The future of biotech looks kinda grim if big labs with unlimited resources keep closing off their model weights like AlphaFold 3, and if LLMs keep cutting you off the moment you ask anything related to biology. AlphaFold 2 was open, and it changed everything. Everyone could build on it. ESM and OF3 still open source, thankfully, and really close to AF3 after years of work. And now general LLMS are getting so jumpy, they won't even answer basic biochem questions, just look at Fabel 5. Then there's GPT-Rosalind, OpenAI's model for biology and drug discovery, looks powerful but again locked behind trusted access and enterprise customers. So, unless you're a big pharma or have the right contacts, you are in the permanent underclass. We need much better bio safety guardrails, refusing to answer doesn't make it automatically safer.

i'm definitely still in a writer's block after that thesis draft

wild to think whales invented the concept of a graveyard millions of years before humans even existed

a lot of people i've talked to are in comp bio because they were really inspired by alphafold this is what a good scientific breakthrough does, it makes an entire field feel like the place to be, and pulls in people from physics, math, cs who never thought about biology before

Stuck with the giants

Meet the members of Proxima: Anindyadeep (@anindyadeeps) is building Litefold. LiteFold combines physics-based simulation and AI to accelerate the design and validation of drug candidates in silico. Their leading product is Rosalind: An AI Co-Scientist for Life Sciences.

To train a GPT class 1T model from scratch - including failed runs, data acq clean rlhf, post-training, team/people will likely req $250M of compute on an aggressive 3-4mo schedule (i.e. more reserved GPUs), $500-600M all-in IF you do a dense one. MoE fp8 will cut costs by 1/10th depending on how many active params you have. If you want SOTA however, the budgets go significantly higher on test-time compute, post-training RL, and data/synthetic generations..and v. high on talent. Maybe $2-4B all-in. After that comes serving the model. The talent is key to get to SOTA/beat it - and then you have to ensure this is useful enough to have inference vol over time - for which the capital will come if there is usage / TAM. So this is not as much about raising $50-60B, or raising it all at once as the OP says - we are investors in mistral, sarvam, reflection and anthropic - and they all scaled capital over time as models got adoption, but the early bottleneck is more on talent GPUs at that scale where you can do interesting things.

yeah bro release opus 4.9 wink wink

It was Amazon all along lol

when the US finds out i'm using fabel 5

been following spacex even before i knew who elon was, it's just a beautiful feeling how far they have come

my mind is absolutely blown away, crazy good sensors from precigenetics

Do single-cell foundation models obey scaling laws? A somewhat thought-provoking new Nature Methods study by the Crawford lab suggests that, for current single-cell foundation models, the answer may be “not really.” Across a broad range of architectures and downstream tasks, increasing pretraining data from hundreds of thousands to tens of millions of cells yielded surprisingly limited gains, with performance often saturating much earlier than expected. This is interesting and provides exactly the kind of rigorous benchmarking our field needs. As Felix Fischer and I commented in the accompanying Research Briefing, such studies help move the discussion beyond model size and computational budgets toward actual scientific utility. At the same time, I am not convinced the key conclusion is that scaling does not work in biology. Rather, it may be that current objectives are not extracting enough information from additional data. Interestingly, in our recent scConcept work, we observe a markedly different scaling behavior, with continued gains as training data grows toward hundreds of millions of cells. The key difference may be the training objective itself: instead of reconstruction-based masked modeling, scConcept uses a contrastive objective that directly optimizes biologically meaningful cell representations. biorxiv.org/content/10.1101/… This raises an interesting question for the field: Have we reached the limits of data scaling, or only the limits of current objectives? -> My guess is that the next generation of biological foundation models will depend less on simply collecting more cells and more on finding the right representation learning principles for biology. Nature Methods paper: nature.com/articles/s41592-0… Research Briefing: nature.com/articles/s41592-0… #SingleCell #FoundationModels #AIforBiology

Together with UC Berkeley we are announcing the laser phase plate - a breakthrough in atomic resolution imaging. This is the brightest continuous wave laser in the world, 100 million times the intensity of the surface of the sun. Phase contrast plays an important role in microscopy, but it was thought close to impossible for electron microscopy, where it would require interfering with an electron beam. Holger Mueller and Robert Glaeser proposed exactly this using a standing wave laser. It has taken over 15 years to make this a reality. Biohub partnered with UC Berkeley and Mueller to support this work and to engineer and build the technology. Contrast has been the critical barrier to achieving atomic resolution imaging of the cell. In cryo-electron tomography, a cellular imaging technology that uses electron microscopy, the low contrast makes it impossible to resolve anything but the largest proteins within their cellular context. The laser phase plate removes that barrier. With advances in AI this breakthrough in contrast will start to open up a new frontier in structural biology, that will allow us to see the molecular machines of the cell, and how they assemble into far more complex and dynamic systems, and understand how they work.

@mihirbafna14 and I are excited to introduce Promera, a co-folding and design model with • best-in-class binder filtering • nanobody design with in-silico success rates matching hallucination • case studies on hantavirus epitope targeting and GPCR agonism (1/8)

still a no for biology if you are wondering, it's a shame