Recently I had a 28-hour day. I felt completely locked in, in launch mode. It’s the clearest picture I have of what AI does to work: it doesn’t hand your afternoon back. It hands you more, faster, and the loop doesn’t end unless you end it.
I’m pro-efficiency, and I want more of that speed in my industry, not less. But I’ve come to think AI made one kind of work cheap and left another as expensive as ever. Doing got cheap. Being right didn’t.
The clearest example of the first half landed last week. On October 27, Anthropic said it had bought Mithril Medicine, a one-person biotech whose founder used Claude to organize contracts, logistics and experiments across 11 outside labs. Anthropic says the program went from an idea to promising results in mice in under 18 months. The mouse study was a small, unrandomized pilot that its authors say needs a blinded repeat. What it shows more clearly is the overhead: the founder made the calls, and the model did the coordination that used to take a team.
The second half shows up in the early record of AI-discovered drugs. In one 2024 analysis, 21 of 24 of them made it through Phase 1, well above the usual rate. In Phase 2, where you learn whether the biology was right, four of 10 did, about the industry average. Ten molecules can’t settle anything, but it fits what I’d expect. AI is getting good at the engineering. The biology is still the hard part.
No model reads biology alone
That’s why I think the number one fallacy in biotech right now is believing a large language model, on its own, will reliably recognize the patterns in biology. Intelligence in drug development is multi-tool and multi-player. Language models are one tool. Cell perturbation data is another. People have to link those systems into one integrated engine and keep refining what comes out. Mithril fits, with a caveat: what Anthropic described is only the language-model part, the coordinating. The experiments ran in 11 outside labs, and a person made the calls.
I tested that view this fall. In a private landscape review, I had 111 AI drug discovery companies scored on public evidence. Of nine kinds of engine, the language-led platforms scored lowest. And only one company passed all four tests of an integrated engine: its own data, human biology, a lab loop that works, and a candidate in the clinic. I won’t name it. Your only clue: it’s a public company, and it cleared three of the four bars with no room to spare.
Research note
Of 111 AI drug discovery companies, one passes all four tests.
I scored each company on four tests of an integrated engine, from public evidence. Most pass one or two.
Tests passed, of 4
Tap or hover a bar.
| Tests passed | All | Own lab | Comp. |
|---|
A company passes a test at level 2 or higher on a 0 to 3 scale. Sources: company sites, papers, filings, trial registries and press, as of 10.14.2026, scored by AI models against a written rubric, with source checks on disputed and top scores. Public disclosure only: a capability a company hasn’t disclosed counts as missing. Not a ranking of companies and not investment advice.
So I’d sort diseases by one question: does most of the cost go to running the program, or to figuring out what to run?
Where the cause is known, point AI at it hard. Rare genetic diseases are the clearest case. One analysis of the Orphanet database counted 6,172 rare diseases, 71.9% of them genetic, most affecting very few people. For a child with a mutation almost nobody else shares, my bet is that most of the remaining cost is operating work, the part that just got cheap. Outbreak viruses belong here too. This year’s Bundibugyo Ebola outbreak in the Democratic Republic of the Congo had no licensed vaccine behind it, and the first trial dose went in more than five months after the outbreak was confirmed. Cheap programs could get candidates ready between outbreaks, while nobody is paying.
Where the biology is unsettled, speed turns into a liability. From 2002 to 2012, 99.6% of the Alzheimer’s drug candidates tested in trials failed. More experiments won’t supply a mechanism nobody has found.
That may be starting to change, and the reason is worth noticing. On Saturday, Eric Topol flagged a new Nature study that found 48 new genetic regions tied to Alzheimer’s risk, about a 60% jump, from roughly 2.8 million people. A high share of them point to the brain’s immune cells, the microglia. I replied by zooming in on one line: 17 of the new genes already interact with drugs approved for other diseases, and mechanisms with human genetic support are about 2.6 times as likely to succeed in the clinic. A real shortlist, once we know which way to push each one. That’s the integrated engine in miniature: the signal came from millions of people’s genomes, not from a language model, and the hard question it leaves, which direction to push each gene, still has to be answered in biology.
