1) a Prime Intellect like RL environment for biology can unlock a lot of value for DNA models, but first needs to be built 2) like most deep learning models, data is king. Curating high quality low sample size task specific data to fine tune a general dna model is the clearest path to realizing value here 3) Specific applications I find interesting are CAR T promoter design, AAV gene therapy regulatory element design, reassigning variants of uncertain significance, bridging the variant to function gap, and biosecurity (details for each one in the post) 4) This is an early technology and the potential exists for two main reasons: a) we have a lot more DNA sequencing data than protein data, b) DNA contains more information than proteins. So if protein models like AlphaFold and ESM can unlock new medicines and discoveries in biology, the possibility for DNA models are orders of magnitude higher