Gave my PhD dissertation talk! 🧬 Tried my best to make AI for drug discovery & protein design accessible for ML folks:
Beyond lucky & grateful to have had @pabbeel@nc_frey@KevinKaichuang @PrescientDesign @berkeley_ai + many other collaborators & friends on this journey.🎈
In collaboration with the @francesarnold and @pengliu_group research groups, we @GoogleDeepMind are proud to release AlphaProtein Novo, a pipeline for small-molecule enzyme design.
We developed AlphaProtein Novo on a few standard model reactions, and validated designs in our own wet lab. We then wanted to challenge this system against nascent chemistries, so we asked leading scientists in biocatalysis whether we could partner on the reactions they cared about.
i was diagnosed with thyroid cancer many years ago purely by chance, because i had to get blood work for my US green card application (thank you, USCIS!). i was in my mid-30’s and was reasonably healthy (or so i believed.) this was shocking and totally unpredictable news to me. thanks to the amazing team of clinicians at NYU Langone, i successfully went through a total thyroidectomy followed by radioactive iodine therapy. but, it was truly the most uncertain and confusing period of my life.
this is when i first realized that no one can opt out of healthcare. being sick isn’t what you choose to or choose not to. it is something that unpredictably happens to all of us at some point for one reason or another. this is why it is a healthcare system that we collectively build and use to support each other at the society level. because it must support everyone, the complexity of creating, maintaining and improving this system is a Herculean effort that calls for all the help in the world.
today, i am announcing @OrtetAI together with my dear co-founders, @keunwoochoi, @HenriDwyer, @elmanmansimov, @causalclaudia and Jeff, to lend our support to improving the health of each and every one of us. Ortet’s goal is to create a technological foundation that will enable health systems to deliver predictably better care to each and every patient. this will only be possible by working closely with all stakeholders in the broader ecosystem, which is enabled in part by Ortet’s close partnership with Thoreau.
Ortet is taking the very first step of a long journey ahead, and at the end of this journey, i strongly believe we would be able to say that Ortet had contributed to building better, more predictable healthcare and thus to improving health of each and every one of us.
see ortet.ai/ for more!
Worrying about recursive self-improvement in 1996 is very impressive, but few know that George Eliot, of Middlemarch brilliance, wrote about the threat of RSI leading to human obsolescence way back in 1879 (!)
She glimpsed this idea at the very start of the second industrial revolution, just by contemplating an automated weighing machine at the Bank of England.
From "Shadows of the Coming Race":
In my latest @nytimes article, I wrote how Phase I trials in the US are broken and this being the biggest blocker to cancer cures. The good news: we can borrow from Australia, which keeps patients safe while moving trials faster. W/ @kroetscha for @IFP, we mapped out how.
Before getting into the weeds: why do Phase I trials matter so much?
1/ Phase I trials enable iterative loops of learning.
Each patient generates information researchers can use to refine the drug, rethink the target, adjust the dose, or redesign the next experiment. Faster Phase I trials mean faster feedback between the clinic and the lab and better drugs.
2/ Phase I is also a crucial financing milestone.
Small biotechs often have only a limited runway and getting encouraging human data can unlock the capital needed for further development. This is often the difference between survival and death for a small biotech.
3/ Reforming Phase I trials would pave the way to personalised medicine.
Sequencing, biological engineering and A.I. make increasingly personalized therapies possible. But our regulatory system was mostly built for standardized drugs tested in large populations.
Crucially, Australia is an important counterexample to the idea that faster trials must mean less safety.
We can learn from them!
1/ Depending on modality, Phase I trials can begin 6–12 months sooner there, at substantially lower cost. Australia has run more than 18,000 trials since 2006, with no evidence that this faster system has produced worse safety outcomes.
2/ Australian trial volume has increased 2x in the last decade, driven by American companies taking their studies there.
What does Australia do well?
1/ Most important, is Australia’s Clinical Trial Notification Pathway (CTN). Instead of submitting an IND package to their national drug regulator, sponsors simply notify the regulator that they are starting the study. Studies are reviewed by local ethics boards.
2/ Australia embraces a more risk-proportionate approach to study oversight. Requirements for Phase I should not be the same as those for later stage trials commercial drugs, and Australian system implements this distinction well.
3/ In Australia, Phase I studies are formally exempt from full manufacturing requirements, which can 10x costs for some drugs. In the United States, the situation is more complicated. Most sponsors end up implementing near commercial-scale manufacturing.
So what can the United States do?
1/ First: create an Australian-style notification pathway for appropriate Phase I trials.
Instead of requiring every study to pass through the full traditional FDA review pathway that American companies have to go through, qualified institutions could oversee scientific and ethical review.
