sonnet 4.5 seems to show major improvements in both alignment and performance in a matter of months. kudos to anthropic for demonstrating that performance and safety are not in fact mutually exclusive. cannot wait to play around with the model.

🎉🎉 Today we launched Claude Sonnet 4.5, which is not only highly capable but also a major improvement on safety and alignment x.com/claudeai/statu…

Sonnet 4.5 is really good, I really like it. I didn't expect such a big jump in (what I assume to still be) the current Sonnet model size class. It feels genuinely better, not benchmaxxed. And I don't mean just coding, also in roleplays it is a fresh and interesting model.

Just found out if you try to delete your Sora app account you will lose your chatgpt account and be banned forever from signing up again.

Sonnet 4.5 is an absolutely beautiful model

you have 2 years to consume as much short form content as possible in order to escape the permanent underclass

the guilded age of slop

Sonnet 4.5 is the least sycophantic AI I've used so far. It'll actually challenge you and share its objective opinions which makes it a great thought partner. Anthropic cooked.

very strong agree here. have to admit I was a bit skeptical before walking into (or more accurately, my skepticism brewed as I stood in line for, and then walked into) the pop-up yesterday. how much brand equity can you really get out of free coffee, a hat, and a free copy of a

spent 30 mins with claude this AM after spending the last ~6 months predominantly with chat. this campaign is so powerful because sonnet 4.5 said things like this to me - which really did make me stop and put my thinking cap on

dwarkesh is experiencing first hand the PhD candidate experience of arguing with your PI on the theoretical implication of your thesis

My thesis, 𝘈 𝘵𝘩𝘦𝘰𝘳𝘺 𝘰𝘧 𝘵𝘩𝘦 𝘤𝘰𝘮𝘱𝘶𝘵𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘱𝘰𝘸𝘦𝘳 𝘢𝘯𝘥 𝘭𝘪𝘮𝘪𝘵𝘢𝘵𝘪𝘰𝘯𝘴 𝘰𝘧 𝘭𝘢𝘯𝘨𝘶𝘢𝘨𝘦 𝘮𝘰𝘥𝘦𝘭𝘪𝘯𝘨 𝘢𝘳𝘤𝘩𝘪𝘵𝘦𝘤𝘵𝘶𝘳𝘦𝘴, is now online:

Excited to share that Reflection has raised a new round to build open-weight LLMs! Joining Reflection post PhD has been a deeply rewarding experience – it’s a privilege to work with this team on something ambitious, interesting and that I believe in. We’re hiring – DM me.

One Tuesday night, as my wife and I sat down for dinner, a sheriff’s deputy knocked on the door to serve me a subpoena from OpenAI. I held back on talking about it because I didn't want to distract from SB 53, but Newsom just signed the bill so... here's what happened: 🧵

- pretrain on math and reasoning dialog - mid train on math and reasoning traces - release as base model LLM RL Researchers: WE HAVE DISCOVERED ALIEN INTELLIGENCE BEYOND OUR COMPREHENSION WITH RLVR!

3 trains an hour for metropolitan transit is tragic

🧵 Haiku 4.5 🧵 Looking at the alignment evidence, Haiku is similar to Sonnet: Very safe, though often eval-aware. I think the most interesting alignment content in the system card is about reasoning faithfulness…

Excited to start OpenAI for Physics w/ Alex Lupsasca Kevin Weil 🇺🇸 Aleksander Madry and Jakub Pachocki! I sat with Alex Lupsasca when GPT-5 reproduced his latest research paper, and we both felt parallels to watching AlphaGo play move 37. It's nearly impossible to be a world class chess player

Fine-tuning APIs are becoming more powerful and widespread, but they're harder to safeguard against misuse than fixed-weight sampling APIs. Excited to share a new paper: Detecting Adversarial Fine-tuning with Auditing Agents (arxiv.org/abs/2510.16255). Auditing agents search