Jacob Trefethen

Makes you think

HT @satvik_93 Source: ir.arrowheadpharma.com/static-files/b…

I believe that 1) siRNA will change medicine, with new treatments for tens of diseases, 2) tens of millions of people will take an siRNA drug in the 2030s, with some chance of >100M people, and 3) I will have taken at least one siRNA drug within a decade. I have never worked on siRNA, and I also don't think of myself as a hypester. So, why the confidence? There's the technical answer, then the structural answer. The technical answer is that siRNA has several attractive properties as a drug modality, and enough of the chemistry problems have been solved that I'm confident more drug approvals are on the way (though there are many important unsolved problems left, e.g. in delivering siRNA to organs other than the liver). The four most attractive properties are that with siRNA you can reduce production of previously undruggable proteins by going after the RNA that codes for them; they’re “programmable” to a gene/protein you’re going after; siRNA can last long, with slow release inside cells; and siRNA drugs have the potential to be relatively cheaply made and self-administered. The structural answer, I believe, is biotechnology’s combination of long timelines with particularly un-juicy, hard-to-remember concepts. If something takes 20 years, you can often spot it 10 years in -- but when it’s an acronym and involves chemical conjugation steps, it’s hard to remember to care. Many important ideas in biology are messy and technical and, unless you've caught the bug (intellectual or literal), boring. That means there are many important secrets lying in plain sight, as per @nayafia's last book. While disagreements and unknowns remain, it is not particularly controversial among scientists that siRNA holds a lot of promise (alongside some realised promise already!) -- and, yet, most people who will benefit from siRNA drugs in the 2030s have never heard of it. (PS: given my job involves stewarding charitable dollars related to science, I don’t invest in individual pharma/biotech stocks personally; this post isn’t about the financial prospects of particular companies.)

More details on our plans in the linked interview. But, this is the gist:

nature.com/articles/d4158…

Here's more on our initial strategy behind these grants, and other grants we're investigating in Public Data for Health, that may be of interest to researchers: Our starting hypotheses for areas OpenAI Foundation support can be most useful are: connected data, scarce data, and direct data. Datasets we fund will usually have one or two of these properties, described in more detail in the next tweets, and in rare cases will share all three properties. Across all datasets, we believe making data broadly accessible, while ensuring privacy whenever that relates to patient data, will become increasingly important. This will allow researchers to build on each other’s work, and scientific progress to emerge from unexpected places.

1. Connected data: Following biology across multiple steps 🔗 Biology unfolds at different scales, but most life science datasets capture only one part of reality. A cell atlas may measure molecules across thousands of cells, while single-molecule imaging follows one cell only. Both views are valuable, but because such datasets are usually created independently, it can take years of follow-on work to determine how scientific results relate to each other. We want to support individual research teams to move deliberately between these scales and to create rich, connected datasets whose different layers reinforce one another. For example, our grant to @UCSF will allow the OpenADMET team to measure key molecular properties and interactions for tens of thousands of compounds. The team will then connect this broad profiling to two views of drug transport: experimentally determined structures of transporter proteins with selected molecules bound, and assays measuring how these proteins functionally interact with the molecules. Finally, they will test a subset of these compounds in models of the human blood-brain barrier, comparing those results with in vivo animal measurements. This approach to data collection will create an unusually complete account of what determines a drug’s movement through the body – one that enables better prediction from AI models at the level we humans care about.

Proud of owl 😊

I talked more about our strategy in Public Data for Health, and these first grants, with @GrinsteinJ: insideprecisionmedicine.com/topics/informa…

At the OpenAI Foundation, we believe breakthroughs against disease will come from pairing intelligence with more observations of the world – in other words, more data. Today, we're announcing over $125M in grants, launching our second science program: openaifoundation.org/news/public-da…

Here are some of the projects we're supporting: OpenADMET will create open datasets, benchmarks, and blinded competitions to test whether AI models can be trained to predict how small molecules are absorbed and distributed across the body, to make drug development more predictable and reduce the failure rate of new drugs. 90% of drug candidates fail in clinical trials⁠. That is often due to the difficulty of predicting how they will be absorbed and move around the body – their “ADMET” properties⁠. Unsolved prediction challenges like this can be a good fit for AI, when paired with high quality data to ground the accuracy of predictions from different models. For example, AlphaFold2 transformed protein structure prediction using data from the Protein Data Bank⁠, in a competition against other modeling approaches run by CASP⁠. Protein structure prediction is now integrated into many drug discovery programs that will likely go on to benefit patients. To better predict ADMET properties of drug candidates, the OpenADMET team will develop high quality datasets to ground new predictive models, and run open challenges that researchers around the world can enter: t.co/nRlhW9a2Hk

This $60M grant will help some of the people with the fewest backup options -- smallholder farmers -- know when to plant, fertilise, harvest. Weather forecasting is traditionally a problem for supercomputers. Now, many useful predictions are cheap and deliverable locally:

Here's a 3,000 word second post on progress across these ten technologies. It was fun to write, especially to learn more about drivers of progress that don't get much coverage. Ever heard of @NIH's RADx ITAP? 2020-22 was frustrating for at-home tests; 2023-26 better:

@NIH Hep B progress continues, and hopefully will be in the news next month with an approval. A great day for some of the 250 million people around the world with chronic infections:

I started a blog in 2024 to write about ten technologies I wished existed. Today, I posted an update on them. Nine out of ten are still unavailable. But I feel differently about that now: blog.jacobtrefethen.com/halfway/

The original:

