
Founder & CEO @becoming_ai Developmental Intelligence - AI for predicting how organisms change over time.
San Francisco, CA
Joined April 2009
- Tweets4.7K
- Following904
- Followers2.4K
- Likes4.7K
Pinned Tweet
After 3 years of quiet building, we are finally sharing that this is now tractable via an engineered placenta. But the implications go far beyond.
I hope this gives you a feeling of ‘awe’ that we feel every day doing our life’s work.
Incredibly grateful to the amazing team we have that enabled this and continues to push the frontier!
Divya Dhar Cohen, MD/MBA/MPA retweeted
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...
Divya Dhar Cohen, MD/MBA/MPA retweeted
Thank you, Neil Mehta for this investment in San Francisco’s kids and their futures.
Eligible families with children ages 2–10 can receive a $100 deposit into an account to help them build a strong financial future.
Learn more here⬇️
sfstandard.com/2026/11/16/ne…
Divya Dhar Cohen, MD/MBA/MPA retweeted
ACE Music v1 just hit #1 on @ArtificialAnlys' music leaderboard. It's the first model on this chart to outrank Suno's latest release.
Try it free at acemusic.ai.
We'll open-source the model weights in the next few weeks. We're making the final tweaks so you can run it locally.
I do this with babies. Lot of people share their baby news with as among the first few people to know. Not sure how I got this reputation but I'm genuinely SOOO happy for them. And can't wait for future play dates.
Underrated life advice: Be someone people want to share their good news with. Get excited. Hype them up. Make a big deal out of small wins. And never compare it to something you or others have done. We always remember who made the happy moments feel even happier. Be that person.
1
6
785
100 agree. The moat is knowing the actual problems patients have.
physician founders are on a tear this year.
@Roche agreed to acquire Andrew Beck’s @Path_AI for up to $1.05B. @R1RCM acquired Jeremy Friese's Humata Health.
yesterday, @ouraring agreed to acquire Muthu Alagappan's @CounselHealth for $625M in cash.
the standard critique of physicians as founders is that "clinical training breeds caution." correlation isn’t causation, but what I take from these outcomes is that when AI makes knowledge free, there is a tremendous moat in hard-won human experience.
never been a better time to be a physician founder.
1
4
474
Divya Dhar Cohen, MD/MBA/MPA retweeted
There is a problem that these people neither understand nor grasp because of lack of knowledge or biology. Unlike molecular biology, Cell and tissue biology/phyisology exhibits emergent behaviours that we're just starting to map and that AI can't deal with. Fo now.
most doctors i know don't do random experiments on their bodies with untested drugs.
but the non-md medtechbros, they love to call it biohacking.
either the doctors know more or they are very risk averse. possibly both. worth noting before you listen to medtechbros.
179
remember those days when tech was 'do no evil' and most people including non-techies thought it was hopeful.
1
185
Divya Dhar Cohen, MD/MBA/MPA retweeted
Bruce Booth, socio de Atlas Venture y uno de los inversores en biotecnología más respetados de Estados Unidos, publicó ayer una predicción que no va a gustar a los entusiastas de la IA en farmacia. Durante los próximos tres a cinco años, la inteligencia artificial hará que la industria sea menos productiva, no más.
¿Por qué? Los modelos aprenden de los fármacos que ya existen. Son muy buenos diseñando moléculas nuevas contra dianas que ya conocemos y siguen siendo flojos en lo difícil, la biología que nadie ha descifrado. Según Booth, más de 100 programas en el mundo persiguen cada una de las ocho dianas más populares, entre ellas PD-1, HER2, KRAS y GLP-1. Al menos otras 30 dianas tienen más de 30 programas cada una. Dos biotecnológicas nacidas con IA que salieron a bolsa este año, Generate e Iambic, trabajan sobre TSLP y HER2. Quizás mejores. Nuevas, no.
