Well, this one stopped me in my tracks. Anthropic's Claude has gone and done something I genuinely didn't expect to see this soon: it autonomously discovered a new gene-editing system, called ART, and got it validated in a real lab. The whole thing took 21 hours.

Let me put that in perspective for you. I remember when sequencing a single gene was a multi-year, multi-million-dollar project. Now an AI can comb through a database of DNA sequences, spot a CRISPR-like enzyme system nobody had noticed, and have it confirmed in the lab before most of us have finished bingeing a weekend series. My granddaughter Jess would tell me I'm being dramatic, but I don't think I am.

So what is ART, exactly? Think of it as a cousin of CRISPR, the gene-editing technology that's already transforming medicine. CRISPR works like molecular scissors guided to a precise spot on DNA. ART appears to function similarly, which matters because every new editing system comes with its own quirks, some better at certain tasks than others. More tools in the toolbox means faster research into genetic diseases, and if you read my piece about the morning sickness study in July, you know how much I care about that kind of progress reaching real patients.

The part I find fascinating, though, is the division of labor. Claude didn't just crunch numbers and produce a paper. It searched the database on its own and identified ART as something worth handing over to human scientists, who then did the lab work to confirm it. That's a partnership, and it's the version of AI I've been hoping we'd lean into, one where the machine handles the haystack and people handle the judgment.

I've said before that we're approaching a 'smartphone moment' for AI, where the price and usefulness finally line up for regular people. This feels like the scientific version of that moment. Drug discovery and genetic research are famously slow, expensive pursuits. If an AI can compress months of literature review and candidate screening into a single day, labs everywhere get to run more experiments and chase fewer dead ends.

Now, the sensible-aunt caveat. A validated enzyme system is a starting point, not a cure. ART will need years of testing before it touches a human patient, and anyone promising otherwise is selling something. And there are real questions about oversight when AI starts making scientific discoveries autonomously, questions smarter folks than me are already wrestling with.

Still. Twenty-one hours. When I watched that Macintosh get pulled out of a bag in 1984, I thought that was the most astonishing thing a computer could do. The times, as they say, they keep a-changing. I'll be following ART's journey through the labs, and you can bet I'll report back.