Science Lead @temple. Biomed Scientist in #longevity. Mentor & Investor in Health startups. Past CTO @BioVivaScience, Researcher @CASMIORG, Founders Health.
A weekly jab in the belly is generating more revenue than the entire AI industry.
Ozempic + Mounjaro: $71B in 2025.
OpenAI + Anthropic: $29B.
And they've barely started. ~2% of the 800 million eligible patients can currently access them.
h/t @DrSamuelBHume
America's dementia map looks a lot like its stroke map. In Medicare data, seven of the 11 stroke belt states, where strokes and heart disease are highest, also rank in the top 15 for dementia diagnoses.
For years, tau protein tangles has gotten much of the blame for Alzheimer's. But in nearly half the patients in a new Rome study, tau looked normal.
In them, the brain's energy loss tracked a leakier blood-brain barrier, the seal around the brain's blood vessels. The study is small. The finding rests on 30 patients at one clinic.
Bottom line: both the map and the scans point to the blood vessels that feed the brain. And high blood pressure is a risk you can treat today.
Sources and Notes
Moraleda, Sackstein et al., Cell 2026, doi.org/10.1016/j.cell.2026.… (PMID 42727576, trial NCT02566655). It was a first-in-human, single-arm, open-label trial. Each woman got one IV infusion of her own marrow stromal cells, coated by an enzyme with sialyl Lewis X so they bind E-selectin on marrow blood vessels.
Group fractures went from 8 a year before to 0.5 a year in the 2 years after (94% fewer, p = 0.018, Fig 2B). Over the whole follow-up of 5.3 to 7.3 years, the rate went from 0.54 to 0.11 fractures per patient-year, about 80% fewer. Two women fractured after the infusion. They had seven fractures in all, and the authors tie one woman's first break to her stopping denosumab.
Biopsy bone tissue area rose about 70% at 4 months and a trabecular score at the wrist rose 23% at 2 years. DXA density at the spine and hip did not change.
There was no control group, and the women joined after recent fractures, so some fall in fractures was expected anyway. Of 13 recruited, 3 were dropped. Their cells failed quality checks. 7 of 10 took osteoporosis drugs, with changes during follow-up, and the cells were not tracked in patients. Two hospital stays were judged unrelated to the cells. Brigham and Women's assigned the rights to the sugar-editing method to Robert Sackstein, the senior author.
The University of Murcia says a larger trial is planned to start in 2027. um.es/web/sala-prensa/-/una-…
Ten women with advanced osteoporosis got one drip of their own stem cells. In the next two years, their fractures dropped 94%.
Before treatment, the group broke 16 bones over two years. Afterward, they broke just one. Eight of the ten went more than five years without a new fragile-bone fracture.
That matters because stem cells from bone marrow can help build bone, but when they are dripped into a vein, they struggle to find it.
So a team in Murcia, Spain, and Miami gave each woman's own marrow cells a sugar tag. The tag binds to a docking protein on blood vessels inside bone, acting like a postal code for the skeleton. No genes were changed, and no serious side effects were linked to the cells.
The women were 51 to 72. Some researchers had questioned whether stem cells from older, diseased bone could still help.
This was a small, uncontrolled study of 10 women. Seven were already on osteoporosis drugs, with the cells added on top. The senior scientist invented the method and holds its rights.
A larger trial is planned for 2027.
Bottom line: older women's own marrow still held cells good enough to help their bones, and one drip was associated with years of fewer fractures.
AI-designed drugs pass their first safety trials 78% of the time, against 52% for drugs overall. When a drug has to treat the disease, they pass about 20% of the time, against 29%.
Those are Phase 1 and Phase 2 results for 129 AI-designed drugs that entered human trials by March 2026, set against an industry review of 2011 to 2020, and only 24 of them have finished Phase 2 so far.
Small numbers, but the right question.
Phase 2 is where most drugs fail, whoever designs them. Isomorphic Labs raised $5 billion today to design drugs this way, and its own drugs are still heading toward their first human trials.
Bottom line: AI is getting drugs safely into people, and the next few years of Phase 2 results will show whether those drugs treat disease better.
Thrilled to announce $5B in funding for @IsomorphicLabs to accelerate our mission to Solve All Disease. We continue to make rapid progress reimagining drug design from first principles with AI to radically cut the time & cost to find new medicines. Gradually, then all at once...