Slop, misdirection and the loop
Unsettled biology is also where AI’s side effects cost the most. The first is slop. AI makes it almost free to produce drafts, memos, analyses and hypotheses that look finished and sound confident. The bad ones don’t announce themselves. Someone still has to read each one and decide.
The second is misdirection, and it’s the expensive one. A wrong answer that arrives well formatted and fully reasoned can send a team down the wrong road for months. Where the biology is unclear, a confident wrong hypothesis is the costliest thing AI can hand you.
I saw it in my own review. Four independent “frontier” AI models scored the same companies on the same rubric, and one model’s average was double another’s. What struck me was where they still lined up. Scoring blind, all four largely agreed on the checkable facts: whether a candidate had reached the clinic, whether a company makes its own data, whether it works with human biology. They split on the judgment call, whether a lab loop actually works. On those, people went back to the sources.
The third is the loop. In 1865, William Stanley Jevons pointed out that more efficient steam engines didn’t cut Britain’s coal use; cheaper power found more uses. I think thinking work follows the same rule. Make a draft nearly free and you don’t get fewer hours. You get more drafts.
I’ve been part of large capacity build-outs in this industry several times. I’ve watched too much capacity crush the burn rate at companies with no revenue, and too little crush demand at one that had it. What I took from both: be surgical about how capacity grows. AI is a capacity build-out for ideas. The capacity to check them hasn’t grown with it.
Where I’d start
The fair pushback is that cheap programs still hit expensive walls: trials, manufacturing, payers. A one-person company can coordinate 11 labs. It can’t make a trial enroll faster or a payer say yes. True, and that’s exactly why I’d aim AI first where those walls are lowest: tiny patient groups, known biology, and outbreaks where the hard part is being ready in time.
As for the side effects, I’d start with three habits. Don’t let the tool that produces the work approve it; give a second model, or a second person, one job: find what’s wrong. Budget attention like cash, because you get a fixed amount each day, even one that feels like a marathon day. And be deliberate about the objective. In biotech, or any product work, be ready to stop things on purpose and recalibrate before you burn the time (or the tokens). When a program costs nothing to start, the one kept alive on hope gets easier to keep, and that has always been this industry’s costly mistake.
Speed up inside the guardrails, never outside them. Start with the diseases nobody could afford to try. And if AI hasn’t taken anything off your plate this quarter, you didn’t buy productivity.
You bought a backlog.
Disclosure: I invest in technology companies and may hold positions, directly or through index funds, in companies mentioned in this piece. This is my personal view, not investment advice, and it does not speak for any company I lead.
Sources
- Early progress on accelerating drug development with Claude (Anthropic, October 27, 2026) and the Mithril technical report (October 26, 2026, not peer reviewed).
- Jayatunga et al., “How successful are AI-discovered drugs in clinical trials?” Drug Discovery Today (2024).
- Nguengang Wakap et al., “Estimating cumulative point prevalence of rare diseases” European Journal of Human Genetics (2020).
- CDC health advisory (May 19, 2026); WHO, Ebola vaccines Q&A; MSF (October 26, 2026).
- Cummings et al., “Alzheimer’s disease drug-development pipeline: few candidates, frequent failures” Alzheimer’s Research & Therapy (2014).
- Uffelmann, Wightman et al., “Alzheimer’s disease genetics converges on microglial and neuronal dysregulation” Nature (2026), and the preprint (medRxiv, 2025), which counts 17 new-locus genes that interact with approved drugs; Eric Topol on X (October 31, 2026).
- Minikel et al., “Refining the impact of genetic evidence on clinical success” Nature (2024).
- W. S. Jevons, The Coal Question (1865), chapter 7.
- Figures and the widget: a private landscape review of 111 AI drug discovery companies, public sources as of October 14, 2026.
Cover: illustration by Torpedo Publishing, in PD brand colors. Figures: Torpedo Publishing.
