A-TIER news
Evo-designed bacteriophages — AI genomes that beat human gene swaps
100k ft — What is this?
Evo-designed bacteriophages — AI genomes that beat human gene swaps
50k ft — Why care?
Science paper: researchers fine-tuned Evo 1 and Evo 2 on 14,466 phage sequences to generate PhiX174-like genomes, and 16 were viable. One (Evo 536) used a G4 packaging protein swap humans had failed to engineer, and recombined AI phages overcame E. coli resistance within 1-5 rounds. Visuals are illustrative, made with Seedance 2.5.
10k ft — Key takeaways
• Genome language models can satisfy overlapping-gene constraints humans struggle with
• Recombination between AI designs produced the resistance breakers
5k ft — Tools & techniques
Claude, Higgsfield
fine-tuning, video generation, whole-genome generation
1k ft — Deep-dive / try it
AI-designed viruses sound terrifying, but these were bacteriophages, meaning viruses that infect bacteria.
A phage is just short for bacteriophage. A genome is the complete set of genetic instructions inside an organism or virus. A mosaic descendant is a later virus whose genome contains pieces inherited from multiple parent phages after recombination.
Researchers used AI to design new phage genomes, tested them in the lab, and found viable designs. Some later recombined into mosaic descendants during experiments. We are now seeing specialized AI models being used to build complex, interrelated systems!
The cinematic visuals in this video were made with SeeDance 2.5 to explore a more unusual use case for AI video generation. They are illustrative entertainment, not scientifically exact …
Transcript excerpt
One of the AI design viruses everyone's talking about today did something much stranger than simply work. It made a biological component swap that previous attempts by researchers had failed to make work. And after this video, you're going to understand what that means. I'm Giga Chen, and I want this to be the most informed community on the whole platform. But I also want you all to be entertained. So I used Higgs Field to generate some visuals using Cdance 2.5, which you all know is one of my favorite video models. And I really want to hear what you think. First, one important correction to the news cycle. The big experiment we're all talking about was actually released as a preprint last September. What's new is that the peer reviewed paper was published in Science yesterday. The researchers used a tiny bacteriophage called Phi X 174. A bacteriophage is just a virus that infects bacteria. Its genome has 5386 DNA letters and 11 genes. But here's the problem. Some of those genes overlap. The same stretch of DNA can help encode multiple proteins. So changing one letter to improve one thing can accidentally break something else. And it's not like they asked Claude or GPT to make them a virus. Rather, they fine-tuned EVO 1 and EVO 2 on 14,466 related phage sequences, guided them towards Phi X 174-like genomes, filtered the results for things like genome structure and which kind of bacteria they should infect, and then physically tested a bunch of designs. 16 of them produced…
More from @gigaqian
- B-TIER Xiaomi AI Cube — 80GB unified-memory box running 120B models locally
- B-TIER Moderna–Merck personalized mRNA cancer therapy — first positive Phase 3
- A-TIER Qwen 3.8 27B — local agentic model that fits on one 24GB GPU
- B-TIER Grok Bot — persistent AI workers sharing one cloud computer
- A-TIER DeepSeek V4 Flash re-post-training — big agentic gains, same model size