Improving Composer through real-time RL

https://cursor.com/blog/real-time-rl-for-composer

by ingve • 4 months ago

98 34 4 months ago

cursor.com

polishdude20• 4 months ago

I'd love to see some data for how much it has improved via this process in the last week

heliumtera• 4 months ago

It would be the same as kimi k2.5, the underlying model

CitrusFruits• 4 months ago

I've been wondering how they've been able to be so generous with Composer usage with it still making business sense. Seems like this is the answer: presumably they think they'll have a competitive advantage in not just the UX space but the model space as well soon. It's a great strategy, but I do wonder if the moat will be big enough with how fast things are moving and how competitive the model landscape is.

ketzo• 4 months ago

After seeing the last few releases for GPT and Claude, I’m not sure how anyone (else) is gonna build a durable advantage on proprietary model quality.

The capabilities of the top labs’ models have improved so much in just the last few releases, and I definitely foresee a world where they gate those models away behind 1st-party harnesses/tooling.

hypercube33• 4 months ago

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 hill to gpt 2 or 3 levels for me.

kgeist• 4 months ago

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. [0]

And still no mention of Kimi in a new blog post :)

Also apparently the inference provider they use, Fireworks AI, already has built-in API for RL tuning Kimi [1], so I wonder which parts are Cursor's own effort and where Fireworks AI actually deserves credit, especially since they repeatedly brag about being able to create a new checkpoint every 5 hours, which would be largely thanks to Fireworks AI's API/training infrastructure.

I mean, I'm genuinely curious how much effort it would actually take me to go from "here, lots of user data" to "the model gains +1% on benchmarks" to produce my own finetune, assuming I already use a good existing foundational model, my inference provider already handles all the tuning infrastructure/logic, and I already have a lot of usage logs.

[0] https://news.ycombinator.com/item?id=47459529

[1] https://fireworks.ai/blog/kimi-k2p5

fzysingularity3• 4 months ago

What do you think actually happened here in the past week?

They used Kimi, failed to acknowledge it in the original Composer announcement. Kimi team probably reached out and asked WTF? Their only recourse was to publicly disclose their whitepaper with Kimi mentioned to win brownie points about being open about their training pipeline, while placating the Kimi team.

fzysingularity3• 4 months ago

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.

ghywertelling• 4 months ago

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.

fzysingularity3• 4 months ago

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).

janalsncm• 4 months ago

Back in my day we called this real time training from implicit user feedback.

The engineering challenge here is an order of magnitude bigger though. An LLM is orders of magnitude bigger than a recommender system model. Kudos.

amazingamazing• 4 months ago

seems expensive. distillation is inherently impossible to defend against. sit back and let your competitors do the hard work. they'll whine and say it's illegal, but they shouldn't complain, they will reap what they sowed.

crazylogger• 4 months ago

This feels so wrong. the LLM should play the role of a very general (but empty & un-opinionated) brain - you don’t want to perform a coding-specific lobotomy on someone every day. The proper target of their RL should have been their harness. That’s what determines the agent's trajectory as much as the base model.

I also wonder since they’re doing constant RL on model weights with today's Cursor design, does that mean they can never change their system prompt & other parts of the harness?

1) Comparison between past trajectories data would be meaningless if they were operating under different instructions.

2) Performance will be terrible the next time they change their tool design, since the model is now "opinionated" based on how a previous version of Cursor was designed.

Anthropic is more sensible with their “constitution” approach to safety. The behaviors (and ultimately the values) you want your model to follow should be a document, not a lobotomy.

hmartin• 4 months ago

Step 1: take an open source model with zero acknowledgement.

Step 2: build on someone else's infrastructure innovations with zero acknowledgement.

Step 3: Write a blog post with "unprecedented" and "100x" and "trillions" in the first paragraph.

Seriously, this seems like cool work and enjoyed the post. But my basic trust in them has completely tanked.

refulgentis2• 4 months ago

I’m not familiar with 2 of these 3 stories so I’m mostly just impressed by the RL turnaround (4 hours!?) and sharing knowledge on the fly off using fine tuned models at scale.

For the gossipy part, I love Kimi, but find it hard to get worked up about them not labelling their model Kimi when Kimi was the base. Especially because Kimi…has had…some issues…being able to distinguish itself from Claude…

m13rar• 4 months ago

I don't though it works and these guys are raking in the money. The wider audience of devs or people paying up for Cursor don't care about the morals old school OSS devs hold.

nickphx• 4 months ago

Define raking in the money. Operating at a profound loss, fueled by untenable investments and hype.. Where are they raking in the money?

htrp• 4 months ago

If the model "improves" every 5 hours, how do you have any guarantee of model consistency across long coding sessions?

annie511266728• 4 months ago

Yeah, this feels tricky

If the model changes every few hours, we’re basically debugging against a moving target - and that gets expensive fast.

dbbk• 4 months ago

What do you need consistency for

pillsburycat3• 4 months ago

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.

natpalmer1776• 4 months ago

Do you have evidence for those claims? I don’t mean to be contrary or subversive, I’d just be interested in seeing how this is actually taking place.

pillsburycat3• 4 months ago

To be very precise about session recording: you can inspect the Cursor binary and see that it comes bundled with rrweb and full telemetry infrastructure setup is in place: mouse movements, clicks, scroll positions, etc. on top of the codebase and prompts being sent over the wire.

However, I have edited my other claims for now and you can consider them provisionally retracted. My original advice about turning off data sharing stands. You are right to ask for more evidence given the severity of the claims. I think this merits a deeper dive, and a throwaway hacker news comment might not be the best channel for it. Stay tuned ;)

TeMPOraL• 4 months ago

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...?

simjnd• 4 months ago

Also from a security perspective. People have been able to extract copyrighted code / API keys some LLMs have been trained on before. If you opt-in to this, your / your company code will be used to train and improve the model. People may then be able to extract that from the model. Another threat vector.

timcobb• 4 months ago

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

pillsburycat3• 4 months ago

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.

vicchenai• 4 months ago

the rl loop here is clever but i wonder how the reward signal degrades over time. if you're optimizing for user acceptance of suggestions, you're inevitably training on a mix of "this was actually correct" and "i accepted because editing the suggestion was more work than accepting it." that second case creates a subtle bias toward suggestions that are close-enough-to-not-bother-fixing rather than actually correct.

also curious whether they see different convergence patterns across languages. my gut says something like python where theres more stylistic variation would be harder to get a clean reward signal vs something like rust where there are fewer idiomatic ways to do things.

addcn• 4 months ago

Yeah. Great callout. You basically need signal all the way to production. Did it get nuked during a code review? Is it still running 2 mo later? How many incidents were caused by this code?

This is tough though because enterprises go absolutely ballistic over “training on our data” - which is understandable, but will also hold us back.

meric_• 4 months ago

What are with these comments? Did no one even read the article. Not once is composer 2, or kimi, or anything mentioned... because the article is about Composer 1.5.

I mean sure the techniques are probably the same in 2 but its not like they're exactly advertising composer 2 here lol

DeathArrow• 4 months ago

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.

Improving Composer through real-time RL

Related Stories