Diets and Ozempic shrink fat cells. They do not remove them. This shot aims to kill them at the injection site, and it is now in final-stage trials.
Caliway says its injection, CBL-514, triggers fat cells to self-destruct. In a 76-person trial published last year, 16 of the 23 treated people who were measured 8 weeks after their last shot had lost at least 150 mL of fat under the belly skin, about two-thirds of a cup. None of the 16 people on placebo did.
About half the trial did not finish it.
Kybella, approved in 2015, already dissolves fat under the chin.
Bottom line: this is a shot for one or two stubborn patches of fat in people who are not obese.
This is fascinating.
We have an injection in Phase 3 that kills fat cells right where it goes in.
It's called CBL-514, and it makes fat cells self-destruct (apoptosis) instead of just shrinking them.
This is interesting because your fat cell number is roughly the same throughout your entire life in both lean and obese people. You can diet and be on GLP-1s to make these cells smaller, but they stay.
In the Phase 2 trial, 69.6% of people lost at least 150 mL of belly fat 8 weeks after the last treatment (compared with none on placebo).
It's not approved yet, and 320 people are currently going through a Phase 3 trial.
Frankly, I think a lot of people who’d never consider liposuction would consider this drug.
A cheap heart pill from the 1960s can blunt the fear in a bad memory. You take propranolol, then deliberately recall the memory. In effect, the fear comes back less attached.
A recalled memory can stay unstable for a few hours before the brain stores it again. Noradrenaline, a cousin of adrenaline, helps write the fear back in. Propranolol blocks that signal, which is also how it slows a racing heart.
Rats showed the effect first, in a New York lab in 2000 and with propranolol in 2004. In 2009, a small Amsterdam experiment wiped out the startle response to a picture healthy volunteers had learned to fear the day before, even though they still knew what it meant.
PTSD is harder. In a 2018 Montreal trial of 60 patients, six weekly sessions of propranolol before reliving the trauma beat a placebo. A 2021 French trial of 66 used the same plan and found no difference.
Bottom line: propranolol can soften the fear in a memory without deleting it, but for PTSD it is still an experiment.
Several of you have asked: Does Metformin also affect strength? Possibly, but the trials are too small to say.
After 14 weeks of lifting, older adults on the dummy pill got 23% stronger in their legs. Those on Metformin got 15% stronger. The trial was designed to measure muscle size, not strength.
In a separate 12-week aerobic trial of 53 older adults, fitness improved by about half as much on Metformin.
Neither difference was large enough to rule out chance (p = 0.055 and 0.08). Both point in the same direction as the loss of muscle.
PMID 31557380, PMID 30548390
Two big heart trials in @NEJM this month show how hard it is to prove a drug works.
The drugs went after IL-6, a driver of inflammation, and lipoprotein(a), an inherited cholesterol particle. Both are suspected causes of heart disease. In more than 14,000 people followed for 3 to 5 years, both drugs lowered what they were aimed at, and neither cut major heart events such as heart attacks and strokes.
Bottom line: AI will change medicine fast where a computer can check the answer, and slowly where only a trial in people can check it. Drugs to slow aging sit at the slow end, because their benefits take years to show.
(8/8)
x.com/EricTopol/status/21191…
Drug development is the slowest of all.
@a16z, the venture firm, estimates that a new drug takes 13.5 years from the first lab work to launch. Eight of those years go to trials in people, and only about 4 in 100 drug candidates make it to launch.
Drugs found with AI face the same trials. By March 2026, 129 of them had reached human trials. By the researchers' count, about 1 in 5 made it through Phase 2, the trial that asks whether a drug works.
(7/8)
x.com/a16z/status/2121290755…
Developing a drug takes years, and most candidates never reach patients.
Amortizing the cost of failed drugs across stages, the present value cost of developing a drug is close to $2B. The process takes 13 years on average, with rising clinical trial costs putting further pressure on R&D productivity.
AI could improve those economics by increasing the probability of success and reducing development time, from identifying promising molecules to designing clinical trials and improving patient recruitment.
That may enable companies to pursue therapeutic targets and indications that might otherwise have been uneconomical.
Full piece on why AI will make pharma even bigger: a16z.news/p/big-pharma-can-r…
Lab work is slower, because each new answer needs an experiment, and experiments can take days.
