmed.ncbi.nlm.nih.gov/?term=hantavir…

@Avi_Marcus But your question is easy to answer though pubmed.ncbi.nlm.nih.gov/?term=hantavir…

@shelbynewsad A year ago I was certain that I will recruit at least one cofounder and a few people. I don’t believe that anymore. Between labs getting automated and AI, I just don’t feel the need to do it

Run an entire company with agents… Underused in bio right now!

@shelbynewsad Relatively small datasets are enough for embedding->function prediction, and in most cases hill climbing of a barely working variant produces good enough results

@shelbynewsad I agree the question is what the cost trade offs are between more data and more compute, but aim for good enough. Absolute maxima/minima is not the threshold we should be aiming for, it isn’t realistic in a 20^(200-300) search space (combinatorics of proteins)

what the field of TechBio is realizing is the untapped opportunity in bio-AI is in newer, small data areas. some areas id love to see updated, scaled, translated to startups are: - AI for self-assembling materials - explosive/energetic materials - synthetic ecology - adhesives/coatings - photonic crystals - AI for flavor (a lot to watch out for here 👀)

@SynBio1 Closest to thing that to that dataset is probably the internal emails of the the Wuhan Institute of Virology

It’s hard to collect data about how bioterrorists might try to use AI. Few people want to create a bioweapon, and those who might aren’t talking. On the other hand, it's easy to predict how the news will cover bioterrorism and how social media responds. We have years of clickbait headlines and viral scareposts to train on. This makes it much simpler to build a biosecurity policy around avoiding bad headlines, rather than installing safeguards that would actually stop bad actors. I have a PhD in Synthetic Biology. I know roughly what it would take to make a bioweapon. It would be enormously difficult and dangerous. Most of the work is in the physical world, where AI tools would be only marginally useful. None of the relevant uses of AI look anything like the examples cited in the NY Times story below. - Printing 8,000 word protocols for methods already in the public domain - Making a list of common cattle diseases - Generating a shopping list of test tubes and media - Describing how to use a weather balloon The actual biosecurity questions that need answers are technical and too boring to cover in a major media outlet. - How can we tell the difference between a dangerous DNA sequence and a harmless one? - What separates a python script used to discover a therapeutic from one used to discover a toxin? - Which practical R&D bottlenecks are being rapidly opened by AI and which are not? Much of the work of biology happens in the real world and doesn’t involve AI much at all. A serious biosecurity policy needs to focus on how bad actors might access physical hardware, specialized facilities and trained personnel. These are infinitely more important barriers than what Claude might tell someone about weather balloons. My point here is that the people telling you to be afraid, and the media outlets who cover them, are putting us all in danger. The big AI shops are going to lock down their models, not to stop bad actors, but to stop bad press. Training models to stop using scary words is easy, the real work of biosecurity is hard. If we don’t push back, we’re going to end up with an industry dedicated to performative biosecurity theater. nytimes.com/2026/04/29/us/…

AI for Low-Pesticide Agriculture @garrytan Farmers are stuck in a bad loop: use more chemicals, get diminishing results, pay more, take on more risk. And they can't just stop, because if pests win, crops die. AI that can identify individual weeds in real time, robotics that can treat one plant instead of blanketing a field, and new biological solutions mean this problem finally looks solvable.

@ycombinator @garrytan The generational company would be the one that replaces plants with a combination of solar panels and vats with proteins that create our food without any cells. The missing link is an enzyme recycling ATP using electricity and that is my mission

AI has stopped being a feature and started being the foundation. We're excited about a new wave of startups rebuilding software, services, and silicon— and pushing AI into the physical world. ycombinator.com/rfs

One day an in silico protein design library will be able to create a pocket for ADP that doesn’t fit ATP. Today is not that day

@tlbtlbtlb Not all exponentials are alike, it won’t be like current AI compute. It will be like Cyanobacteria in a pond

Lots of things grow exponentially but aren't free because demand growth exceeds supply growth. In a world where robots are building Dyson sphere of solar panels, there'll also be robots building things that use energy. Everyone who needs energy will be willing to pay for more of it. That sets the price.

People should stop predicting that the marginal cost energy will go to zero. That can only be true if the marginal demand becomes zero, ie people can't think of anything useful to do with more energy. I don't think I'll run out of ideas.

Have you wondered what the wet lab success rates are for current AI-driven protein design models? Look no further! In our new open access review, @KevinKaichuang, @avapamini, @SarahAlamdari, and I report wet lab success rates for *over 200* different protein design tasks 🧬💻

@kosonocky @DdelAlamo @KevinKaichuang @avapamini @SarahAlamdari Any stats about the correlation between the predictions of Boltz2/Chai/OpenFold-3 and success? I care about the tools used to design less than the correlation of the compass used to navigate to reality