Astute
Largely agree with these thoughts, and have some perhaps discordant points to add.
- Solving biological diseases in the next few years will benefit mostly from the usefulness of the language models, substantially more than molecular ML.
- Proponents of the bitter lesson for biological models (not LLM) aren't wrong, just (possibly pathologically) too early. We don't have enough data, and the data we can collect quickly is probably not the right kind (we may get lucky with biology being easy, but it's usually not).
- I think just scaling models and public (passively accumulating) data is a cost ineffective direction (it's like scaling dense NNs on the internet). Today's model outputs are as good as GPT2. A single bioinformatician with linear regression, sometimes even without any additional data, can beat these models in many applications (people rarely benchmark AI advances against a simple Potts model). The most promising approaches for molecular models are innovating heavily on architecture (e.g. Baker Lab, Alphafold, etc). This may automate/speed up with LLMs.
- Few companies (hopefully Dyno included) are betting that we've triangulated enough pieces from multiple experimental channels to predict (with scale) how things translate in humans. Having worked on translating in vitro things to primates, I'm very cautious with the idea that current models, trained on current public data, can do anything more than manipulating already working solutions, either explicitly or through implicit remixing of what has worked. It may still work, but the underlying bets offer no specific reason why there is any epistemic risk advantage other than being "made with AI".
- Short term, there is more promise in combining molecular models with LLMs rather than scaling them. The two sources of lift in AI discovery elsewhere has been (1) Breaking Silos: AI synthesizes a bunch of knowledge/expertise that is rarely present in the same room (2) Data mining : AI pushes through volumes of data, with an intelligent eye, that humans or simple algorithms simply can't scale or adapt to. I'm curious what will come out of multimodal training of LLMs and biological data, which seems to be something the big players are pursuing. I think latent reasoning in the language (or better yet, + code) domain will reveal a lot more biology than scaling on nucleotides or proteins alone.
- My view is that the next opportunity of combining lab and AI will be massive-scale hypothesis testing with LLMs. Currently a multiplexed assay generally tries to test a handful of ideas because the humans are bandwidth limited in designing properly powered experiments. LLM's can scale this very well.
I have no doubt lots of good news will come on the impact of AI in biology in the next five years. I can't tell what paradigm will win, but my prediction is that it will not be the same paradigm imported from the LLM land (which will get a ton of funding anyways, because it's easy to pattern match for tech VCs).
1
0
6
3,238
Arjun Raj retweeted
It is hilarious that super intelligence costs the same as Netflix
139
415
7,558
156,239
Arjun Raj retweeted
What fraction of causal human biology do we actually know?
My answer: there isn't one answer. It depends on the biological question we're asking, and that distinction matters for the role we expect AI to play in drug discovery.
Explore Part 1 via:
plengegen.com/blog/chb-all-y…
My answer: there isn't one answer. It depends on the biological question we're asking, and that distinction matters for the role we expect AI to play in drug discovery.
Explore Part 1 via:
plengegen.com/blog/chb-all-y…
9
15
79
10,238
Executing with high levels uncertainty and pure hype can look very similar on the surface. Discriminating between them is the real superpower.
0
0
17
2,301
I am proud of the clear, logically consistent writing we pore over by hand in the lab. AI still can't quite automate that, at least not yet. The problem is… I don't think logical consistency is valued by humans as much anymore anyway.
The problem of AI for academic writing, and why you don't notice it in your own AI writing:
AI loves to use vague language. It will produce something that’s consistent with what you want it to say. The problem is that it will also be consistent with other interpretations that you didn’t intend. This confuses readers: which interpretation was meant?
But you won’t notice the problem when you proofread your AI’s output. That's because you’re checking whether the output is consistent with what you meant (which it will be), without thinking about how it could be misunderstood (which we're usually unskilled at, due to the curse of knowledge)
1
5
44
6,617
Arjun Raj retweeted
Happy to share our preprint with an exciting and kind of wild result about how much lineage can be extracted from scRNA-seq alone. biorxiv.org/content/10.648…. That is, could a lineage tracing experiment be done virtually to enhance cell fate inference, at least at specific timescales? (1/14)
1
7
51
6,192
Arjun Raj retweeted
New nature paper on the anatomy/architecture of personal health agents drops. The gold standard benchmarks is how good experts (MDs, coaches etc) think the results are. When expert human judgement is the gold standard agents are trying to approach... replacement feels pretty far away indeed.
