One of the most interesting panels at the Ai4 conference this year was about building the future of medicine with AI, and I came away feeling like I had watched two conversations happen at once. One was about speed: how much faster and cheaper we can make drugs now that machine learning touches every stage of discovery. The other was quieter and more unsettling: what it means that we are starting to write the language of life the way we write software, and whether we are ready for that.

The panel moved between longevity, generative biology, drug development timelines, and biosecurity, and the through line was that biology is becoming programmable. The second conversation matters more than the first, and it is the one most people are not having yet.

The two problems AI is actually solving in biology

AI in drug development gets described as if it were a single breakthrough, when it is really two very different problems wearing the same coat.

The first is scientific. Biology is complicated, and designing a molecule that does what you want inside a living system is genuinely hard. AI is already making a real dent here, helping design better molecules and predict how they behave.

The second problem is the process itself. Turning an idea into an approved medicine is not one step, it is somewhere between one and two thousand of them, spanning target identification, chemistry, manufacturing, regulatory work, and clinical evidence. You have to be good at all of them. Failing at any single one kills the program.

The scale of that process problem is easy to underestimate. Bringing a single new drug to market takes roughly 10 to 15 years and costs somewhere between $1 billion and $2.6 billion per approved medicine once you account for all the failures along the way (figures from RAND’s 2025 JAMA analysis, the US Congressional Budget Office, and Tufts Center for the Study of Drug Development). The scientific wins are the ones that make headlines, but the process wins are where the compounding value lives. The maximum impact comes from bringing all of that context together in one place at one time, and AI is uniquely suited to that, because it can hold the whole pipeline in view rather than optimizing one heroic step and leaving the rest untouched. That is a different claim than “AI designs better drugs.” It is “AI runs the factory that makes drugs,” and that is the more important idea.

Generative biology means treating DNA like code

We built large language models that generate human language by learning its patterns. Now the same idea is being pointed at DNA. Instead of tuning individual molecules by hand, the ambition is to design biological function directly, programmatically, and repeatably. Large language models for the language of life.

Generative biology borrows modularity from engineering. In software and hardware we do not reinvent every component from scratch. We assemble reusable parts with predictable behavior and compose them into something new. Generative biology asks whether we can do the same with living systems: define reusable biological components, understand how they behave, and combine them to produce a desired outcome with reliability. That is a profound shift in mindset. It moves biology from something we discover to something we design.

There is an honest caveat woven through this. Modularity and prediction only get you so far, because at the end of the day the clinic tells you the truth. The pragmatic version is a loop: use predictive models to narrow the possibilities, build the thing, observe what actually happens in humans, and feed that back into the next design. The engineering mindset supplies the hypotheses. Reality supplies the verdict. That framing keeps the ambition grounded. We are not replacing biology with math. We are using math to search biology more intelligently.

The dual-use problem is not a footnote

A team on the panel is known for building foundation models trained on DNA, and they described generating a genome from scratch that turned out to be a functioning organism, a bacteriophage, a virus that infects bacteria. Let that sit for a second. An AI system designed a genome, and when it was built in the lab, it worked.

This is not hypothetical, and the specifics are worth knowing. In work published in Science in August 2026 (with a preprint out in September 2025), researchers at Stanford and the Arc Institute, led by Brian Hie and graduate student Samuel King, used two genome language models called Evo 1 and Evo 2 to design complete bacteriophage genomes modeled on ΦX174, a small, well-studied virus that infects E. coli and cannot infect humans. They generated hundreds of thousands of candidate genomes, synthesized roughly 300, and 16 of them assembled into viable, functioning viruses, some of which infected bacteria more effectively than the natural version (Science; Arc Institute; reporting in Axios and Science News). The team deliberately excluded genetic sequences from anything that infects humans, animals, or plants, precisely because of the risk.

