ceo @ y-trap, advisor @ centivax | previously: herophilus, distributed bio, stanford, gilman
San Francisco, CA
Joined March 2009
- Tweets141
- Following416
- Followers347
- Likes348
Pinned Tweet
This is what makes coding agents powerful: tight loops between proposal, execution, error signal, and revision.
Drug discovery needs the same loop:
perturb disease-relevant systems → measure state changes → train models → choose the next perturbation.
This is what we need to make better biological bets, not just faster molecules. (5/5)
rishibedi.com/posts/ai-drug-…
Rishi Bedi retweeted
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.
But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.
So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.
TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.
tao wants human understanding of a treatment to guard against reward hacking. bio-twitter (correctly) laughs: many approved drugs have unclear MoAs.
but there is an important (different) point here: clinical trial design can be susceptible to reward-hacking (including by humans), especially in non-randomized Ph I/II studies which bear the weight of enormous investment decisions. you can optimize who gets enrolled and what gets measured to get a positive readout without actually establishing the top-line claimed clinical benefit.
trial design / patient selection getting black-box-AI’fied carries real risks if the system is optimizing for a positive readout. humans already know how to game that objective; AI will do it better.
so we don't need to understand MoA perfectly, but we do need to be hyper-vigilant about hill-climbing behavior of AI-driven trial design and patient selection
concrete work that moves the needle for lean teams moving drugs into the clinic, great work from the edison team accelerating a critical and often thankless part of the process
i've tried a bunch of ai assistants in the last ~year. dots blows them all out of the water by being connected to codex and being able to access context about all my work. this has been a very productive evening.
"proving out unique biological hypotheses" is the interesting part of Max's thread, but the rest of the narrative is focused on claiming "exquisite engineering" of a more developable candidate for a known target.
demonstrating that this engineering is actually enabling unique biological hypotheses will determine Iso's impact
Rishi Bedi retweeted
As someone who did this kind of genome mining work during my PhD, some thoughts on this Anthropic announcement:
First, the very simplified version of what they did is that they noticed two genes (one known, one new) sitting next to a weird repeating piece of DNA. More specifically, they described an unusual reverse transcriptase (RT) associated with a repetitive DNA array and an unknown accessory protein. This kind of process was used to understand CRISPR back in 2002 and was key to the gene editing tools we use today.
To put this into context, though, people have been finding RTs associated with CRISPR arrays since 2008, and this general kind of genome-neighborhood mining has been used to discover new biological systems for decades. The basic genome-mining strategy is well established, and there are now mature tools and published pipelines for doing much of this. There are papers that discover and experimentally validate dozens of new systems using this approach in a single study. Doing it in bacteriophage genomes is also nothing new (eg CasPhi).
Finding a weird cluster of genes and repeats is often the easy part. The hard part, and where the real discoveries come from, is figuring out what the system actually does. Eg for the bridge-RNA discovery in 2024 from @arcinstitute or the discovery of CasPhi in 2020 from @DoudnaJennifer they figured out the pieces of the system and the rules for what makes it work so it can be used.
Anthropic does not yet know what this does. They’ve shown that the repeat array produces RNAs, but not what those RNAs do, what the RT does with them, or whether the system has any of the programmable properties that make the CRISPR comparison justified.
I’m genuinely rooting for all of the frontier labs to seriously get into biological discovery, and I’m excited about what comes out of it. But announcing these very early, incremental findings with the framing of a major discovery doesn’t help. I’d much rather they set the bar high now, so that when an AI actually discovers a fundamentally new biological mechanism, everyone appreciates how big a deal it is.
Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to CRISPR.
We don’t yet understand what this system does, but only a handful of known systems share its features, and all of them are able to cut, copy, and paste DNA. Historically, the discovery of such programmable systems has helped revolutionize medicine. CRISPR, for instance, is now the foundation of genetic medicines. But it will take much more work to learn what this system does, and whether it can be put to similar use.
Read more: anthropic.com/news/claude-di…
"violated its own AI policy" misses the point. "additional training for all staff" won't fix the insidious yearning for performative diversity that thrives in the abyss of make-work bureaucracy.
the real story is: "administrator whose job shouldn't exist at all does dumb thing"
Stanford University confirmed it violated its owns AI policy by turning a Hispanic student into a fake Black woman in a dining hall ad. It also modified the faces of two other students sfchronicle.com/bayarea/arti…
Rishi Bedi retweeted
very cool result showing how wet lab data enables a specialized model to beat gpt-6 astra at a task at the frontier of science!
in general scaling is great and obviously i am a believer in it, but probably the more we approach the frontier of science, the more specialized data matters and that gives task-specific models a chance. this specialized data is usually private and is probably a real moat
it should be in principle true that a task specific model will probably do better at scientific discovery just because it can use more of its parameters for the task you care about
We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next.
Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon.
This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials.
Read our blog posts below.
open-sourcing CTDs is extremely exciting. we have found clinical protocols available on clinicaltrials.gov to be extremely helpful at y-trap. this will be an OOM jump beyond that in utility, especially for lean teams without decades of regulatory experience.
CTD Commons will preserve and publish regulatory knowledge from failed drug development programs so future teams can learn from precedent that is usually locked away, to help uncover effective treatments for patients faster.
A CTD, or Common Technical Document, is a document that compiles the full journey of an investigational drug – everything from animal toxicology, manufacturing details, and, perhaps most usefully, correspondence with the FDA. Yet only a small fraction of that work appears in published papers.
CTD Commons will test the possibility of acquiring CTDs from failed or shelved drug programs, and make them openly available for research and analysis, for everyone to use.
@antonioregalado wrote about the project, and @RuxandraTeslo's essay on the topic, here: xitter.cf/antonioregalado/status…
Rishi Bedi retweeted
I must be among an extremely small group of people (n=1?) that have both 1) trained a frontier LLM and 2) designed and synthesized custom viruses in a lab with my own two hands.
And I think that the takes on AI killing us all by creating dangerous viruses is total bogus.
Rishi Bedi retweeted
Mathematics has now fully morphed into biology:
Project costs millions $ ✅
Put out paper that nobody has read ✅
Massive paper supplement nobody will read ✅
Advertise with pretty art ✅
Vicious authorship fight ✅
Say you're curing cancer ✅
Rishi Bedi retweeted
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.
The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.
The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
From a YC-backed healthtech launch this week: "ILLUSTRATIVE RECORD · NO PHI" beneath a fictional marketing graphic. Obvious AI slop.
It’s like a bank bragging that the fake account statement in its ad has "NO REAL CUSTOMER ACCOUNT NUMBERS": meaningless.
I have nothing against AI-generated website or marketing copy in general. But used indiscriminately, it gives off strong "safety theater" vibes which is a bad look in healthcare and biotech.
The problem isn't just bad copy - it suggests that the people running the business don't understand the actual risk profile for their patients... and are instead using nonsense slogans to put on an (amateurish) performance of safety.
it is very exciting for codex et al. to obviate the need for scientists to look at every line of code analyzing their data
do not let this lull you into not looking at the _data_
noticed for the first time today agents in codex communicating with each other; happened automatically when i tagged one active task in another
"Sent by ChatGPT from another task"
Rishi Bedi retweeted
> "what's the easiest way to fix our comms issues?"
> "maybe we should just cure cancer"
Elegant demonstration that immune cells from the periphery infiltrate the aging brain. The possibility that modulating peripheral immunity could affect neurodegenerative disease opens up a whole new control surface for therapeutic intervention.
Gaining therapeutic access to the human brain is one of the biggest unsolved problems in biomedical science.
Today @nature, we uncover a massive influx of immune cells into the human brain during aging, revealing that the brain is more accessible than previously thought. 1/
nature.com/articles/s41586-0…
Rishi Bedi retweeted
Catching skin cancer early is a home robotics problem.
Melanoma is highly treatable when detected early, yet today’s screening process depends heavily on patients noticing tiny changes across their entire skin surface. This requires patients to solve a near-impossible visual-memory and registration problem.
I built OpenDerm, an open-source 4-DOF robot that captures high-resolution images of the skin and uses them to reconstruct and track the skin surface in 3D over time.
The best way to make skin screening truly routine is to bring it into the home. OpenDerm shows that inexpensive robotic skin imaging is possible, but the path to scale is not a dedicated screening robot in every household—it is to make skin screening one of the many useful things a general-purpose home robot can do.
Read more about why I built OpenDerm and how it works here:
Blog: marionlepert.github.io/blog/…
Project: openderm.github.io/
besides how awesome it is that a century-old conjecture is disproven by LLM, it is also very cool to see mathematics-in-public on X: open discourse from experts on how to interpret/contextualize/generalize the result
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)