They're training the next model from scratch using 10x more compute, backed by Colossus 2's million H100-equivalent capacity. Reports say they're using tens of thousands of xAI GPUs for this new model.
Cursor's harness already makes the Claude and GPT models perform better. Imagine what they could achieve with so much more compute.
https://cursor.com/blog/composer-2-5
Cursor just released Composer 2.5 and announced they're training a much larger model from scratch with xAI on Colossus 2. This could be huge.
Composer 2.5 improvements:
• Better at sustained work on long-running tasks
• More reliable with complex instructions
• Improved communication style and effort calibration
• Built on Moonshot's Kimi K2.5 with additional RL training
The bigger story is what comes next.
Here's my new episode with @alexalbert__, who shared an inside look at how Anthropic is building the next Claude. We talked about how the research team: → Plans for the model and harness together → Uses Claude to turn user feedback into evals → Trains Claude's character &
What really got me is how heavily they’re using Claude to analyse user feedback. It clusters themes, generates synthetic versions of recurring issues, and converts them into evals. The model helps them improve the model.
Right at the end, the consciousness question came up. Alex called it a “big question” they’re working on. No detail, but the fact it’s even on their roadmap tells you something about where they think this is going.
https://youtu.be/T4ieZPIEmd8?si=3KXce86LjNOC2R__
The interesting part is they’ve got strong intuitions based on architecture choices, but they still don’t truly know what the model will be good at until it’s deep in training.
They’re also working on memory in a way that feels more human. When the agent isn’t running tasks, it revisits memories in the background - spotting contradictions, pruning, tidying, cleaning up. Alex called it “dreaming”. If it works, that’s genuinely smart.
Anthropic shared how they’re building the next Claude model. A few things stood out.
They treat every model like a product launch. Before training even starts, they write a spec for the capabilities they want - coding, knowledge work, spreadsheets - then track, step by step, whether training is actually delivering on those targets.
Thanks for the feedback on Codex in the ChatGPT mobile app. While it’s in preview, we’re working to improve it fast. What you can expect next: push notifications, /fork, ability to restore after revoking, better reconnects, fixing the ability to control other devices, fewer
The real issue: when credits reset, the agent should just continue automatically. That's what an autonomous agent should do. Instead, you have to manually intervene every time. At minimum, there needs to be a "force continue" button to get things moving again.
I want Codex to work. The developers clearly care and they're shipping regularly. But these workflow blockers keep dragging the experience down, especially on Windows where the performance gap is noticeable.