Here is the uncomfortable symmetry the panel named directly: the same technology that lets you program useful function into an organism also lowers the barrier to making something dangerous. You cannot separate the two, because they are the same capability. As one panelist put it, the same models that let you program a genome into an organism also lower the barrier to build something harmful. This is different from most technologies of concern. You cannot easily defend against a nuclear weapon by building a bigger one, but biology is different, because the tools that create biological threats are also the tools that build biological defenses. That gives me a strange kind of hope, but only if we actually invest in the defensive side.

The warning is already concrete. Alongside the Science paper, biosecurity specialists at Johns Hopkins pointed out that the current infrastructure for screening synthetic DNA orders cannot reliably detect novel sequences like the ones these models produce, and no US law forces synthesis companies to screen for them (reporting in The Washington Post and Tech Times). The people describing this risk were not alarmists. They were the ones building the frontier, and they were the ones saying the frontier needs guardrails. The company on the panel framed its own mission as pushing capability and safety at the same time, treating them as inseparable rather than as a tradeoff. That is the right instinct. The wrong instinct, which is more common, is to build the capability and assume someone else will handle the safety.

Longevity as the highest-leverage bet

After roughly 250 years of pharmacological research, we do not have a single product that reliably gives a healthy person one extra year of life beyond average expectancy. Not diet, not exercise, not diagnostics, not the drugs we already celebrate. The newest metabolic therapies, the GLP-1 drugs like semaglutide that reshaped the weight-loss market, help with weight and may nudge things in the right direction, but the core claim stood: we have never bought a healthy person a guaranteed year.

The number that reframes the stakes is this. If you give one additional year of life to every person alive, you generate roughly 8.2 billion life-years at once, since there are about 8.2 billion people on Earth today (UN World Population Prospects). That scale dwarfs almost any other intervention humanity could make. Put that way, aging stops looking like a niche interest and starts looking like the largest addressable problem in medicine, because it is the one condition every single person shares. Aging takes everything away from you regardless of how hard you work. Targeting it is not vanity. It is the highest-leverage bet available.

This does not mean we should abandon disease-specific medicine. But the argument that aging is upstream of most age-related disease, and that hitting it could address many conditions at once, is more serious than the anti-aging marketing that has cheapened the field. The panel’s version was scientific and economic, not cosmetic.

Speed is real, but the bottleneck moved

The acceleration numbers are real. Getting from zero to a development candidate, the milestone just before human trials, has traditionally taken a large pharmaceutical company around 4.5 years and hundreds of millions of dollars. Moving that work to China cuts it to roughly two and a half years. Adding frontier AI on top can save another year or more. One panelist reported an average of nine to thirteen months to a development candidate across 31 candidates, with a failure rate lower than historical norms. This tracks with the company profile of Insilico Medicine, which has publicly described AI-designed candidates reaching the clinic on compressed timelines. That is not incremental. That is a different tempo of science.

But one sentence from the panel mattered more than the headline numbers and got the least attention. After the development candidate, once you enter human trials, you move at the speed of the regulated system, and that part does not accelerate. Clinical trials are slow and expensive on purpose, because they exist to protect people. AI compresses the front of the pipeline dramatically and then hands off to a back half that runs on the same clock it always has. Which brings me to the assumption underneath all of this.

The assumption nobody quite examined

The whole optimistic case rests on a premise that was stated only in passing: that accelerating and improving the preclinical stage will translate into more approved medicines reaching patients. That does not obviously follow.

Most drugs fail, and they mostly fail in the clinic. Around 90 percent of drug candidates that enter human trials never reach approval (ScienceDirect, 2022; BIO Industry Analysis). When you break down why, roughly 40 to 50 percent of failures come from lack of efficacy and about 30 percent from unmanageable toxicity (published analyses summarized in medRxiv, 2025). These are exactly the failure modes that preclinical work, including AI-designed preclinical work, often cannot predict, because a biological hypothesis that holds in animals frequently breaks in humans. If the real valley of death is clinical rather than preclinical, then making the preclinical stage ten times faster and cheaper produces ten times more shots on goal, but at the same goal, defended by the same goalkeeper. You get more candidates entering trials, not necessarily more drugs coming out the other end. The bottleneck does not disappear. It moves, and it becomes more visible.