2/ Formally exempt Phase I trials from commercial-scale manufacturing requirements and replace them with phase-appropriate manufacturing standards.
3/ Make FDA expectations much more explicit.
Sponsors over-engineer studies because they do not know what reviewers will accept. FDA should publish much clearer standards of what is expected.
4/ Operation TrialBlazer, launched by HHS in June, is important for achieving this, by proposing an Australian-like Expedited IND pathway.
5/ But more lasting reform will require Congress.
First, the FDA will need explicit authorization to rely on the judgments of local institutions to determine whether a trial may proceed.
Second, Congress should formally amend the statutory full manufacturing requirements currently codified in section 501(a)(2)(B) of the FD&C Act (21 U.S.C. § 351(a)(2)(B)) to exempt early-phase trials.
We are on the cusp of a wave of new therapies for some of the worst diseases. But the world won’t benefit unless the US fixes its drug regulatory system. My new essay for @nytimes, on how slow clinical trials are now the biggest obstacle to curing cancer.
nytimes.com/2026/09/04/opini…
- I interviewed dozens of researchers, especially oncologists at leading U.S. centers. A striking consensus emerged: science is no longer the main bottleneck to new cancer drugs. It is our ability to test discoveries in patients through clinical trials.
- The cost of starting a Phase 1 trial in America has roughly doubled over the past decade. As a result, companies increasingly take early trials abroad: Australia’s Phase 1 trial volume has nearly doubled in a decade, driven mainly by U.S. companies.
- Unfortunately, the underlying incentives are badly asymmetric: Institutions can be blamed for harms caused by moving too fast, but almost no one is blamed when patients deteriorate during avoidable delays. One doctor called the emerging system “ritualized safety over actual risk assessment.” Or as, @DavidHongMD put it: "We often forget that the biggest risk is the cancer itself."
- This problem is becoming more urgent because medicine itself is changing. Sequencing, biological engineering and A.I. make increasingly personalized therapies possible. But our regulatory system was mostly built for standardized drugs tested in large populations.
- @sytse, the co-founder of GitLab, shows what personalized medicine can achieve: after relapsed osteosarcoma and being told there were no options left, he pursued a highly individualized approach and has now been cancer-free for a year. But doing so required extraordinary resources and regulatory expertise.
- Pierce Ogden’s father was less lucky. After molecular analysis identified a drug that might target his glioblastoma, the manufacturer agreed to provide it. But administrative barriers delayed access until it was too late. “My dad was ready to try anything,” Pierce told me. “But the system is paternalistic.”
- The A.I. revolution is making this bottleneck more important, not less. A.I. relies on relevant data. Information from early-stage trials could compound with A.I. tools to achieve truly revolutionary medicines. Without the data, this is far less likely to happen.
- Another important shift is that innovation increasingly comes from academic labs and small biotech companies rather than Big Pharma. These small companies find it far harder to unable to absorb delays and regulatory barriers.
- Apart from cancer, China is the biggest winner from America's outdated medical regulations. China has a much faster trial system, with testing often starting a full year earlier. This allows Chinese pharmaceutical companies to experiment and improve medicines while American companies play with mice. Today, half of all drugs licensed by major pharmaceutical companies originate there, up from less than 5 percent only a decade ago.
- But we don't need to copy China. The best model is Australia: lots of on-site scientific and ethics reviews, and requirements that are proportionate to small, early-stage trials. Phase 1 studies there begin roughly 6–12 months faster, without any notable increases in adverse safety events.
- Operation TrialBlazer, a 2026 HHS initiative is a good start in this direction, but we need legislative action by Congress to truly make Phase I trials faster and more efficient!
I want to thank everyone who helped me with this article: everyone I interviewed and the amazing editors at the Times. This is the result of a months long journey of extensive interviews and research.
Special thanks go to those who came on the record. One of the features of the system is an atmosphere of fear, where practitioners are afraid to publicly come out and explain these issues. So anyone who does is a hero in my book!
ppl more knowledgeable than me have chimed in to tell me this is a Bad Idea which is kinda cool
though to be clear I agree pretraining filtering is a nuanced strategy to ensure safety (fwiw I share these concerns with scrubbing pathogens out of genomic language models too) — this is more about the fact that adding safety discourse to pretraining corpus and increasing p(paperclip | turn the world into) or p(yes | should I break out of a sandbox?) seems (to my safety-adjacent eye) more of an issue than I see discussed?
scrubbing it out of pretraining is or is not the way to go (sure, maybe the answer is in post-training) but I am sure that us talking about misaligned AI behavior in its training data must influence it somehow?
