Truth
30 posts
Wait, Vercel has lower "ARR" than Lovable? Is this real?
English
@sethbannon This is not correct.
They simply play the spray and pray game.
It is NOT about "giving as many founders as possible a launching pad".
No idea who ingrained this idealistic fallacy into you, but it is what it is: A fallacy.
English
It's not fair to blame YC for backing apparent frauds like Delve.
They only spend 10 minutes with founders over zoom before investing. It's a cursory screen.
They're not trying to deeply vet founders -- they're trying to give as many founders as possible a launching pad.
English
@ZixuanLi_ But I appreciate that you mentioned "luck".
So it appears that you recognize that any positive returns would converge to negative returns if one lets the law of large numbers play out reality.
English
@ZixuanLi_ This isn't how financial markets work.
It's the ever-returning fallacy of comparing publicly accessible financial markets with other markets.
Financial markets are the most competitive markets by far.
Any edge that a public AI model gives becomes worthless almost instantly.
English
Whether driven by luck or analytical capabilities, GLM-5 is currently the only model outperforming the Human Baseline on predictionarena.ai.
Anyone using GLM-5 for trading? Does it feel capable to you?
English
@toddsaunders I think the YC model is obsolete.
Not - will be.
It is already.
LLMs changed the whole game.
English
YC’s model works for a specific founder.
Technical, Bay Area adjacent founders who want to play a zero sum game.
But there’s a massive founder group being ignored.
The operator who built a $5M-$50M services business and knows exactly what software their industry needs.
They know the pain and workflows better than anyone else.
And that founder has an industry following / friends to sell the product to.
Claude Code just gave them all the power.
And I’m going to surround them with the best people I know to help turn their domain expertise into a giant software company.
English
The fact that AIs tend to answer you in bulleted lists tells us something important, though somewhat depressing: people can't read. They don't do this by accident. What you're seeing is an implicit portrait of the median user.
English
After using Composer 2 pretty much all day, I gotta say... this is an amazing model! ❤️
English
@ctjlewis Stanford's non-PhD programs have been far from research-level for quite some time now.
They shifted towards practically oriented skill learning.
Most CS undergraduates don't even do any proof heavy math anymore, and theoretical CS is also cut short severely.
It's not elite.
English
@TeksEdge @novita_labs @OpenRouter Yeah. Novita may be ZDR indeed. However does Novita confirm that all their models are hosted by themselves? Because obviously they cannot control logging of third party inference providers.
That said, it may well be that Novita has a non-public commercial agreement with MiniMax.
English
English
📉 Bad News: MiniMax M2.7 is closed weights. No HuggingFace release = no local inferencing.
🛡️ Good News, you can still use it with 100% privacy.
Use @novita_labs (US SF-based) via @OpenRouter.
They operate under a Zero Retention agreement:
🚫 No logging inputs
🚫 No storing chat history
🚫 No training on your prompts
📜 “You acknowledge and agree that we shall not log, store, or retain any User Content, including but not limited to your Inputs, prompts, chat histories, or any Outputs generated by the Services, unless we have obtained your prior written consent for specific instances.”
If we can't get open weights, zero retention is better than nothing. 🤝
English
@bennpeifert Crazy how y'all are obessed w bro. What about the 1000+ other successes from YC?
English
new rule: jail time for Garry Tan every time a YC startup is convicted of fraud
English
@jessfraz Kimi appears to be VERY generous here that they permit this
Composer 2 is not available in the publicly accessible Fireworks AI Model Library.
By my understanding, this would mean that Cursor uses Fireworks as a compute provider and so Cursor would have to expose it themselves.
English
This is nuts to me, the one thing Moonshot (kimi creators) asks you do is say that they are the base. Like just say it, what's the big deal, everyone already knows! It's insane. At this point not saying it makes you look like the baddies.
English
@leerob They are VERY generous here.
But congratulations.
Based purely on the license wording, legally speaking, they could probably demand much more out of this, since you are primarily using Fireworks as a compute provider for a closed weight model derived from Kimi K2.5.
English
Here's confirmation the license is correct from the Kimi team.
