The Week the Impossible Quietly Stopped Being Impossible
There was a week, late in April of 2026, when a number of long-standing walls came down without a parade. A twenty-three-year-old asked a chatbot for help with a sixty-year-old number theory problem and got a real answer. An algorithm that had stood since the moon landings was beaten by a piece of software that designs other software. A laboratory in Cambridge taught a machine to do the scientific method on its own. A child who would have grown up deaf can hear her mother. A trial that nearly ended started over. None of these things, taken alone, would alarm anyone. Taken together they are a kind of weekly report on what repair looks like now.
A young man with no advanced math training opened a chatbot and asked a question. His name is Liam Price, he is twenty-three, and the chatbot was OpenAI's GPT-5.4 Pro. The question concerned Erdős Problem #1196, a sixty-year-old conjecture from Erdős, Sárközy, and Szemerédi about primitive sets — sets of whole numbers in which no number divides any other. After roughly eighty minutes of internal reasoning, the model returned a proof. It did not brute-force its way through the problem. It introduced a Markov-chain construction with von Mangoldt weights — a tool that lives in a neighboring corner of number theory and that, as Oxford's Jared Lichtman put it, no human had thought to bring across the road. Lichtman called it "the first AI result at the level of Erdős's Book." A small footnote: an AI researcher had gifted Price a ChatGPT subscription and called what he was doing "vibe-mathing." Vibe-mathing now has a citation.
For fifty-six years, the answer to a particular question in matrix multiplication was Strassen's. Volker Strassen had shown in 1969 that you could multiply 2×2 matrices using seven scalar multiplications instead of eight, and the recursive ghost of that result had set the upper limit on every machine-learning chip ever built. This month the wall moved. Google DeepMind's AlphaEvolve, a Gemini-powered coding agent that writes and tests its own algorithms in a loop, found a way to multiply 4×4 complex-valued matrices using forty-eight scalar multiplications instead of forty-nine. One multiplication. The kind of improvement that sounds like nothing until you remember that every training run, every inference, every model anyone uses sits on top of it. The same agent was set to work on more than fifty open problems in analysis, geometry, and combinatorics, and improved the best known answer in twenty percent of them. It is, in a phrase Lewis Thomas would have liked, a kind of cell that has just learned a new trick of repair.
Three million dollars and six months, against two-point-one billion dollars and a decade. That is the comparison Lila Sciences likes to make about its CAR-T cell therapy program, and it is worth dwelling on because the numbers do not look like a typo. Lila is a Flagship Pioneering company — the same firm that built Moderna — and it is led by Geoffrey von Maltzahn. Its premise is that the bottleneck in modern science is not ideas but cycles, and that an AI that runs the scientific method autonomously, around the clock, across chemistry, materials, and biology, will compound knowledge the way Charlie Munger said money compounds. They have closed a Series A of $235 million on top of an earlier $200 million. Their mRNA program is reporting expression that lasts fifteen days where conventional approaches manage one and a half. Three hundred thousand designs versus thirteen. We have argued for years over whether AI would help science. The question is now what to do with a colleague who never sleeps.
Fifty babies a year. The kind of number that does not get a press conference of its own. They are born in the United States with a particular mutation in a gene called OTOF, and until this month they would have grown up deaf or nearly so. On April 23, the Food and Drug Administration approved Otarmeni — Regeneron's lunsotogene parvec-cwha — the first-ever gene therapy for a form of genetic hearing loss. The molecule is a dual adeno-associated virus carrying a healthy copy of the OTOF gene. The procedure is a small surgery, of the sort already done for cochlear implants, but it leaves nothing behind in the ear. What is uncommon about this approval is the speed: sixty-one days from the filing of the Biologics License Application to clearance, under a new program called the Commissioner's National Priority Voucher. The drug will be offered at no cost to eligible families. The babies, of course, do not know about any of this. They will simply find, at some point in their first year, that they can hear their mother's voice.
The story closes with a trial that almost ended. In October of 2025 the FDA placed a clinical hold on Intellia Therapeutics' MAGNITUDE program — the company's Phase 3 in-vivo CRISPR trial for transthyretin amyloidosis, the slow, fatal disease in which a misfolded protein silts up the heart and the nerves. A patient had developed Grade 4 liver enzyme elevations. Most companies, most drugs, do not come back from a hold like that. The agency lifted the hold on the polyneuropathy trial in January and on the cardiomyopathy trial in March, with new safety monitoring in place. Long-term Phase 1 data continues to show deep, durable reductions of the offending protein from a single dose. We have seen this pattern before in biology — a system bumped off course, a reasonable correction, a return to the road. The therapy may yet fail. What matters today is that it is allowed to keep trying.