This framing assumes over prioritising safety has no downside. It means telling a stage 4 cancer patient they can’t take a drug that passed a phase 3 trial, and must die of their cancer instead, while experts debate how it works. It’s their body and their choice. For that patient the choice isn’t wait safely vs take a risky drug. It’s die waiting vs a chance of living. That’s why patients already inject untested compounds at unregulated clinics in Mexico. Millions die while we wait and nobody gets blamed. If 100 people died from an AI designed drug it would be called a catastrophe, even though most of them would have been dead within months anyway. Patients know new drugs can go wrong. The alternative is extreme suffering or death. We aren’t impatient because we want the next shiny update, we’re impatient because many of us will die before it arrives. Drug regulation is already too strict. With AI now moving biology this fast, the priority should be loosening it, not tightening it. Right now a patient who will die from their disease in a few months, or take their own life because of the suffering, usually can’t get access to promising drugs still in trials or try combinations that might work better together but carry more side effect risk. They just have to die. If we were being logical, those patients could enroll in a program, try the drug or any combination they want, and record the outcome so it feeds back into the science. Relying on slow, admin heavy trials with tightly selected groups as the only way to learn about new treatments is the least efficient and least ethical option we have. It exists to protect pharma and regulators from blame, not to reduce suffering and death. Now that some people could realistically be cured in the next decade instead of waiting 20 years for trial funding and long term outcome studies, we should be moving away from it fast. For weight loss or baldness, be as rigorous as you like. Framing this around stage 4 cancer is detached from what patients actually face. The people debating it don’t pay for the delay, the patient does.

As a patient, I would inject any substance on earth if it had passed a phase 3 clinical trial for my disease successfully. I’ll go even further. I’d probably inject it if it had positive phase 1 signals. I don’t need anyone to know why it works, I just need to not be ill.

The simplest way to understand what's wrong with the letter is to swap their words for oncology. "The push by AI companies to cure cancers as a benchmark is detrimental to the science of oncology. For students we often suggest cancers to work on with the intention of developing their skills. Curing cancers is only a proxy for the primary goal of understanding tumour biology. The mass production of cures at a faster and faster pace could destroy fertile ground." You could argue that isn't a fair switch, but they concede in the letter itself that maths is the basis for progress in science and technology. Maths feels abstract, but its outputs flow straight into engineering and medicine. There's just enough distance between the two that the connection stops feeling attributable. I do understand where they're coming from. Spending years inside a problem, then the talks and the arguments and the slow process of a result becoming something a student can learn from, is a wonderful way to live. But if AI can solve the maths, and then use that maths to solve disease, then a preference for the culture of a mathematical profession doesn't hold up against the millions of people who die from delay. If the AI can go the whole way, we should let it go the whole way, and then take our time understanding what it did afterwards, for the intellectual pleasure of it, without slowing it down (unless for safety reasons).

I'm a patient. Whether I'm alive in ten years depends on how fast science moves, and nothing moves it faster than AI. Chasing a problem for the understanding, not the answer, is beautiful. It's also a preference held by people whose lives don't depend on the answers arriving. I think every scientific field is about to face a question it has never had to say out loud. Is the work for the people outside the field whose lives it changes, or is it for the people inside it, who get to spend their lives on the most satisfying kind of thinking there is? For most of history you never had to choose, because the satisfaction drove the work and the benefit came along behind it. When you're the patient the two come apart, because the benefit stops being an abstraction and becomes you. Every year a cure arrives later, people who would have lived instead die. We don't really count them, because we're very good at separating deaths from doing and deaths from not doing. Clinical trials are the clearest example. Keeping a risky combination therapy away from a dying patient protects them from a death by action while quietly accepting a death by inaction, and we weight the first heavily and barely count the second. Slowing AI down works the same way. Pushing the cure for most major diseases from the next decade to the one after is not a careful default. It kills people, it's just that nobody has to sign their name to it. Banning frontier AI in any meaningful capacity would be a decision to let millions of people die on schedule. The better path is infrastructure, meaning verified access for good actors, government capacity to audit what the leading systems are doing, and real consequences for labs that cut corners. On the letter itself, the mathematicians are right about the specifics. A proof announced in 88 hours with no proper writeup and no credit to the work it stands on is a bad way to do mathematics. But that's an output problem rather than a speed problem. Require the model to cite prior work and write the full argument and it's fixed within a quarter. What the letter is really protecting is a culture, and I understand that, but I still think the weight falls on adaptation rather than protection. Traditions built over centuries matter a great deal and they still aren't worth millions of lives. The job for mathematics now is to become the field that turns machine-generated results into ideas humans can actually understand and use, which is still mathematics, and might be the most important version of it. If the machines eventually get better at that too, I'd look at retirees. Most go through a period of agitation and then slowly lose the belief that meaning requires a nine to five. They keep reading, keep learning and keep arguing about ideas, they just stop doing it for a paycheck. I don't think that's a tragedy. I think it might be where all of this ends up. So for anyone weighing how fast this should go, put the hidden deaths from delay next to the visible risks from action every time. It's the only way to account for everyone with a stake in this, including the ones who won't be around to complain if we get the timeline wrong.