leerob

We plan to eventually have another model at that weight class, but right now trying to train the best possible model.

Probably can't advance frontier math yet, yeah. But please let us know other places you want to see Grok improve for future models!

We are and will continue to.

(I work on Grok) This isn't allowed. CSAM / deepfakes are against our acceptable use policy.

(I work on Grok) We've been working on teaching the model how to reason about great visual design principles. Obviously this is hard and somewhat subjective, but through a combination of writing down these principles (e.g. how to think about systems, not just "use this italic serif font on marketing pages"), and then creating a lot of data to pairwise compare designs/outputs, we've made a notable improvement over G4.5 and see a path to improving much further in the next model.

It works with the highest tier Cursor/Grok accounts during the beta.

It's quite a bit different, namely that ChatGPT Work has both local conversations and cloud agents. But for each cloud agent, you are spinning up and tearing down a new VM each time. This is an always-on Linux box, which stays logged in. Additionally, your bots can talk to each other (although Codex did have the ability to reference threads, I am not sure if one thread could send messages to other threads).

You can read our full technical report here: https://cursor.com/blog/composer-2-technical-report

(I work at Cursor) We score well on Terminal-Bench and SWE-bench Multilingual. DeepSWE, not so great yet, as it's more for very long-horizon tasks. We're planning to include more public benchmarks in our next model release.

(I work at Cursor) When Composer 2.5 launched, we initially scored very competitively on AA's composite benchmark. I believe 3rd place overall. They have recently updated to use DeepSWE, which has more of a focus on very long-horizon tasks, and Composer isn't as good at those yet. We're aware and working on this for our next model.

Overall, some benchmarks show Composer doing well, others not so much. We think the model is very capable at the given price point. There's lots to improve! If you see any specific behaviors or places the model isn't very good, lmk here or can email me lrobinson at cursor.com.

(I work at Cursor) CursorBench includes many evals from actual engineering tasks from the Cursor team, which include our private codebase. This codebase is held-out from training so models haven't seen it, including Composer.

Will do. On it.

(I work at Cursor) Sorry about this, we should have made this more clear. The new privacy mode is needed because we have to store some state to enable running agents in the cloud. If you don't want to use cloud agents, you can continue using the legacy privacy mode. Currently the mobile app requires this new privacy mode and won't work without it. We're pushing an update right now to make this more clear in app and can help you get reverted back to the legacy version on the support thread.

Glass was a codename while the UI was in early alpha with testers. It redirects to download now because there is no special link anymore. It's just part of Cursor 3 itself.

I'm an engineer at Cursor, can try to clarify questions here.

> I wish they'd keep the old philosophy of letting the developer drive and the agent assist. Even when I'm using AI agents to write code, I still find myself spending most of my time reading and reasoning about code.

We very much still believe this, which is why even in this new interface, you can still view/edit files, do remote SSH, go to definition and use LSPs, etc. It's hard to drive and ship real changes without those things in our opinion, even as agents continue to get better at writing code.

> I'm hoping in this new UI in v3 I can still get that experience (maybe it's just hidden behind a toggle somewhere for power users / not shown off in the marketing materials).

This new interface is a separate window, so if you prefer the Cursor 2 style, that continues to exist (and is also getting better).

We used a Kimi base, with midtraining and RL on top. Going forward, we'll include the base used in our blog posts, that was a miss. Also, the license is through Fireworks: https://x.com/Kimi_Moonshot/status/2035074972943831491

Are there other coding benchmarks we should include next time? We included Teminal-Bench 2.0 and SWE-bench Mulitilingual.

We don't plan on reporting SWE-bench Verified, for similar reasons to OpenAI: https://openai.com/index/why-we-no-longer-evaluate-swe-bench...

You can disable this if you want, it's under "Inline Diffs" in the Cursor settings.

We've found it to be a strong mix of speed and intelligence. It scores higher than Sonnet 4.5 on Terminal-Bench 2, maybe we will post more on this later.

There's a setting to turn this off if you prefer (Cursor Settings > Agent > Attribution).

> in particular the extra layer of diff-review of AI changes (red/green) which is not integrated into git

We're making this better very soon! In the coming weeks hopefully.

(I work at Cursor) We have all these! Plan mode with a GUI + ability to edit plans inline. Todos. A tool for asking the user questions, which will be automatically called or you can manually ask for it. Hooks. And you can use Opus or any other models with these.

> The JS engine used a custom JS VM being developed in vendor/ecma-rs as part of the browser, which is a copy of my personal JS parser project vendored to make it easier to commit to.

https://news.ycombinator.com/item?id=46650998

Should compile now: https://news.ycombinator.com/item?id=46650998