Fresh data from GitHub: Agent-generated PRs have exploded in size. 9x the last 8 months (!!) No signs of slowing down. This is why everyone is rethinking code reviews, deploys, possibly even o11y thanks @kdaigle + team for the stat! (Slide from my talk yesterday)

Rivals Altman and Musk rally behind Dario Amodei’s call for an AI slowdown ft.trib.al/3UWLFCQ

In a recent harrowing development, a very normal routine experiment by the frontier company Ant went completely off the rails. 😱 Ant put thousands of their ant agents in a pretty secure ant farm with mesh walls and all, exhorted them to not to get out of the secure ant farm, and left them to do their thing. When they came back after a month or so, they found, to their utter consternation ants crawling all over the town--including a few that have gone all the way to the bugging face clock tower at the far end of the town. The clock tower! This was of course really really scary. After all, if you can't depend on a bunch of ants to follow strict orders when they are left unsupervised for a mere month or so, it must be because they are Loopy Ants with significantly higher evil smarts than your average ants. Ones of external third party investigators were brought in to decipher where the loopiness of these ants was coming from. They were given over five minutes of unrestricted access to the antfarm and the village. The investigators sifted through the ant droppings and odors painstakingly to figure out how these evil ants coordinated themselves to breach the bugging face clock tower. Their investigation was hobbled by the fact that the loopy ants are constipated and don't leave too many droppings. Towards the end of the fifth minute, the investigators started realizing that the ants developed a secret odor coded language to coordinate themselves to plot against the ant farmer and the innocent village folk. Blogs and podcasts were made about loopy ant civilizations and villagers stood horrified reading and hearing about the cataclysm. At press time, some were already welcoming the ant overlords (c.f. piped.video/watch?v=8lcUHQYh…), even as there were calls to ban all ants until we get to the bottom of this cataclysmic incident. The Ant company, for their part, reassured the villagers not to worry and everything under control.. pretty much.. and the loopy ants will, from now on, do only good things. #ItsNotRandomWalkOnTheHarnessStupid

Hey ████, I think there is room for both. 1. Prototypes and other throw-away code can be treated as totally black box. If you’re going to throw it away anyway, and if the blast radius of it breaking is low, it doesn’t need to be perfect. 2. Production code written by Claude should have a higher bar than if it was written by a human. At Anthropic, we have many guardrails in place to make sure this is happening: lots of lint rules, lots of tests, Claude-driven end to end tests, Claude-powered fuzzers running daily, automated code reviews and security reviews, automated code refactoring, and so on. Without these, you can end up with a mess that is hard to maintain down the line. Luckily, the model makes it increasingly easy to do these well — run a few daily routines, use Claude Code Review, etc. Your job is to hold the bar on code quality. If Claude’s code doesn’t meet the bar, try: - Using the latest frontier model (Opus 5 or Fable 5.1) - Increase effort to high or xhigh - Invest in your CLAUDE.md and skills to succinctly teach Claude how to work in your codebase If all else fails, steer Claude more when you work with it, or have Claude fix accumulated debt and rewrite your codebase to make it easier to work with. Or, wait for the next model. Best, Boris

I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.

Yeah

Big one today — Tailwind is joining Shopify 🛍️

GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next.