Claude Agents Find ART, a Phage Enzyme System With CRISPR-Like DNA Repeats

Cool result, but where are the actual results? With the function still unknown I wonder whether they have a Nature letter queued up or know another AI lab is about to publish the same find.

In a year or two posts like this are either artifacts of peak hype or the opening shot of the singularity.

I'd bet on the singularity reading. OpenAI now has an AI working as a research intern, so the recursive self-improvement loop Hinton describes may already be starting.

That's black and white thinking. I expect a midgularity: slow, partial, neither hype nor apotheosis.

My worry is the next post is some lab saying they were about to publish this and had Claude proofread the paper. Wonder how any of that ends up in training data.

The post itself says the reverse transcriptase was noticed before. Only Claude commented on the repeating subsequence, so a lot of this is dots that were hidden in plain sight.

I can't shake how much the Alpöge and Buckmaster episode damaged the perception of AI-driven science. Whatever you think of them, every claim since has been filtered through that anxiety.

This shows biology is a much harder problem area for LLMs than math. Finding RTs is tedious but doable, so they scoped the problem way down from anything like a Navier-Stokes of biology. Glad they're trying, even if it's marketing.

It's really about iteration and validation. Models progress in code and math because they can write tests and proofs. Fields that need physical experiments lack that feedback loop, so closing it is the whole game, and Anthropic building its own lab is how they push partners to invest.

This domain can get far more dangerous than some datacenter hacking. Where did the bio doomsday fear go?

LLMs are very good at finding patterns in huge datasets. Google had similar breakthroughs and simply stopped marketing them.

Yes and no. You can't dump DNA into a context window and call it a day. The blog describes a tool-calling agent session, with actual ML models doing the sequence work as tools the LLM calls.

Clear read on the strategy: Dario thinks the fix for AI's PR problem is curing cancer. Expect heavy promotion of every step, however small or far from commercialization.

Even with recursive self-improvement you don't get new treatments to market instantly. Real world testing takes a long time and I see no shortcut around it.

On 'unavoidable': physics says nothing that forces cancer cures to be decades away. Plenty of places will wave trials through for a fistful of dollars if it buys them an edge.

One more worry: Anthropic using Claude's reputation to launder known approaches to aging and cancer that society hasn't accepted yet, with enough marketing to get people trying them.

Either way it's an uphill battle. Job loss stories, datacenter fights, eminent domain blowups, slop everywhere. They have to show the goal is helping humanity and that everyone else isn't collateral damage, because those people vote.