Some more about these four roles, since I will work closely with everyone we hire into them. (Salary ranges are posted on the job descriptions for these and other roles across the Foundation: t.co/LSnbwIok6Q) First, I am excited to share that, 8 weeks ago, Vidya Vasu-Devan joined the OpenAI Foundation to lead our work in High-Burden Diseases. Before this, Vidya led the investment team for a global health biotech VC fund; before that, she was Chief Strategy Officer at a vaccine company; and before that she was Director of the $2.5 billion portfolio of strategic investments at the Gates Foundation. In an earlier draft of this tweet I wrote two more paragraphs bragging about some of the things Vidya achieved in those roles, but unfortunately for me she doesn’t like being bragged about. Which relates to why you may have a lot of fun working with her if you join our team -- Vidya is unusually humble, has helped protect an unusual number of people from infectious diseases through her work, and most importantly has complementary skills to mine in bar trivia. We are hiring a Program Officer and/or Program Director on High-Burden Diseases. These are open-ended roles with a broad mandate to help solve some of the world's biggest problems in health. They will work with Vidya to figure out: what is newly possible due to AI progress, in tuberculosis drug development? In newborn health, where roughly 2 million infants die each year? In stroke, in hepatitis B, in other diseases facing market failures where we may be able to unlock rapid progress for patients? If you have a scientific or technical background and would like to spend your time thinking about those questions, we would love to receive your application. One of the other two open roles on our team is a Chief of Staff reporting to me. If you've made it this far into a long tweet, that bodes well for your attention to detail, so perhaps you should apply (put "ATN" in the "Anything else?" field on the form). I am looking for someone with excellent judgement and organisational skills who 1) moves fast, 2) is comfortable with ambiguity, 3) takes ownership without waiting to be asked, and 4) enjoys thinking about how teams can operate at their best in the presence of growth + time constraints. I'm open to a range of backgrounds, though some science in your past helps. Vibecoding addiction is a positive too. Finally, we are hiring a Head of AI for Alzheimer's. Our goal is straightforward and extremely hard: help scientists invent new tools to finally prevent and treat Alzheimer’s. We have made some initial investments (t.co/wOkiE78KUo), but much of what we do next remains deliberately undefined. We are seeking a flexible and highly autonomous leader to help us figure out the path. This role is similar to what Nan Ransohoff describes as a General Manager: t.co/9pavjfFt1z. If the role sounds exciting but you are not sure you fit the profile, I recommend reading that essay to see how much you resonate with applying something like it to AI for Alzheimer’s.

I've made it to organoids, danger up ahead. (Great listen so far, recommend.)

corememory.com/p/ai-will-cure…

We are building a team to help define the future of the OpenAI Foundation. This is our first major hiring round, and what we are able to accomplish as an institution will be shaped by who joins now. openaifoundation.org/careers Thread on roles and areas:

In AI Resilience, we are funding efforts to reduce risks that could prevent society from realising AI’s benefits. That includes securing cyber infrastructure, making sure the AI models themselves are safe, preventing biological misuse, ensuring AI’s impact on young people is positive, and more. Solving these problems requires an ecosystem of talented people and capable organisations. We have started to fund some of those organisations over the last five months, and are hiring so that we can do more, quickly:

This first grant from our Civil Society division aims to double hepatitis C cure rates over 2 years, starting in Alabama, Illinois, Louisiana, and Massachusetts. The theory behind these grants -- which will range public health and other areas -- is that organisations with local context are best placed to reach people for whom the current system isn't working, and AI can help those organisations do their best work. In many areas/contexts, the tools are ready. So, financial and technical support is powerful in combination:

@mhdempsey great footnoting style

@TheAnnaGat 🙏

In the back half of this episode Saloni and I talk through which drugs we're most hopeful could be repurposed against new diseases, as well as how prizes, patents, grants can be used to get there. Happy anniversary, @salonium! x.com/salonium/statu…

It was fun to be one of the judges for this essay competition, mostly as an excuse to read essays about science, and secondly to reflect on how AI is affecting scientific writing. The essays that read like they were 100% LLM-generated usually didn't contain the best ideas, according to my human judgement. But, some of the best essays were LLM-assisted at the level of sentences (according to Pangram's judgement). Perhaps most surprising for me was that some *good* essays contained LLM sentences AND typos!

Congratulations to the winners:

The owl has landed 🦉 x.com/owl_posting/st…

Keep posting, folks

Big and happy news. Congratulations to @hebebbington and the rest of Frontier. Interesting to read about the progress over the last 4 years from technologists working on different carbon removal paths. Long way to go, but the fog is clearing: x.com/hebebbington/s…

We've opened applications for a new $50M commitment to: * Community Support Services (e.g. accessing legal aid & public benefits) * Community Arts & Cultural Organizations * Community Journalism & Media (e.g. local media, nonprofit newsrooms) Please share with groups who may be eligible and interested:

Here's where we are starting, on AI resilience: * Bio-resilience * Cyber-resilience * AI model safety * AI's impact on young people We are grateful for what we've learned so far from practitioners in each of these fields. Building resilience will take a broad community of people x.com/FoundationOAI/…

openaifoundation.org/news/resilienc…

Saloni and I started Hard Drugs because we like inventors and scientific discoveries. It’s a treat to record our first interview with an inventor of a difficult drug. This episode goes deep on a story that mostly has not been told before. The R21 malaria vaccine will help x.com/salonium/statu…

open.spotify.com/episode/79Foyj…

Welcome to the OpenAI Foundation, Yo! x.com/yonashav/statu…

it's great to have yo

An excellent day at the OpenAI Foundation. Welcome, Bianca! x.com/bianca__martin…