Agree with the feedback we should have mentioned the base up front, we will do that for the next model!
x.com/Kimi_Moonshot/…
Kimi.ai@Kimi_Moonshot
Congrats to the @cursor_ai team on the launch of Composer 2! We are proud to see Kimi-k2.5 provide the foundation. Seeing our model integrated effectively through Cursor's continued pretraining & high-compute RL training is the open model ecosystem we love to support. Note: Cursor accesses Kimi-k2.5 via Fireworks' hosted RL and inference platform as part of an authorized commercial partnership.
English
Yep, Composer 2 started from an open-source base! We will do full pretraining in the future.
Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training. This is why evals are very different.
And yes, we are following the license through our inference partner terms.
Fynn@fynnso
was messing with the OpenAI base URL in Cursor and caught this accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast so composer 2 is just Kimi K2.5 with RL at least rename the model ID
English
@dzhulgakov @FireworksAI_HQ @leerob Thanks for clarifying.
Kimi appears to be VERY generous here tbh.
Legally speaking they could demand much more out of this based on the license wording most likely.
English
@NumCupboard @FireworksAI_HQ @leerob We've been working with Kimi and Cursor from the beginning, it's all good
x.com/lqiao/status/2…
Lin Qiao@lqiao
I’m extremely excited about working together with @cursor_ai and @Kimi_Moonshot to pushing the most cutting edge quality and inference performance – Kimi-k2.5 served as the foundation for Cursor continued pretraining & high-compute RL, powered by @FireworksAI_HQ RL infra & ultra-fast inference stack, delivering frontier coding performance.
English
Data flywheel is the moat. Open base models + fine-tuning enable frontier quality/speed/cost tailored specifically for *your* AI product
@FireworksAI_HQ RL platform takes care of GPUs/algo, you focus on crafting the best model quality. We powered Composer 2 RL run at a huge scale and would love to help your use case.
clem 🤗@ClementDelangue
Looks like it’s confirmed Cursor’s new model is based on Kimi! It reinforces a couple of things: - open-source keeps being the greatest competition enabler - another validation for chinese open-source that is now the biggest force shaping the global AI stack - the frontier is no longer just about who trains from scratch, but who adapts, fine-tunes, and productizes fastest (seeing the same thing with OpenClaw for example).
English
@scottwww @adcock_brett @mcuban > They are ...
No they are not.
> It makes zero sense ...
It does make sense.
You are satisfied with finding local optima among infinitely many possible local optima.
I want to find the global optimum.
Big difference.
English
@NumCupboard @adcock_brett @mcuban They are for any tasks done by humans. Because the world was already designed for humans. It makes zero sense to create 1,000 specialized robots for different tasks that a human can do now.
English
@ericzakariasson That is nothing.
Running an interpreted code editor in Chromium in a CLI-first agentic AI era is insane.
I know you'll make the move at some point.
But it surprises me that it hasn't happened yet.
This could also fix your `com(.)todesktop(.)xyz` ID and Hilary issue.
English
this could have been real
cursor.com/glass
eric zakariasson@ericzakariasson
ok, hear me out on this one..
English
@eliebakouch I don't believe that it is this what he is saying.
He is using ambiguous wording to say the following:
25% of THEIR compute was spent on fine-tuning the base, and 75% on RL.
English
so 3x the training compute gets you 1% improvement on swe bench multilingual and 21% on terminal bench 2.0 but k2.5 is in non thinking mode? if those benchmarks are useless, it's weird that they are the ones reported in cursor blog then?
something is wrong
Lee Robinson@leerob
Yep, Composer 2 started from an open-source base! We will do full pretraining in the future. Only ~1/4 of the compute spent on the final model came from the base, the rest is from our training. This is why evals are very different. And yes, we are following the license through our inference partner terms.
English
@sceptical_panda @teortaxesTex You are a freaking top-moron. You never looked at hardware company financials.
You may be even more of a pseudo-intellectual moron than @teortaxesTex .
If you'd surpass a 130 IQ threshold (which is not high), you'd realize that most of his tweets are slop.
English
@NumCupboard @teortaxesTex I don’t believe you had any education. Not possible to be this dumb if you were educated
English