This matters because the two big promises of the panel, radical speed and radical safety prediction, are in tension with each other.

Speed comes from the parts of the pipeline AI can compress. Real safety and efficacy signals come from the part it cannot, the human trial. The panel gestured at a future where regulators might eventually accept AI-generated evidence and deregulate some of the toolbox in five or six years. Maybe. But that is a policy bet layered on top of a scientific bet, and it is doing a lot of quiet work in the optimistic timeline. If regulators do not move, the clinical bottleneck holds, and the acceleration story becomes a story about cheaper failures rather than more cures. Anyone building in this space should be honest about which of those two futures they are actually betting on.

Testing what has never existed

When AI designs a molecule or a genome that has no precedent in nature, you cannot look up whether it is safe. The truth only arrives when you physically build it and test its function in a lab. There are heuristics you can lean on, like measuring how similar a design is to known natural structures, but the genuine bottleneck is empirical testing, and that is slow. The Stanford and Arc phage work makes the point: hundreds of thousands of candidate genomes on a computer collapsed to about 300 worth synthesizing, and only 16 of those actually worked.

The right role for AI in the near term is not to replace the test but to narrow the search. Instead of exploring a near-infinite space of possible designs, you use models to identify the small set worth building and testing. AI does not remove the need for physical validation. It makes each experiment count for more by pointing it at better candidates. That is a more accurate picture of where we are than “AI designs the cure,” and it is the version worth trusting, because it matches how the results actually came in.

Where this leaves me

Programmable biology is not a distant idea, it is a present capability, and the conversation has to grow up fast to match it. We spent decades treating drug discovery as a slow, artisanal craft. We are now treating it as an engineering discipline, with all the speed, modularity, and dual-use risk that engineering brings. The recombinant DNA debates of the 1970s offer a decent template. In 1974, leading molecular biologists called a voluntary moratorium on certain experiments, and in February 1975 about 140 scientists met at the Asilomar Conference Center in California to write safety guidelines that the NIH formalized in 1976 (Nature; NobelPrize.org; National Library of Medicine). The community paused, set rules, and moved forward responsibly. We may need a moment like that again, and we should reach it before a crisis forces it rather than after.

The practical takeaway is to hold two things at once without collapsing into either. The upside is enormous, measured in billions of life-years and diseases we could finally address at their root. The risk is real, concentrated in the same tools that deliver the upside, and currently under-defended. Anyone serious about this field has to fund both sides of that equation. Building the frontier without building the defenses is not optimism. It is negligence dressed as progress.

Key takeaways

Here are the takeaways I am leaving with:

1. AI solves two separate problems in drug development, the scientific one of designing molecules and the process one of coordinating a thousand-plus steps across a 10-to-15-year, billion-dollar-plus pipeline. The second is less glamorous and more valuable.
2. Generative biology is the attempt to treat DNA like code, using engineering principles of modularity and predictability to design biological function rather than discover it.
3. The dual-use problem is inseparable from the capability. The same models that designed 16 working AI-generated viruses lower the barrier to dangerous ones, and current DNA-synthesis screening cannot reliably catch novel sequences.
4. Aging is arguably the highest-leverage target in all of medicine, because it is the one condition every person shares and because a single year of added life across humanity generates over 8 billion life-years.
5. AI compresses the preclinical stage dramatically, cutting years and hundreds of millions of dollars, but the clinical stage does not accelerate, so the bottleneck moves rather than vanishes.
6. The hidden assumption in the optimistic case is that faster preclinical work yields more approved drugs. Since roughly 90 percent of failure is clinical, we may just get more cheap shots at the same hard target.
7. For genuinely novel designs, physical lab testing remains the ground truth. AI’s near-term value is narrowing the search space so each experiment counts for more, not eliminating the need to test.