Anthropic AI Finds New Enzyme System in 21 Hours
Anthropic announced its AI model discovered a new DNA-cutting system in 21 hours, though humans performed all physical lab work.
Key points
- Anthropic's AI model Claude identified a new DNA-editing enzyme system in 21 hours using 210 million tokens.
- Humans performed all physical experiments in the Bay Area lab, limiting the AI to computational analysis.
- The discovery requires independent validation, as similar systems were previously found by Stanford researchers.
Anthropic announced this week that its internal biology lab made a significant discovery. The company stated that its AI model, Claude, identified a new enzyme system with properties similar to CRISPR. This work was done in a wet lab located in the Bay Area.
The discovery involves a previously unknown system hidden in the DNA of bacteriophages. These are viruses that infect bacteria. The new system can cut, copy, and paste DNA, much like the well-known CRISPR technology used by researchers worldwide.
Rapid discovery with human oversight
Anthropic claims the AI did most of the heavy lifting. Claude used roughly 950 agents to search through data. This process consumed 210 million tokens and took 21 hours of concerted effort. The lab itself was only established this spring.
However, humans performed all the physical experiments. The company stated that its scientists handle only lower-level biosafety risks. They do not work with pathogens that infect humans. This trade-off ensures safety but limits the scope of what the AI can do directly.
Validation remains an open question
Anthropic CEO Dario Amodei acknowledged that the finding builds on prior work. He noted that a Stanford team previously found a similar system. The broader research community must now validate how novel this discovery truly is. Independent verification is the next critical step.
The announcement follows public admissions about AI safety risks. Amodei has expressed concerns about bioterrorism and the need for slower development. Yet, he also believes AI could cure most diseases within five to ten years. This suggests Anthropic views the potential rewards as outweighing the risks.
A competitive landscape in bio-research
Anthropic is not alone in this field. Stanford researchers recently published work on using large language models for gene editing. UC San Francisco scientists used AI to design enzymes from scratch. Google launched its AlphaFold tool back in 2020. AI is becoming a standard tool in biological research.
Despite this, Anthropic is cautious about automation. Amodei said it might be possible for AI to control lab equipment in the future. He emphasized that they are not doing this today. This careful approach reflects the industry's ongoing balance between speed and safety.