Esto ya pasó. En 2001, en plena euforia del genoma, Lehman Brothers y McKinsey publicaron un informe que avisaba de una indigestión: demasiadas dianas nuevas sin validar, más fracasos, más costes y más años. Acertaron, y la productividad de la I+D farmacéutica cayó durante buena parte de este siglo. Nadie diría hoy que el Proyecto Genoma Humano no valió la pena.
Lo que más me preocupa es lo que viene después. Una de las mejores habilidades de la IA es leer una patente, adivinar la molécula principal y diseñar otra que la esquive. Booth aconseja a los fundadores con ciencia de verdad nueva patentar más tarde, publicar menos y hablar poco. Eli Lilly, la mayor farmacéutica del mundo por valor en bolsa, dejó hace poco de publicar su cartera de fase 1.
La medicina avanzó porque publicar compensaba. Si copiar se vuelve instantáneo, lo racional es esconder. Eso no lo arregla la tecnología, lo arregla la ley. Booth sugiere incluso que Estados Unidos vuelva a dar la patente al primero que inventa, el sistema que abandonó en 2013.
Más moléculas no significa más medicinas.
Incredible foresight and execution @martinvars
Gameto my sixth unicorn was just sold to Ferring for $1.38bn here is the story.
In 2020 I endowed a fellowship at Harvard in my father's name, Carlos M. Varsavsky, in the lab of George Church, one of the world's leading geneticists. Its goal: to make human eggs in the lab. Gameto grew out of it.
My father was an astrophysicist and built the Instituto Argentino de Radioastronomía, Argentina's radio astronomy institute. He was a great teacher, at the university and at home. In a time without artificial intelligence, he was the artificial intelligence for me: he knew about everything. He died young. The fellowship carries his name. The fellowship paid for the PhD of Christian Kramme, a Californian geneticist who worked out how to make in the lab the cells that surround an egg in the ovary and help it mature. He is now Gameto's chief scientific officer.
Ferring Pharmaceuticals, the Swiss fertility drug maker, is acquiring Gameto for up to $1.3825 billion. I am a co-founder, and this is the story of a company that set out to do one thing and was bought for another.
I came to fertility in 2016 with Prelude Fertility, a network of US clinics I built, now part of Inception, the largest fertility network in North America. Inside those clinics I learned that IVF works, but it is brutal. For two weeks a woman injects herself with hormones so her ovaries release many eggs at once: around 28 shots, ultrasounds, blood tests, bloating and mood swings.
Many women go through it once and never again. Every one who stops is a pregnancy that never happens. One question obsessed me: what if the treatment took three days instead of two weeks.
I brought in Dina Radenkovic Turner as co-founder and CEO. She was 26, a physician trained at University College London and one of the brightest people I have ever worked with. Many questioned the bet on someone that young. Time proved us right. I had a moonshot and a fellowship. Dina turned them into a company. She looked at Christian's cells and saw a product: if the egg was out of reach, we would do the ovary's work in the lab. That became Fertilo.
With Fertilo a woman takes hormones for two or three days, about 80% fewer shots. Her eggs are collected immature and spend a day in the lab with those cells, which do the ovary's job. Then they are fertilized as in any IVF cycle. The cells never enter her body.
On December 16, 2024, the first baby conceived this way was born in Lima. By August there were five babies and more than twenty pregnancies. Three gentle cycles beat one brutal cycle nobody wants to repeat.
With Dina at the helm, Gameto raised about $127 million in total from Overwater Ventures, Insight Partners, RA Capital, Two Sigma Ventures, BOLD Capital Partners, Future Ventures, Olivia Walton's Ingeborg Investments, Arcadia Investment Partners, PAGS Group, Pontiva Healthcare Partners and Portfolia, and from Anne Wojcicki of 23andMe, Bob Nelsen of ARCH Venture Partners, Stacey Bendet Eisner of Alice + Olivia and Chelsea Hirschhorn of Frida.
Gameto is the sixth company I have built to pass a billion dollars, after Viatel, Yacom, Jazztel, Eolia Renovables, sold in 2017, and Inception/Prelude Fertility, Telecoms, internet, clean energy, clinics, biotech: nothing in common, and in every one the team mattered more than me.
