1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.

Why is no one talking about this? 🤯 Claude just solved a 25 year old internet mystery, and then doubted itself > The security behind your bank, logins and the websites you use every day was built using secret numbers nobody could explain > For 25 years people feared they hid a backdoor > Claude cracked it in under 30 minutes: they came from a basic old computer trick, no sign of anything hidden Then Claude stopped and asked itself, hmm, is this even real? It even wondered why something this obvious wasn't found in 25 years. I genuinely think that's the scariest and coolest part.

Hot take: The more you let AI code for you, the less you’ll know how to code yourself.

We are 48 months into AI being 6 months away from taking your job.

i know a dude who doesn’t read code or even use a review tool he just chats with an llm all day, tells it to deploy everything when he’s done and walks away from his computer

SITUATION EXPLAINED: Ex-OpenAI researcher Nick Cammarata says we can now understand most neurons inside AI models, even at trillions of parameters. • Nick Cammarata, now at Goodfire, co-discovered the famous "cat and car" neuron with Chris Olah in 2019 • Neurons often seemed to fire for unrelated things like cat faces and car fronts. Superposition blamed models cramming in more concepts than neurons, which would make them nearly unreadable • His new essay says that was misread. The cat-and-car neuron detects a V shape, and most mixed neurons grow by "recursive generalization," like "iy" to "DIY" to "do it yourself" • After ordinary architecture fixes, almost every neuron in Inception v1's last layer became readable, and about 90% of Qwen3-32B's neurons have one main job • His "bitter lesson for interpretability" is that scaling ordinary investigation with AI research agents beats clever new methods @theojaffee: "Certainly we have not been researching interpretability at anywhere near the scale with which we have been researching large language models, scaling AI capabilities. So how would we know that ordinary methods don't scale? We don't."