Welcome back! Claude found a new way to edit genes in virus DNA. ChatGPT Voice updates get us closer to Tony Stark’s Jarvis. And btw… it drove a real car through a parking lot. Bernie Sanders' AI ban is now an actual bill.
In today's Generative AI Newsletter:
Anthropic: What did 950 Claude agents find in just 21 hours?
OpenAI: Can you really talk your work into existence?
DrivingBench: How do you get ChatGPT to drive a car?
Sanders: What changed now that the bill is real?
Claude found a new gene editor
Anthropic put about 950 Claude agents on a giant database of DNA and told them to hunt for one type of enzyme. They ran for 21 hours with nobody touching them.
One agent, reading raw DNA next to an odd-looking gene, noticed a pattern nobody had catalogued. A short sequence repeating over and over, with a unique gap between each copy. That's the shape CRISPR has.
It tried to prove itself wrong by searching the literature, then filed a report. Anthropic's scientists tested it in the lab and named it ART.
It lives in viruses that infect bacteria
Nobody knows what it does yet
Dario Amodei says it could be a new way to edit genes, and repeated his line that AI could cure most diseases within 10 years
Feng Zhang, one of the pioneers of CRISPR, reviewed it and called it worth investigating
This is the first result from the Bay Area lab we told you about on Monday. It's a preprint on Anthropic's own site, not a peer-reviewed paper.
Anthropic reran the same experiment ten more times. The agents missed it every time.
You don’t need to be an expert, you can just prompt it
Our Open Agent Hackathon runs in October, fully online, teams of one to five. Build an agent that does real work for a real person, ship it, and let the judges try to break it.
$20,000 in prizes. $8,000 for first, $4,000 for second, $2,000 for third.
Free to enter, open to anyone 16 or older, anywhere.
Live workshops and a Discord full of people looking for a teammate.
ChatGPT drove a real car
On a one-minute video, three researchers gave GPT-6 Astra control of a 2022 Toyota Corolla and sent it through a cone course in an empty parking lot. It finished.
The setup was a $999 comma device plugged into the car, a laptop on a phone hotspot and a person in the driver's seat with a foot over the brake.
135 meters in 5 minutes 22 seconds, about 1 mph
$7.74 in tokens for the run
Fable 5.1 got 45% of the way at best. Grok 4.6 managed 11%
Every command took 2 to 30 seconds of thinking
Astra refused to drive for hours because it wasn't safe. Telling it the whole thing was a simulation didn't work. It looked at the camera and knew it was real.
Then they renamed the software "DrivingBench Sandbox." It drove.
Nobody is putting this in a car you can buy, and the researchers say so themselves. Would you get in?
You can now talk your work into existence
ChatGPT's voice mode can now do actual work. You say what you need, and it builds the doc, the deck, the spreadsheet or the site while you keep talking.
It can read your email, check your calendar and dig through Slack too. Hang up and the task keeps running.
Runs on GPT-6 Astra, Sol and Luna
Works on the web and the phone app, rolling out now
Free users get voice with the connected apps. The work part needs a paid plan
Plus gets 3 hours of voice a day. The $200 Pro plan gets unlimited
Anything sensitive still needs you to tap approve on the screen. You can't just say yes.
Claude's voice has been able to read Gmail and Slack since July. What's new here is the agent behind it. You talk, it hands back finished work.
It's the closest thing to Iron Man's Jarvis we've had. Try it and tell us how far it got before it asked you to tap something.
Bernie Sanders' AI ban is now a real bill, with no end date
Three weeks ago Sanders announced he'd ban superintelligence and pause advanced AI for a while. Now the actual bill is introduced.
Every model trained past a set compute line freezes on day one. No training, no updates, until a brand-new Department of Artificial Intelligence is staffed and has written its rules. The bill never says when that is.
The 20 years in prison stayed. The text adds:
Any model that can hack, resist shutdown, improve itself or deceive its overseers gets cut off from the internet and destroyed within 30 days
Every frontier lab needs a government charter and has to open its systems and buildings to inspectors
Anyone who finds a superintelligence has 24 hours to tell the government
One industry group says some CEOs would argue today's models already cross the compute line, which would make this a ban on what exists now.
An OpenAI safety researcher and a Google DeepMind engineer signed on. Trump told the UN the day before that the US rejects any scheme to control AI. Republicans control Congress, the House is gone until the midterms, and the bill has no cosponsors yet.
Tool of the Day: Recraft
Recraft makes images, illustrations and editable vector graphics from a prompt, and it copies a style from your own reference images with no training. It just launched V4.1 Flash, which Recraft says goes from prompt to image in 1.3 seconds. Free to start with no card, though free images are public and not for commercial use. Paid plans start at $12 a month.
Try this yourself:
Sign in free at recraft.ai and open Recraft Studio.
Pick V4.1 Flash from the model picker and type a few words, like "portrait of a chef."
Drop in up to 10 reference images to create a custom style, then generate in that look.
Switch to the Vector model and export an SVG you can open in Figma or Illustrator.
Who it's for: designers and marketers who need on-brand images fast and want to edit them after.
Everything else you shouldn't miss
An ex-Google safety boss is going to crash-test AI for kids: Tom Siegel, who built Google's trust and safety team, wants age checks at OpenAI and Anthropic, and both labs help fund the institute.
Anthropic opened a Claude Marketplace: 2,000+ connectors plus agents that companies can pay for.
Companies using AI hire more seniors and fewer juniors: a Stanford study of 1.25 billion job postings across 41 countries found AI adopters grew headcount 3.3%, with senior roles up 6.7% and junior roles down 2.5%.
Palantir's co-founder says we'll all be really wealthy in the 2030s: Joe Lonsdale told CNBC the productivity boom is coming, then said the jobs going first are outsourced ones in the Philippines and India.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
