ATX - Health AI, AGI, longevity 🛸 CEO Health Universe (Kleiner Perkins) 🚀 Co-Founder @arrivehealth (Acq Bain Capital) 🤖 Hacked insulin pump. FDA approved.
Latent Space
Joined December 2008
- Tweets881
- Following858
- Followers444
- Likes5.3K
Dan Caron⚡ retweeted
If you’re a physician sitting across from a patient and you hear this:
“My LLM suspects I have XYZ diagnosis based on these symptoms, labs, and so on…”
What is your reaction?
Dan Caron⚡ retweeted
Enzyme design just got fully automated🧪🤖
AI protein engineering is great but you still have to synthesize the designs in the lab, get data, and update your model.
The latest work from @cobanbrooks, @NotinPascal, and @romerolab1 changes this.
📝Paper:
biorxiv.org/content/10.64898…
I’m really excited to share our new preprint, out today, where we built a fully autonomous system for enzyme engineering — integrating a self-driving lab with a generative protein language model that worked together over a month to engineer substrate specificity in glycoside hydrolase (GH1) enzymes.
biorxiv.org/content/10.64898…
[1/n]
Dan Caron⚡ retweeted
This is probably the best paper I have read about causal reasoning for quite some time. Really a great weekend read!
"Causal Persuasion" (Burkovskaya & Starkov) models how much evidence you need to establish vs. rule out a causal link. The result is stark:
To prove X causes Y: 1-2 well-chosen variables often suffice.
To prove X does NOT cause Y: you must account for every possible common cause. Arbitrarily many confounders. Practically unfalsifiable.
This inverts the Humean intuition: in causal reasoning, positive claims are cheap to sell and negative ones are almost impossible to rebut.
Now think about what this means for Virtual Cell models.
Most perturbation datasets cover a thin slice of the combinatorial space — a few hundred gene knockouts, maybe a few contexts. A model trained on that data can confidently "learn" gene X drives phenotype Y. But if the true structure is X←C→Y , and C was never systematically varied — the model will never see its own confounding. It has no mechanism to distinguish causal signal from correlated noise.
The paper formalizes exactly why: the model is a sophisticated receiver that accepts whatever causal story is consistent with the data it's seen. And if the data omits the right confounders, even a "sophisticated" model is manipulable.
This is the deepest argument for perturbation diversity. Not just more data, but also more axes of variation. Vary the context. Vary the genetic background. Vary the timing. You're not just collecting samples; you're systematically eliminating alternative causal explanations.
This is why we need “scale” the training data with more contexts including cell types, spatial, and temporal variations.
Paper: aburkovskaya.com/pdf/causali…
Dan Caron⚡ retweeted
Federal funding for US biomedical research is moribund.
Since October 1 2025, NIH is -80% in new grants and -70% in values (total dollars).
Labs are closing down and researchers are leaving science.
To what end?
Buckle up!
History just happened.
An AI system has autonomously solved a long standing open conjecture in pure mathematics (Fel’s open conjecture on syzygies of numerical semigroups) building a complete formal proof in Lean with zero human input.
This is the first time in history an unsolved theory level math problem has fallen to AI.
I built some causal models directly from LLM data earlier this year. People have no idea how powerful this technique is. There are only a handful of folks in the world thinking about this rn. Huge opportunity for those that understand it. Will definitely change the world.
Holy shit… this paper might be the most important shift in how we use LLMs this entire year.
“Large Causal Models from Large Language Models.”
It shows you can grow full causal models directly out of an LLM not approximations, not vibes actual causal graphs, counterfactuals, interventions, and constraint-checked structures.
And the way they do it is wild:
Instead of training a specialized causal model, they interrogate the LLM like a scientist:
→ extract a candidate causal graph from text
→ ask the model to check conditional independencies
→ detect contradictions
→ revise the structure
→ test counterfactuals and interventional predictions
→ iterate until the causal model stabilizes
The result is something we’ve never had before:
a causal system built inside the LLM using its own latent world knowledge.
Across benchmarks synthetic, real-world, messy domains these LCMs beat classical causal discovery methods because they pull from the LLM’s massive prior knowledge instead of just local correlations.
And the counterfactual reasoning?
Shockingly strong.
The model can answer “what if” questions that standard algorithms completely fail on, simply because it already “knows” things about the world those algorithms can’t infer from data alone.
This paper hints at a future where LLMs aren’t just pattern machines.
