The first time I saw AlphaFold, it felt like stepping into the future, a moment where something previously thought impossible had suddenly become inevitable. AI had cracked the protein-folding problem. Well, not cracked exactly, but solved it well enough that it rewrote the playbook for structural biology. Overnight, the way we thought about proteins, drug discovery, and molecular engineering shifted. That was 2018.
Since then, I’ve been watching the field evolve with the same kind of anticipation, waiting for the next AlphaFold moment, the next leap forward that will shake the foundations of biotech. We’ve had AlphaFold2, AlphaFold3, ESMFold, and a couple of Nobel Prizes. But lately, something’s changed. The pace that once felt unstoppable, the unrelenting acceleration of AI-driven biological discovery, seems to have hit resistance. The latest report from Epoch AI confirms what I and a few colleagues I’ve chatted with have been feeling: AI in biology is slowing down.
From 2017 to 2021, the trajectory was almost absurd in its growth, compute scaling at breakneck speed, model complexity exploding, datasets ballooning in size. I started working in the field in 2015 and it was hard to keep up, even in the early days.
AI-driven breakthroughs were piling up so fast it felt like we were racing toward some inevitable singularity where life itself would be programmable. But since 2021, that feverish pace has cooled. The scaling curves that once shot upward like rocket trails are bending, flattening, tapering off. The field is still advancing, but no longer with the same relentless momentum. Despite all the hype and hubbub that has ballooned since the 2022 release of ChatGPT, the biology world is not ending with AI, yet.
So what happened? And more importantly, what happens next?
Why Biology Felt Like the Perfect AI Playground
For a while, it seemed like biology and AI were a match made in heaven. The raw ingredients were all there: vast amounts of genetic data, molecular structures encoded like natural language, the kind of massive, messy, high-dimensional problems that deep learning thrives on. Biology, at its core, is just information in physical form, and AI is exceptionally good at wrangling information.
That realization led to an explosion of work in the space. It wasn’t just AlphaFold. There were protein language models that treated amino acid sequences like sentences, drawing meaning from evolutionary history. There were generative models capable of designing enzymes that nature never thought to create. There were AI-driven drug discovery platforms promising to upend the way we find and optimize molecules.
For a while, it felt like we were unlocking something fundamental. Compute was cheap, data was plentiful, and researchers were riding the high of early successes. The mindset was simple: if a model wasn’t working, just make it bigger, feed it more data, throw more compute at the problem.
And then reality set in.
The Bottlenecks That No One Wanted to Talk About
Scaling brute-force solutions works, until it doesn’t. Biology is not like language or images, where vast public datasets exist for AI to feast on. There’s no equivalent of the internet’s infinite text supply when it comes to molecular biology. High-quality experimental data is scarce, expensive, and often proprietary. There’s also the simple, stubborn fact that biology is messy. It refuses to conform to neat mathematical representations. Predicting how a protein folds is one thing, understanding how it actually behaves in a living system is another beast entirely.
Beyond data, the cost of AI development itself became impossible to ignore. Training biological AI models isn’t cheap, and as the world’s appetite for AI grew, so did the price tag on the chips that power it. The money was still flowing, but investors were becoming more cautious. Venture capitalists who once threw cash at anything labeled “AI for drug discovery” started asking harder questions. What was the path to commercialization? Could these models actually replace wet lab experiments, or were they just very expensive guesswork?
As a side, that is true for most startups. Some, however, seem to raise money like it grows on trees. e.g. EvolutionaryScale, Xaira.
Then came the security questions. AI in biology isn’t just a tool for discovery, it’s also a potential threat. If we can use machine learning to design new enzymes, new proteins, new molecules…well, it doesn’t take a dystopian imagination to see the risks. Governments started paying closer attention. Regulation, once a distant afterthought, became a serious consideration. Some research groups started keeping their models behind closed doors. Others implemented restrictions, limiting access to prevent misuse. I happen to be one that believes we should have a healthy skepticism of fanatical promises and doomsday scenarios.
According to Epoch AI’s dataset, fewer than 3% of biological AI models include meaningful safeguards. That’s an alarmingly low number, considering how powerful some of these models have become. And yet, the most advanced systems, the AlphaFold 3s, the massive foundation models trained on the full depth of evolutionary biology, tend to be the ones with the most security controls in place. That’s not a coincidence. The teams building these systems understand what’s at stake. The recent release of Evo2 is a great example of teams trying to build responsibly.
The End of the Scaling Era And the Beginning of Something Else
So here we are. The wild early days of unrestrained growth are behind us. AI for biology is no longer the frontier free-for-all it was a few years ago. The question isn’t how big can we go? anymore. It’s how do we go smarter?
The future won’t be about throwing more GPUs at the problem. That model is already running out of steam. Instead, we’ll need systems that make better use of less data. We’ll need models that aren’t just memorizing vast biological datasets, but actually reasoning about them, forming hypotheses, designing experiments, making connections that human scientists might miss.
And we’ll need to integrate these systems directly into the experimental loop. The most exciting thing happening right now isn’t AI models themselves, it’s what happens when you put AI in the driver’s seat of an automated lab. Imagine a system that doesn’t just predict a protein’s structure, but designs it, tests it in real-time, refines its approach, and learns from each experiment. That’s the real future of AI in biology: closing the loop between computation and experimentation.
There’s also the question of governance. The days of building these models in the open, releasing them on GitHub, and hoping for the best are hotly contested now. If AI is going to transform biology the way we all hope it will, it needs to do so in a way that’s responsible, transparent, and secure. That means useful safeguards, clearer guidelines, and an understanding that just because we can build something doesn’t always mean we should.
This Isn’t a Plateau, It’s a Pivotal Moment
If you’re in this field, it might be tempting to feel a little disillusioned. The energy of the last few years, the rapid-fire breakthroughs, the feeling that every month brought something new and revolutionary, it’s not gone, but it’s changed. The gold rush phase is over. But that doesn’t mean we’ve peaked.
What we’re seeing now is a natural evolution. The first wave was about proving what was possible. This next phase will be about proving what’s practical, scalable, and truly transformative. The work ahead won’t always be as headline-grabbing as an AlphaFold moment, but it will be deeper, more thoughtful, and ultimately more impactful.
We’re not slowing down. We’re getting smarter. And that’s what’s going to make all the difference.
Keep pushing boundaries and stay connected.
Cheers,
— Titus
Share this post