Jason S.’s Post

It's pretty interesting to see the progression of AI models. It was once thought that AI would first be utilized in basic tasks and not in creative or advanced logic tasks. But in practice AI has excelled where 1) there is discrete and large quantities of data for training and 2) where formal systems exist that provide structure to a learning process. The former is important for generation, and the latter for search and reinforcement. Formal mathematics falls squarely in the latter category and combined with generative content provides the structure in which to find and reinforce theorem proving policies. I think we should expect this type of approach to continue to advance rapidly Some additional thoughts on this topic are here: https://lnkd.in/gRyEgxHc #ai https://lnkd.in/gei3Z394

The Llama3 technical report is chalk full of useful technical content. Doesn't have training data, but has a lot of training and infrastructure details. Interesting read. #ai

There is a tremendous amount of froth in the AI space that is also bleeding into the techbio space. Much of it is about trying to predict the pace of technology development and long-term value capture potential of products and platforms. It's easy to get caught up in narrative threads and nuances of particular markets and technologies, but it's often useful to step back at look at the bigger picture. This article does that and makes the case that understanding market drivers is as critical as building cutting-edge tech. #ai #techbio

Does it amaze anyone else that BI is still a private, family owned company? Financial discipline in chemical and drug development at its best. 

The topic of developing AI models for virtual cells has continued to attract attention.  Indeed, the idea of being able to effectively simulate the comprehensive behavior and trajectories of cells promises to accelerate many industries including drug discovery. The magnitude of the challenge is significant, and there are key difference between the biological setting and the techniques that have been successful in the digital world. This article is a follow up to earlier article on Virtual Cells that has more depth on the technical intuition for building complex simulators of biology. As with other articles in the series, it is designed to a be a semi-technical introduction accessible to a broader interested audience. For more on these topics see here: https://lnkd.in/g9g4AD7Q #ai #virtualcells #drugdiscovery

I read the entire class action lawsuit against various parties related to Illumina and Grail that was assembled and filed on Friday. I am very cognizant that these documents are highly one sided, however, we watched much of this story play out in the public media and the filing was an extensive dive into the same narrative. It would be a disservice to the many people at Grail who worked genuinely hard to develop a transformative product to discount the whole effort, but from the executive and governance perspective, this appears like a Theranos level event in magnitude. However it plays out it is likely (or should) have significant reverberations for the field.

AI is doing for knowledge what cloud computing has done for hardware The combination of the two is yielding a productivity explosion This article details the experience of building AI powered tools that is being unlocked Just a few years ago, approaching this would have been intractable, now it's at your fingertips. It's fascinating to imagine what will be built and by who. #ai #llms #research

If anyone is down in SD at BIO this week and wants to meet up send me a note.