They become causal engines systems that form, test, and refine structural explanations of reality.
If this scales, every field that relies on causal inference economics, medicine, policy, science is about to get rewritten.
LLMs won’t just tell you what happens.
They’ll tell you why.
💊 Wild. There’s a new AI model called DrugReflector that uses gene activity data from human cells to predict promising drug compounds — way faster than brute-force screening.
It’s up to 17× better at finding real hits than random library picks. And when it retrains on its first results, the success rate doubles.
Traditionally, drug discovery throws giant chemical libraries at simple cell readouts — slow, narrow, and blind to the full gene network.
DrugReflector flips that on its head: it learns how ~9,600 compounds shift gene expression across 50+ cell types, then asks which ones push a cell’s program toward a desired state.
If a compound moves the gene-expression “signature” toward a healthy target, it’s a strong candidate.
They tested it on blood formation, aiming to boost platelet and red-blood-cell production — and the AI-selected compounds crushed random screening.
The beauty is that it learns general patterns of cellular change, so one dataset teaches it transferable insights for new diseases.
🔗 nature.com/articles/d41586-0…
This seems like meaningful signal.
BREAKING: Google's quantum chip computed a 150-year problem in 2 hours—and proved it.
Nature confirmed today: Willow achieved verifiable quantum advantage, meaning physics validates its answers. At 13,000× classical speed with 99.9% fidelity across 105 qubits, it's not just calculating faster—it's perceiving molecular structures that were mathematically invisible until this morning.
Dan Caron⚡ retweeted
If a simple coffee shop has a bot farm with 100s of phones to amplify their message, please consider what a foreign agency or adversarial operator is running on your favorite social media platform.
Especially today, please consider that the opinions you read, the calls to violence you hear, and the news you digest, are all an operation done to sow hatred in your mind and your soul.
Do not let anyone online turn you into something that you are not. Rationalize everything, take everything with a grain of salt, question everything for yourself, not based on how someone online is questioning it.
Turn off all distractions and focus on the mission you have, not the mission of someone trying to control your mind.
Dan Caron⚡ retweeted
Summary Table of the massive defunding of biomedical research in the United States nytimes.com/interactive/2025…
Excited to pick up Proof by
@adamjkucharski which talks about the pitfalls of p-values, the dangers of loose priors and the difficulty of communicating the efficacy of interventions.
Clearly we need to get better at communicating complex and nuanced interpretations of data, or else simplistic, misinformed views will prevail.
If you're interested in science or the nature of Truth, this is sure to become a must read. Thanks to @EricTopol for surfacing.
Dan Caron⚡ retweeted
DeepMind just dropped a 106-page paper unveiling AlphaGenome.
This single model could completely redefine how we discover disease-causing mutations and drug targets.
This is massive. 🧵
Dan Caron⚡ retweeted
A big day – we are making the largest dataset of chest X-rays available to the world. This is the first dataset to cross 100k patients.
huggingface.co/datasets/rajp…
Dan Caron⚡ retweeted
It's not every day that you wake up to a @Forbes article covering the great work of your team.
We at @tahoe_ai open sourced Tahoe-100M to start a movement in open data in biology. Thanks @Amelia_Palermo for taking note of that and the thoughtful piece.
"By providing unprecedented access to high-quality, large-scale single-cell data, Tahoe is promoting a more open, collaborative approach to scientific discovery. ... The release of Tahoe 100M may represent a first step towards creating the “internet of biology”, laying the foundation for the development of truly transformative AI models to integrate and understand cellular biology and drug development at high speed."
Dan Caron⚡ retweeted
Does RAG solve hallucination?
Even w/ RAG, we found that >30% of LLMs' medical statements are not fully supported by (sometimes contradict) the cited refs @NatureComms nature.com/articles/s41467-0…
We present #SourceCheckUp agent to verify faithfulness of LM to source info. Great job @kevinywu @ericwu93 w/ awesome collaborators👏
Dan Caron⚡ retweeted
" I think maybe in the next 10, 15 years we can actually have a real crack at solving all disease."
Nobel Prize Winner and DeepMind CEO Demis Hassabis on how AI can revolutionize drug discovery doing "science at digital speed."
Dan Caron⚡ retweeted
About a month after Donald Trump took office, almost all grant-review meetings remain suspended at the US National Institutes of Health, preventing the world’s largest public funder of biomedical research from spending much of its US$47 billion annual budget.
go.nature.com/3EQlDjy