1
239
I love that some of our investors are pregnant themselves. And still investing in the best companies.
Having babies makes you more hungry. More ambitious. More likely to want to build things of consequence.
5
197
Divya Dhar Cohen, MD/MBA/MPA retweeted
Functional recovery of vitrified and nanowarmed human-scale pig kidneys
biorxiv.org/content/10.64898…
Divya Dhar Cohen, MD/MBA/MPA retweeted
有点震撼。。。我们可能真的快要搞清楚大模型内部到底是怎么运作的了。
OpenAI早期机械可解释性研究的参与者,昨天突然提出了一个很大胆的结论:
我们可能一直误解了神经网络。
昨天, @nickcammarata 发了一篇很有意思的文章,叫《Neural Networks Are Already Clean》。
这哥们早年就在OpenAI做机械可解释性研究,2023 年还牵头启动了@OpenAI 第一个大规模自动化神经元解释项目。
当时他们发现,现有工具只能解释大约 2% 的神经元。
Nick因此得出了一个很悲观的判断:神经网络内部存在大量纠缠,想搞清楚它们到底在干什么,可能非常困难。
结果三年后,他带着更先进的工具和Agent重新研究,发现情况完全不同。
在他们测试的现代视觉模型中,最后一层 99%~100% 的神经元都能找到相当清晰的解释。
甚至在Qwen3-32B这样的语言模型中,初步研究也发现,约90%的神经元都有一个可以识别的主要功能。
最有意思的是,他还重新研究了过去的一些经典案例。
比如曾经有一个神经元,既会对猫脸产生反应,也会对汽车前脸产生反应。
以前大家觉得,这是典型的Superposition,多个不相关的概念纠缠在了同一个神经元里。
结果重新研究后发现,这个神经元可能只是在识别一种共同的V形视觉结构。
猫脸和汽车前脸恰好都有这种特征。
全世界的家人们折腾了这么多年,可能连最基础的观察方式都有问题。
而且我觉得这篇文章最值得关注的,其实是他们使用 Agent的方式。
他们让Agent自动观察神经元、提出功能假设、设计实验、分析电路,再通过新的输入验证假设,甚至自动审查了10 万张神经元归因图。
这其实也是我最近一直关注的一个方向:让 AI 自己去做科学研究,把大量观察、假设、验证的工作自动化。
过去需要研究员花大量时间逐个分析的神经元,现在已经可以让Agent成规模地开展研究了。
当然,目前这些结果还处于早期阶段,距离完整解释大模型的内部机制,还有很多工作要做。
但我觉得Nick提出的这个观点非常有意思:
神经网络内部可能一直都很有规律,只是过去我们的工具太原始,根本看不清楚。
如果这个方向最终被验证,机械可解释性领域过去很多基础假设,恐怕都需要重新审视了。
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Amazing work Nabiha and Marina. Super useful.
In partnership with Vinod Khosla (@vkhosla), I’m proud to introduce Reservoir.
Billionaires have owned their cells for years. Starting today, you can too.
Sign up today: withreservoir.com
Divya Dhar Cohen, MD/MBA/MPA retweeted
A good example of why open weight and open source models are important - they permit users to fine-tune specific use cases with greater cost efficiency. As has been the case in other technology markets, these open models will drive lower costs for all consumers and enhance security.
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Divya Dhar Cohen, MD/MBA/MPA retweeted
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Divya Dhar Cohen, MD/MBA/MPA retweeted
There’s a strange paradox to living with an abundance mindset: sometimes having too much to give becomes painful when you have nowhere to put it.
Love, curiosity, generosity, wonder. They accumulate, looking for somewhere to go.
I’ve discovered that abundance can create its own peculiar form of suffering. The irony isn’t lost on me.
I’m working through this in my mindfulness practice. Learning that not everything I have to give must be given immediately. Sometimes abundance is something to sit with, not something to spend.






