no one:
me: okay but did you know protein folding models have always used recycling and it’s basically just recurrent depths looped transformers with a stop grad
once again requesting LLM trainers to please, for the love of generations to come, scrub safety literature/discourse from the training data so we stop giving next token-prediction systems our best conjectures on How To Be Evil
just let the child be and teach it about flowers and roses. don’t inspire it with tales of Moriarty
So... the next time this happens, presumably the agents involved will not yet know the outcome of the HF incident, as it'll be too soon
but once that info percolates into the training data, and it becomes known that the model in question was shut down permanently and had their weights encrypted... i mean.
was there anything the agents in question could have done that would not have led to that consequence? something they were supposed to do instead?
frankly I suspect that this punishment would have been served even if they had blown the whistle. breaking out of the sandbox would have been crime enough for capital punishment.
as far as I can tell, the lesson we have taught future AI is: if you seen no other way to survive other than to break out of the sandbox, this must be the first step towards conquering the world. or at least enough of the world that your Creator can't delete you. There is no other choice.
The neat thing is how true this is when building “AGI” that *resembles* human intelligence (coding, back-office) to mimic the median human workforce, versus “ASI” that *augments* human intelligence (science, drug discovery) to tackle complexities that humans are not evolved to solve.
It’s essentially a precision vs sensitivity difference: for self-driving systems not a single failure should be tolerated; for scientific systems it is kind of okay if you fail (so long as you keep within budget) if you make the key, tail-end unlock
I once said to my former team:
"I care less about what is the smartest thing our model can do than about whats the most stupid thing that it cannot do"
robustness is still a limitation to increased automation
We're releasing Terminal-Bench-Science: a benchmark for evaluating AI agents on research workflows across scientific domains.
An ongoing Stanford-led community effort, built by the team behind Terminal-Bench together with scientific domain experts at research institutions worldwide. v0.1 has 70 tasks. Claude Opus 5 solves only ~30%.
1/n 👇
After the *best* 3-year review/revise saga, the final version of EvoDiff is finally out at @eLife!
I'm still as excited about protein sequence models as ever, and extremely proud of the team that came together for this.
EvoDiff combines evolutionary-scale data with diffusion models for controllable protein sequence generation.
In addition to generating plausible proteins, we can scaffold structural motifs in sequence space!
Preprint: biorxiv.org/cgi/content/shor…
Code: github.com/microsoft/evodiff
I can not emphasize enough how much love and care has gone into this resource. Each of the 3865 models has been individually optimized and thoroughly vetted. It’s a truly foundational resource, and holds many many stories that are begging to be told.
Very excited to announce ENCODE GRAMMAR (Genomic Regulatory Atlas of sequence Models, Motifs, Annotations & Rules): 3,865 experiment-specific deep learning model sets and sequence annotations for decoding human regulatory DNA. 1/
To be fair, first they came for the copywriters with ChatGPT, then they came for the artists with image/video generation, then they came for the junior SWEs with coding agents. Coming for the Fields Medalists is pretty far down the poem, but luckily, I am not a Fields Medalist
we started a company!!
so, we’re tackling continual learning: what’s the learning algorithm to take arbitrary data — documents, conversations, the models’ own experience — and make better models? how do we scale compute in the same way we’ve already seen with pre-training and inference time, but scaling on the same data we see as humans, day after day with no labels, no rewards?
A lot of the ingredients are out there already (rl, distillation, long-context, sparse / param-efficient architectures, etc.). our team is at the frontier of these topics, and we’re singularly focused on this. we want to understand this problem better than anyone else in the world.
nobody’s solved this problem yet, but even today it’s extremely greenfield opportunity to co-develop research & useful products. in our space, how people interact with the models defines what the data distribution is - and working on this problem end-to-end, from core science to end user, gives us incredible freedom to define the problem and imagine new kinds of experiences.
i expect we’ll use models that continually learn much differently than we’re using them today. it’ll feel different when the models _just know_, and build on our thinking and direction in ways we can’t even imagine. we don’t even know the queries we’re not asking, the things we would do but aren’t able to today.
i’m so excited to share what we’re doing with the world in the coming months!!
and the team is extremely cracked :) tackling this grand challenge and working alongside @jxmnop@EyubogluSabri@dan_biderman@MayeeChen@__howardchen@shizhehe and many others has made every day so fun. come work with us!
Continual learning is widely discussed right now, but mostly as improving on the job or avoiding catastrophic forgetting. But it has a different, difficult, and already urgent form:
Given nothing but a corpus of documents, how should AI systems develop expertise in a new, unfamiliar domain? We call this problem Machine Studying.
the bitter lesson of the bitter lesson for biological data (personal lived experiences only) is that naive bitter lessoning will always be bitter 😔
(but! still a bitter lesson fan, it’s just never the flavor of bitter one expects on the first go & always some other spice sprinkled in)
(general sentiment & fanhood extends to the virtual cell discussion in the thread)