Live data from Hacker News

Improving Composer through real-time RL

cursor.com

41–45 of 45 posts

Re: Improving Composer through real-time RL

#41

Important disclaimer for anyone using Cursor: make sure to disable "data sharing" in your account settings, as it is enabled by default and old accounts are automatically opted into it.

Devil's advocate: why? Why do people need to make sure to disable it? Is this the "but privacy!111" stance, or the Dog in the Manger stance ("I'm not being paid for it, so why should anyone else benefit?"), or...?

Always appreciate people seeking clarity on positions that are not clearly elaborate/taken for granted. Kind of stresses me out sometimes.

Re: Improving Composer through real-time RL

#42
What training do they claim to make since Composer 2 is just Kimi K2.5? Do they have a collaboration with Kimi team to help with training.

I use Cursor heavily and pushed for its adoption at work a year ago, but with every day passing I like and trust it less and I am beginning to think about alternatives.

Re: Improving Composer through real-time RL

#43

Important disclaimer for anyone using Cursor: make sure to disable "data sharing" in your account settings, as it is enabled by default and old accounts are automatically opted into it.

Devil's advocate: why? Why do people need to make sure to disable it? Is this the "but privacy!111" stance, or the Dog in the Manger stance ("I'm not being paid for it, so why should anyone else benefit?"), or...?

> Is this the "but privacy!111" stance

Mostly, yes. But let's unpack this a little: these companies often claim a massive amount of data is being collected "to improve our products" and "to make the world a better place." In reality, the data is more often used to deanonymize the user and build a precise profile of them. It is subsequently sold to third-party companies who specialize in this sort of business and are equipped to extract the most value possible from the data. In the most benign form, this is done by targeting the user with personalized ads.

Nowadays, this data-mining process is almost completely automated, and there are ways to legally cover your ass by stripping the datasets of too-directly identifiable information. If you audit these databases, your name, phone number, email, and what you like having for dinner are probably not going to appear together in a single row. However, a direct mapping is trivial to recover by correlation and inference. The profile that can be built from this data is usually very precise.

People who say "muh privacy!!111" tend to find this whole process fundamentally icky. I personally don't like opening an IDE like Cursor and feeling like I'm stepping into some KGB-era hotel room with microphones in the walls and a two-way mirror in the bathroom.

Of course, it is completely up to you to disagree and continue uploading your codebase, prompts, and session data to Cursor's servers and give Cursor permission to inspect this data and send it to third parties at their leisure. I am merely giving my advice, which I admit is very biased.

Personally, what I find the most reprehensible is that this feature is opt-out instead of opt-in. I strongly suspect that most users would not agree to enable data sharing if they were asked directly.

Re: Improving Composer through real-time RL

#44
post #31

Earlier quoted context omitted.

Across my 4 different gpt subscriptions (personal, personal cursor, GitHub Copilot and cursor) all gpt5 models are junk compared to v4 - constantly ignore prompts, skills, can't write c# or powershell properly the first go, up to 5 tries. Qwen3 hands down beat it on a ryzen 5800 and 6700xt GPU even though it's slow it got the code right first try. I feel like the v5.0 preview did ok but it's slid all the way down the…

Saying gpt 5.4 is like gpt2 is wild.

Lol, audibly.

I'm glad AI curmudgeonry on HN has shifted from "it doesn't work, scam, they made the deployed model worse with 0 communication" to something more akin to "why does anyone use mac or windows, nix is peak personal computing"

Re: Improving Composer through real-time RL

#45

Real-time or continuous learning is great on paper, but to get this to work without extremely expensive regression testing and catastrophic forgetting is a real challenge. Credit to the team for taking this on, but I’d be skeptical of announcements like this without at least 3–6 months of proven production deployments. Definitely curious how this plays out.

Can this be also used as an attack vector? A small seed percentage of users constantly choosing a particular poisoned pypi library to achieve a niche task which gets rled into the model suggestions and recommendations.

The recent claude code leak also revealed that they're poisoning their competitors via anti-distillation policies baked in claude code CLI (fake tool calls, adding noise etc).
Post reply on HN