DeepSeek R2's specifications:
- Unit cost reduced by 97.3%, ready for immediate release
- Its self-developed distributed training framework achieves 82% utilization of 910B chip clusters
- Reaches 512 PetaFLOPS computing power under FP16 precision
- Achieves 91% efficiency compared to same-scale A100 clusters (data verified by Huawei laboratory)
It’s pretty weird that M3 Ultra is more powerful than M4 Max #MacStudio
I think Gemini 2.5 Pro is also an excellent model, but Claude 3.7 Sonnet generates responses that are closest to what I have in mind.
I was quite intrigued by the Surface Laptop 7, but after some research, I found out they only have models with Snapdragon processors. As the ARM64 version of Cursor hasn’t been released yet, I think I’ll have to pass on the Surface Laptop 7.
Just updated a .cursorrules file for Go projects
There's a rumour that Claude Code will be bundled with the Claude Max plan. If this is true, I'm more than happy. Since Claude Code is quite expensive, not having to pay for the subscription would help me a lot.
Personally, I reckon Claude Code is much better than similar AI coding agents like Codename Goose and OpenAI's Codex CLI in terms of functionality.
Claude 3.7 Sonnet Max in Cursor is a significantly improved version of the standard Claude 3.7 Sonnet. It features an expanded context window of 200k tokens and the capability to chain up to 200 tool calls, enabling thorough analysis and enhancement of extensive codebases that were previously too large to handle in a single operation.
I'm a big fan on TIDAL because it offers higher royalty rates to artists, paying approximately $0.013 per stream as of 2025. This is more generous than other services like Apple Music and Spotify. TIDAL has also introduced a system where a portion of subscription fees can be directed to specific artists chosen by subscribers.
starting the day with x japan!
@jack i gave codename goose a go in cli mode and found it very good, but i’ve got a little feedback. i reckon the readability of prompts and responses needs a bit of improvement. i find it a bit hard to find out where responses start from because my prompts don’t stand out.
When I request AI to create code for a feature I need, Claude 3.7 Sonnet typically generates code that I like the most.
What OpenAI models are optimised for agentic coding?
Starting the new week with X JAPAN!
I’m thinking of building a desktop app for chatting with OpenAI’s models using its API, and I’m planning to code with Codename Goose.
I reckon the CLI is better than the GUI when it comes to agentic coding.
Having been coding with Claude Code since its launch, I find GUI-based IDEs like Cursor and Windsurf rather troublesome.
you can use google ai mode through a vpn even if you’re not in the us
it appears OpenAI’s Codex CLI is a rival to Anthropic’s Claude Code. as far as the Codex CLI GitHub page can tell, Claude Code seems to be a bit ahead. i’m going to give the Codex CLI a go myself and see how it performs.
what i love about gemini 2.5 pro is that it outputs a markdown document in a code block properly
