Why We're Skeptical AI Will Compress 50 Years of Biology Into Five
We're optimistic about AI in the lab. The timeline is the part we doubt, and the fastest gains today are in how lab work is organized.
By Yaoyu Yang ·
Any biologist who uses AI should weigh what a model can do against how messy biological research is.
In October 2024, Dario Amodei published "Machines of Loving Grace," an essay arguing that powerful AI could "compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years." The claim is inspiring. It's also ambitious. Two years on, we're still skeptical of the timeline, even with the capable models we now have at our fingertips.
WHY ARE WE SKEPTICAL OF THE 5-TO-10-YEAR TIMELINE?
Our skepticism comes from the bench. Our team has spent more than a decade building software for labs, much of it at Ginkgo Bioworks, across DNA synthesis, assembly, high-throughput screening and fermentation. We helped make a lot of products that work. Getting to each one cost far more than anyone planned.
The hard parts were rarely about intelligence. Getting a single clean plot for a PhD thesis once took us about a year. The cells didn't grow, a time point got missed, a culture overgrew, the media got contaminated, and a machine ran out of maintenance liquid. None of that was a reasoning problem.
WHAT DOES "MACHINES OF LOVING GRACE" GET RIGHT?
The essay gets a lot right, including its own caveats. Amodei trained as a biophysicist, with a PhD from Princeton and a postdoc at the Stanford University School of Medicine. Those caveats come from someone who has done the science.
"Cells and animals run at a fixed speed," Amodei writes, and clinical trials "involve a lot of bureaucracy and regulatory requirements." The essay even says its forecast "is not based on any rigorous methodology." We respect that candor.
The essay's answer is that AI can work around these limits, with experiments run in parallel, better cell and animal models and a better trial system. Where we differ is on how hard the limits are. The gap between the predictable logic of computing and the behavior of living systems looks wider to us than the essay allows.
WHICH LIMITS CAN'T A MODEL REASON AWAY?
Biology has physical constraints that even the smartest model can't reason away. Three of them sit largely outside what AI can change.
The first is time. Cells grow at the speed of cells. Running experiments in parallel helps once a protocol works, but getting a protocol to work is usually the slow part.
The second is causal data. Biology has few clean, unambiguous datasets. Decades of patient records tell you which treatments went with which outcomes. For a new drug, that tells you little about how it will act.
The third is trial throughput. Regulation, clinical sites and patient enrollment cap how many trials can run, however many candidates come out of a model. Unless you can predict the trial result itself, you still have to run the trial. Nobody can predict it yet.
HOW EXPENSIVE IS BEING WRONG IN BIOLOGY?
Being wrong in biology is expensive, and most bets lose. A single pivotal trial, the late-stage study an approval rests on, had a median estimated cost of $19 million in a 2018 JAMA Internal Medicine study of drugs approved in 2015 and 2016.
Only 7.9% of drugs entering Phase I between 2011 and 2020 went on to FDA approval, according to BIO, Informa Pharma Intelligence and QLS Advisors. That's how a biotech can spend all of its money and come out of the clinic with nothing.
Better candidates from AI could raise those odds, and that would matter, but they still need a full dosing period and enough patients to show an effect.
WHERE CAN SCIENCE MOVE FASTER TODAY?
We think the fastest gains are in how science is organized. The scientists we work with still struggle with things the tech world takes for granted, like standard protocols, shared records and coordinating work across a team. A lot of research is slow because nothing ties those pieces together. That's the part we work on.
At an AI for Science evening during SF Tech Week this month, a panel made a point we keep coming back to. A cell-binding assay might take four or five days to run, but optimizing it can take months. Every parameter matters, from reagent concentrations to pipetting technique.
Picture three scientists optimizing that assay over twenty runs. When it finally works, the conditions behind the good runs are split across ELN entries, a spreadsheet and someone's memory. When the assay moves to a new hire or a CRO, part of those months gets spent again.
The published record shows the same gap. The Reproducibility Project: Cancer Biology set out to repeat 193 experiments from 53 high-impact papers. None of the 193 was described in enough detail to design a replication without asking the original authors.
This is fixable today. It takes protocols written in a standard form, each run recorded against the protocol version it used, and results kept where the next person can find them. A model can only reason about what was written down, so complete records also get a lab better answers from AI.
Cypher AI is digital infrastructure for life-science R&D: an ELN, a LIMS, data analysis and a services Marketplace in one platform, with an AI agent that works across them. Protocols in Cypher AI are versioned, and each run records the protocol version it used. The agent can draft a protocol from a description, and the scientist reviews it before it's saved.
FREQUENTLY ASKED QUESTIONS
What does "compressed 21st century" mean? It's Dario Amodei's term, from "Machines of Loving Grace," for the 5 to 10 years in which powerful AI could deliver 50 to 100 years of biological progress.
How long does a drug take to go from Phase I to approval? About 10.5 years on average for drugs in development between 2011 and 2020, according to BIO, Informa Pharma Intelligence and QLS Advisors.
Can AI speed up assay optimization? Yes, in part. AI can help plan the next conditions to test and analyze each run faster. Each run still takes as long as the cells take.
HOW CAN YOUR TEAM START?
We're hopeful about AI in biotech. We just don't see models closing the gap between computing and biology within five years. The teams that move fastest will be the ones whose records let every run build on the last one.
A good first step is one protocol your team runs often, ideally one you're still optimizing. Record the next few runs against it and see what the history shows. We're happy to walk you through it, or give you access to try it yourself. Start with a free account at cypherbio.ai and book the 15-minute onboarding call from the home screen.
For R&D organizations that need their own cloud, SSO or a forward-deployed engineer working alongside the team, the Enterprise plan covers that. You can book a conversation through the demo link on the site.