>When tested in their secure lab, the researchers found that these novel phages were more effective at killing the common microbe E. coli than the natural phage ΦX174, which served as a template for Evo 2 to work on. The results were published on Thursday in the journal Science.>“Our study is a proof of concept showing for the first time that generative design can generate entire functional genomes,” said Brian Hie, the project leader. “We chose to work with phages because they are small viruses, well studied and have broad applications in molecular biology and therapeutics.”>The approach could be extended later to larger organisms such as bacteria and then to more complex forms of life, he added. https://www.ft.com/content/5ab33fb4-2636-4bb0-aa0e-3da6a8f71838?syn-25a6b1a6=1https://www.science.org/doi/10.1126/science.aec2657
>>17032677that's pretty cool
>>17032677Finally something that isn't LLM slop. >AIYeah, "AI" in the same sense that alphafold was "AI". The only worrying thing is that it was published in science. You know how high impact journals are.
>>17032705It is, in fact LLM slop (a language model trained on millions of genomes)
>>17032732Specifically>Evo 2 is trained with 40 billion parameters and 1 megabase context length on over 9 trillion nucleotides of diverse eukaryotic and prokaryotic genomes.>Evo 2 learns to accurately predict the functional impacts of genetic variation—from noncoding pathogenic mutations to clinically significant BRCA1 variants—without task-specific fine-tuning. Mechanistic interpretability analyses reveal that Evo 2 learns representations associated with biological features, including exon–intron boundaries, transcription factor binding sites, protein structural elements and prophage genomic regions. The generative abilities of Evo 2 produce mitochondrial, prokaryotic and eukaryotic sequences at genome scale with greater naturalness and coherence than previous methods. Evo 2 also generates experimentally validated chromatin accessibility patterns when guided by predictive models3,4 and inference-time search.
>>17032732what I mean specifically is that it's not a compressed reddit database but actually encoding something relevant to the application
>>17032738what if you trained it on millions of genomes AND reddit tho?
>>17032741>what if you fit a polynomial and add a ton of extra dimensions that all model canine excrement
>>17032735A statistical model would have achieved the same effect with 5 parameters.
>>17032745Obviously not, which is why they're bothering to do this.
https://www.biorxiv.org/content/10.1101/2025.09.12.675911v1.fullpreprint is actually ancient in AI termsi'm assuming they've grown a humanoid by now
>>17032735the next interesting step will be to compress that param count by a few orders of magnitude by optimizing architecture
>>17032753what if, instead, you just grew it by 1000x?
>>17032755>>17032762fuck off>>17032757Making models more dogshit? That's what private enterprise is for.
>>17032677Am I the only one seeing an ewok
Did it work? IDK, feed it to some test mice to see what happens. It kills an sentient creature off the universe completely. Life was here for a reason and you did the opposite to it.
>>17032747If they actually had a statistician, they wouldn't need to do this.
now imagine it being used to engineer unkillable viruses
This is Wuhan-dangerous but also very perspective, cool and futuristic.
>>17032753Evo 2 has a 7B variant btw