1
2
6
954
Build bricks, not dashboards
TL;DR:
AI has given us the ability to realize an enormous amount of implicit knowledge—ideas, analyses, and applications that were always within reach of what we already knew, but previously took considerable effort to produce. This is tremendously useful, but it doesn't necessarily give us the foundations needed to push further. For that, we need to focus on building new, reusable pieces of knowledge that make the next round of progress possible. Let's spend our tokens building bricks, not dashboards. 😀
AI has given us a crazy new world of apparently analytic and information abundance. Now, with a few vague lines of instruction, I can create not just a dashboard, but whole worlds, on demand, in hours or even minutes. In principle, this development represents an utter and complete transformation of knowledge work. And yet… I can’t help but ask myself sometimes: where’s the meat? Do I really need another dashboard? I mean, yes, there are seemingly huge gains in so many things, but it also feels like there is this invisible wall at the boundary (many other AI-pilled people I know also feel this boundary).
One way I’ve been thinking about it is that we currently have a lot of implicit, unrealized knowledge. Imagine the world of knowledge as a sphere. Normally, people work really hard to extend human knowledge by pushing past the boundary of this sphere in a particular direction (perhaps that’s your PhD, ala Matt Might). It takes a lot of work even to make a small dent in this universe. Now, it seems like we can push the entire sphere out all at once. AI lets us bring ideas to life faster than could have ever been imagined, realizing all that implicit knowledge. But I think it will hit a wall sooner than we think, because what we build with it is often more of a cloud than a solid layer.
It’s worth asking HOW AI is able to realize implicit knowledge. Take coding. The reason that it can whip up a dashboard in minutes is because it is able to build on foundational, battle-tested frameworks. Sure, you can now build crazy things completely from scratch, but again, that’s generally following a playbook (and spending a ton of tokens). But it seems that many of these gains are relatively shallow, and consequently are often very hard to build upon. I feel like the same is true in my neck of the woods in science. We are now extremely capable with AI of filling in a lot of blanks using all the various scientific packages and knowledge out there. This capability gives you “AI euphoria”. But is that what we should be doing with our time?
I think that building the next layer of knowledge requires building bricks, not dashboards. What makes something a brick? To me, a brick is something that changes not just what is possible now, but what is possible next. A new conceptual framework, or a new, well-tested package that solves an important problem robustly. Making a brick requires synthesis, focus, and careful, deep thought. A bunch of half-baked dashboards, semi-playable video games, and minor scientific connections are not going to get us to the next step. I suggest we spend our efforts and tokens towards building these foundations instead of fluff.
Note: I am very much NOT saying that building foundations requires humans per se. I do think that as of right now, building these bricks does, to some extent, still seem to require humans. But that might all change in a month, and I would be a fool to bet against that. My point is more that as we guide our own intelligence or our purchased intelligence (or purchased intelligence guides itself, or whatever), we should be pushing it towards a common goal of building foundational pieces of knowledge that we can build upon. I worry a little that we are optimizing AI towards making flashy demos that give us a bit of dopamine and a nice Tweet, but don’t actually move the needle for knowledge. Regardless of who makes it, maybe the measure of a new piece of knowledge might not be how much it tells us here and now, but how much additional knowledge it makes accessible.
TL;DR:
AI has given us the ability to realize an enormous amount of implicit knowledge—ideas, analyses, and applications that were always within reach of what we already knew, but previously took considerable effort to produce. This is tremendously useful, but it doesn't necessarily give us the foundations needed to push further. For that, we need to focus on building new, reusable pieces of knowledge that make the next round of progress possible. Let's spend our tokens building bricks, not dashboards. 😀
AI has given us a crazy new world of apparently analytic and information abundance. Now, with a few vague lines of instruction, I can create not just a dashboard, but whole worlds, on demand, in hours or even minutes. In principle, this development represents an utter and complete transformation of knowledge work. And yet… I can’t help but ask myself sometimes: where’s the meat? Do I really need another dashboard? I mean, yes, there are seemingly huge gains in so many things, but it also feels like there is this invisible wall at the boundary (many other AI-pilled people I know also feel this boundary).
One way I’ve been thinking about it is that we currently have a lot of implicit, unrealized knowledge. Imagine the world of knowledge as a sphere. Normally, people work really hard to extend human knowledge by pushing past the boundary of this sphere in a particular direction (perhaps that’s your PhD, ala Matt Might). It takes a lot of work even to make a small dent in this universe. Now, it seems like we can push the entire sphere out all at once. AI lets us bring ideas to life faster than could have ever been imagined, realizing all that implicit knowledge. But I think it will hit a wall sooner than we think, because what we build with it is often more of a cloud than a solid layer.
It’s worth asking HOW AI is able to realize implicit knowledge. Take coding. The reason that it can whip up a dashboard in minutes is because it is able to build on foundational, battle-tested frameworks. Sure, you can now build crazy things completely from scratch, but again, that’s generally following a playbook (and spending a ton of tokens). But it seems that many of these gains are relatively shallow, and consequently are often very hard to build upon. I feel like the same is true in my neck of the woods in science. We are now extremely capable with AI of filling in a lot of blanks using all the various scientific packages and knowledge out there. This capability gives you “AI euphoria”. But is that what we should be doing with our time?
