We open-sourced an implementation of an LLM-Research manager that @OpenAI Codex built based on the Gist we published earlier this year. github.com/GPTomics/cofr It's been a lightweight useful tool for managing different research projects our team works on.
When we think of evaluating agents for research, we really need to start peeling apart:
1. do they know/can reason about science at PhD+ level
2. can they generate/retrieve evidence and operate scientific tools
3. can they execute/recover through a long workflow
4. are its conclusions scientifically valid, calibrated, traceable
5. do they actually helps a researcher do better work
A researcher does all of this and jumps between the levels constantly. We need evals that help us better understand which of these an agent is good at.
We're moving BioSkills to an archive state on github. We feel that it has served its purpose of making people think about agent skills from a task perspective. The skills are still useful, we just will no longer make updates or fix bugs.
While a few examples may be from movies, it's important to remember that most of the great discoveries arise when empirical evidence contradicts expectations.
AI4bio is accelerating quickly, so it's important to remember that the wet lab is the ultimate validator. @BMLauro19 points out that even in a fully automated DBTL, the available assays present a bias that requires experienced bench scientist oversight gptomics.com/blog/the-wet-la…
AI4bio is accelerating quickly, so it's important to remember that the wet lab is the ultimate validator. @BMLauro19 points out that even in a fully automated DBTL, the available assays present a bias that requires experienced bench scientist oversight gptomics.com/blog/the-wet-la…
It's been fun iterating on this model architecture and seeing how some of our partners have adapted it to their proprietary data along the way. It's available under an Apache 2.0 license and the training code is available in the repo. Thanks @BMLauro19 for collaborating.
We're released the v1.0 of BioJEPA-AC, a JEPA based cell state model that uses action conditioning to learn both genetic and chemical perturbation response in a unified model. This represents a core extensible framework that can easily be expanded github.com/GPTomics/biojepa/…
We're released the v1.0 of BioJEPA-AC, a JEPA based cell state model that uses action conditioning to learn both genetic and chemical perturbation response in a unified model. This represents a core extensible framework that can easily be expanded github.com/GPTomics/biojepa/…
Physics ditched the ether and unlocked a deeper understanding of space-time. Genomics still clings to its equivalent: the reference genome. What if ditching it unlocks a deeper understanding of the genome itself?
We've shown genomic position needs reform. It's based on a broken reference based system. Through special relativity, physics ditched the idea of absolute rest, so @infoecho and @domenjemec propose that we need Genomic Relativity gptomics.com/blog/relative-t…
Part 2 of the series where we look at how physics dealt with absolution position and it's parallel in genomics. On the surface the idea of genomic relativity seems like semantics but we think that it would unlock a new way of looking at the hardest genomics problems.
We've shown genomic position needs reform. It's based on a broken reference based system. Through special relativity, physics ditched the idea of absolute rest, so @infoecho and @domenjemec propose that we need Genomic Relativity gptomics.com/blog/relative-t…
We've shown genomic position needs reform. It's based on a broken reference based system. Through special relativity, physics ditched the idea of absolute rest, so @infoecho and @domenjemec propose that we need Genomic Relativity gptomics.com/blog/relative-t…
Having helped bring long-read technology from its dark age to a new dawn by tackling genome assembly, I've been reflecting on what it all means.
I'll never forget the first time I understood the beauty of Einstein's relativity. I hope we can get there with genomics soon.
@infoecho and @domenjemec continued to think about how genomics is defined by the reference. They found there was a parallel with physics and its attempts to understand what the center of the universe was gptomics.com/blog/some-guy-f…
Physics for a long time focused on swapping out what the reference for object position was, from earth, to the sun, to the ether, and eventually Einstein just removed the need for absolute position. We're seeing a lot of similarity in genomics with position/variants.
@infoecho and @domenjemec continued to think about how genomics is defined by the reference. They found there was a parallel with physics and its attempts to understand what the center of the universe was gptomics.com/blog/some-guy-f…
I remembered when I was in Applied Biosystems as a software engineer discussing about "copy number variation". I was so confused about the concept.... 1/N
The reference genome has been useful for advancing genomics, but, it's starting to become a systemic limitation. @infoecho and @domenjemec explored what a "variant" actually means and why the relying on references is holding the space back. gptomics.com/blog/your-genom…
It is somewhat like if I ask a radiologist give me the X-ray slide of my broken wrist bone, he or she would give me a reference image of other's bone and where the break is on the reference image. It could be useful, but it is not the my bone x-ray image 😆 3/N
The reference genome has been useful for advancing genomics, but, it's starting to become a systemic limitation. @infoecho and @domenjemec explored what a "variant" actually means and why the relying on references is holding the space back. gptomics.com/blog/your-genom…
If you've wondered how the layer by layer mechanics of protein folding models work, I created a series of notebooks that walks through each layer. This was way easier to do thanks to @biohub who kept the ESM models open source. github.com/djemec/descriptiv…