AI is getting better at making sense of the results. On @OpenAI's new test of 100 routine problems in reading lab data, its two best models score 62% and 67%. Its older GPT-5 models scored between 31% and 44%.
(6/8)
x.com/OpenAI/status/21187827…
A new way to measure how AI agents handle the routine analyses of wet-lab science—BioBasicsBench.
Developed with Ligo, a biotech company focused on enzyme design, the benchmark draws on 100 tasks from real experiments to test whether agents can interpret raw lab data, identify errors, and recognize when the evidence doesn’t support a conclusion.
Accelerating biological research starts with getting the everyday work right. BioBasicsBench shows where models need to improve, and helps us track what those gains mean for scientific progress, biosafety, and biosecurity.
openai.com/index/measuring-h…
Diagnosis is changing now, because AI can practice on past cases where the right answer is known.
One case written up in @NEJM shows it. A woman had lived with stomach pain for 10 years. Her doctor asked a chatbot, and largely on the basis of its answer named the cause and started a treatment that worked.
(5/8)
x.com/EricTopol/status/21216…
A 37-year-old woman with 10 years of abdominal pain.
—The AI makes the diagnosis and recommends a successful treatment.
—The physician "diagnosed a patient with a condition I had never diagnosed before and started treatment largely on the basis of the output of a chatbot."
In today's @NEJMnejm.org/doi/full/10.1056/NE…
.@dwarkesh_sp says AI research is the next field to change, because a computer can also check AI experiments. His best guess is that AI matches the best AI researchers by 2028.
A footnote in his essay shows why speeding up most of a job is not enough.
Say a job takes 100 hours, and AI does 90 of those hours 50 times faster. Those 90 hours shrink to under 2. The other 10 stay. The job now takes about 12 hours, so it is about 8.5 times faster overall.
In medicine, the part that does not shrink is trials in people.
(4/8)
x.com/dwarkesh_sp/status/212…
Among the sciences, math changed first, because a computer can check a proof line by line.
Navier-Stokes is a famous prize problem about how liquids and gases flow. @OpenAI's AI wrote a 165-page proof for the version where an outside force pushes the fluid, and a proof-checking program passed it.
Terence Tao, one of the world's top mathematicians, says no expert he has talked to understands the whole argument. For now, the proof is trusted because a computer checked it.
(3/8)
x.com/matharium/status/21198…
Terence Tao says that the AI proof on the Navier–Stokes problem is correct, yet mathematicians learned almost nothing from it. The proof is 100 pages, AI-generated and full of issues in its text, but a formal certificate in the language Lean confirms it.
Navier–Stokes is a Millennium Prize problem that asks whether the equations for fluids like water can start smooth and then blow up, with velocity or pressure becoming infinite. Human mathematicians came very close, and OpenAI ran its biggest model, with millions of dollars of compute, to get there first.
A human proof normally comes with talks on what was tried, what worked and what it connects to, like picking fruit and planting the seeds. Tao calls this one strip mining: the fruit taken, the landscape left barren. One colleague, using lots of AI, could explain only half of it after days of work.
The tests built to measure AI show the same pace.
@emollick tracked the 173 tests AI labs quote most, and AI now scores 85% or higher on 63% of them, which his tracker counts as beating a test.
Newer tests are beaten sooner. Tests released in 2023 and 2024 lasted about 20 months before AI beat them. Tests released in 2025 lasted about 14 months.
(2/8)
x.com/emollick/status/212175…
AI is getting better faster than our measurements are
I threw together SatBench, which measures how quickly major AI benchmarks become saturated (effectively beaten). Of 173 benchmarks cited by the Labs since 2023, 63% are now saturated. Saturation is happening faster each year.
Two numbers show where AI will change medicine and longevity fast, and where it will take years.
The first is under two months. That is how fast the length of task AI can do on its own now doubles, by one estimate built on @AnthropicAI's own test scores. It used to take about seven months.
The second is 13.5 years, the time a new drug still takes to reach patients.
AI improves fastest where a computer can check the answer, and slowest where the only check is a trial in people. So math has already changed, AI research is next, diagnosis is changing now, lab work is slower, and drug trials are slowest.
(1/8)
x.com/ben_j_todd/status/2122…