I think that building the next layer of knowledge requires building bricks, not dashboards. What makes something a brick? To me, a brick is something that changes not just what is possible now, but what is possible next. A new conceptual framework, or a new, well-tested package that solves an important problem robustly. Making a brick requires synthesis, focus, and careful, deep thought. A bunch of half-baked dashboards, semi-playable video games, and minor scientific connections are not going to get us to the next step. I suggest we spend our efforts and tokens towards building these foundations instead of fluff.
Note: I am very much NOT saying that building foundations requires humans per se. I do think that as of right now, building these bricks does, to some extent, still seem to require humans. But that might all change in a month, and I would be a fool to bet against that. My point is more that as we guide our own intelligence or our purchased intelligence (or purchased intelligence guides itself, or whatever), we should be pushing it towards a common goal of building foundational pieces of knowledge that we can build upon. I worry a little that we are optimizing AI towards making flashy demos that give us a bit of dopamine and a nice Tweet, but don’t actually move the needle for knowledge. Regardless of who makes it, maybe the measure of a new piece of knowledge might not be how much it tells us here and now, but how much additional knowledge it makes accessible.
0
3
19
3,212
Arjun Raj retweeted
manuscript central be like do you want to reset your password right away or shall we go through the motions of you trying to log in
2
8
110
5,607
Arjun Raj retweeted
It’s indeed time to transition from mRNA and protein levels do not correlate.
Don't correlate is a weak statement, especially in the absence of independent measurements & reliability accounting.
What's compelling for me are direct measurements of what regulates protein abundance.
doi.org/10.1101/2025.0…
1/
0
7
40
7,499
There is a certain untethering from reality and logic in science these days that I find troubling.
This is a terrible trend in AI & applied AI. Scientists in industry & academia shud seriously just cut this bullshit. U will be called out every time u do this. U might get away with it wrt more naive audience but u increasingly lose credibility. 3/
1
4
27
8,314
Arjun Raj retweeted
We’ve just made public 125,091 DICOM radiology images from across my cancer journey (link below). The data is both longitudinal and multi-modal, dated and linked to my treatment timeline, so you can compare imaging across initial diagnosis, standard of care, first remission, re”, and second remission. It’s viewable in OHIF, the standard open-source viewer used in radiology research.
Given the long history of applying deep learning to radiology, we’re hoping to see a variety of methods and models applied to my data. If you find something interesting, please reach out to cancer@sytse.com, or if you know a researcher who should see this, tag them below.
Thanks to Terra LaFerriere for organizing these images, and Kamil Slowikowski for building osteosarc.com into the resource it is.
Given the long history of applying deep learning to radiology, we’re hoping to see a variety of methods and models applied to my data. If you find something interesting, please reach out to cancer@sytse.com, or if you know a researcher who should see this, tag them below.
Thanks to Terra LaFerriere for organizing these images, and Kamil Slowikowski for building osteosarc.com into the resource it is.
19
74
504
28,511
Arjun Raj retweeted
1/ How do cells divide when they're stuck in something squishy? Most of what we know about the cell cycle comes from yeast in liquid or on flat plates, but in nature they live in dough, mucus, soil and tissue. This new paper asks what that does to division. Love this one!
4
20
91
7,694
Arjun Raj retweeted
Our new work using virus-specific T cells as biosensors in Long COVID. These data point to viral persistence detected by T cells and some level of T cell exhaustion. Excellent work by @markmpainter biorxiv.org/content/10.648…
21
124
456
38,476
At some point, will it be hard for us to even judge whether a new model is better than the one before due to our puny meatbag brains?
6
2
19
3,016
Arjun Raj retweeted
In fact, most of these comrades who I gauged to be more brilliant than I have gone on to become distinguished mathematicians.
They’ve done all things, often beautiful things in a context that was already set out before them, which they had no inclination to disturb.
Without being aware of it, they’ve remained prisoners of those invisible and despotic circles which delimit the universe of a certain milieu in a given era.
To have broken these bounds they would have to rediscover in themselves that capability which was their birthright, as it was mine: the capacity to be alone.
- Alexander Grothendieck, Récoltes et Semailles
They’ve done all things, often beautiful things in a context that was already set out before them, which they had no inclination to disturb.
Without being aware of it, they’ve remained prisoners of those invisible and despotic circles which delimit the universe of a certain milieu in a given era.
To have broken these bounds they would have to rediscover in themselves that capability which was their birthright, as it was mine: the capacity to be alone.
- Alexander Grothendieck, Récoltes et Semailles
3
11
111
7,318