Developers just caught a company hiding something inside of their new model.

@casey.aicreates

16.8K followers · usually 2.3K views a post

Views

22.2K

Likes

455

Comments

12

Shares

46

Selling · SaaS › AI Tool

Cursorseen in 62 videos →

Why it worked

The video leverages a recent controversy and a "gotcha" moment in the AI community, where a company is accused of misrepresenting its product's origins. This taps into public interest in AI ethics and transparency, making the content highly shareable.

Summary

Developers discovered that Cursor's new AI model, Composer 2, was built upon an open-source model from Moonshot AI called Kimi K 2.5, contradicting Cursor's initial claims of developing a proprietary model. This revelation sparked controversy and a debate about AI development practices.

Structure

1Cursor releases Composer 2, claiming advanced AI capabilities.
2Developers discover the model ID references Kimi K 2.5, an open-source model.
3Evidence suggests Composer 2 is based on Kimi K 2.5 with additional training.
4Cursor admits to using Kimi's open-source base, but claims significant proprietary training.
5The situation ignites a debate about AI model development and transparency.

Product placement

Cursor is the main subject, with its new AI model 'Composer 2' being analyzed. The video also shows screenshots of benchmark charts comparing Composer 2 to models from Anthropic and OpenAI, and a tweet from 'Fynn' detailing the discovery, as well as a tweet from 'Kimi.ai' congratulating Cursor.

Original caption

Developers found a hidden model in Cursors new Composer 2 model (Cursor states the use of Kimi was consistent with the terms of its license) #vibecoding #aiagent #cursor

More from @casey.aicreates

Their other posts in the index, biggest breakout first.

Call to action

Join the Discord in the link in my bio if you want to keep up with this AI space.

On-screen text

DEVELOPERS JUST CAUGHT A COMPANY HIDING SOMETHING INSIDE OF THEIR NEW MODEL. RELEASED A NEW CODING A.I. MODEL CALLED COMPOSER IT HAD FRONTIER LEVEL CODING INTELLIGENCE. BENCHMARK CHARTS COMPARING IT TO MODELS FROM COMPANIES LIKE ANTHROPIC AND OPEN A.I. THEY THEN USED THOSE CHARTS AND NUMBERS TO JUSTIFY RAISING MORE MONEY FOR THE COMPANY. BUT THEN DEVELOPERS STARTED DIGGING AROUND IN THE API. FOUND SOMETHING VERY THE INTERNAL MODEL ID LITERALLY REFERENCE KIMI K 2.5, OPEN SOURCE A MODEL BUILT BY THE CHINESE STARTUP MOONSHOT A.I. THE NEW MODEL MIGHT ACTUALLY BE KIMI K2.5 WITH REINFORCEMENT LEARNING ADDED ON TOP. CRAZY PART. THE DISCOVERY HAPPENED A DAY AFTER LAUNCH. DEVELOPERS NOTICED THE RAN TOKENIZER TEST, AND STARTED POSTING THE EVIDENCE ONLINE. THE CONTROVERSY EXPLODED THAT CURSOR EVENTUALLY ADMITTED THAT BUILT ON TOP OF KIMI'S OPEN SOURCE BASE. ONLY ABOUT A COMPUTE ACTUALLY CAME FROM THE ORIGINAL AND THE REST OWN TRAINING. BUT THE WHOLE SITUATION SPARKED A THE A.I. ARE COMPANIES ACTUALLY BUILDING NEW A.I. MODELS, JUST TAKING OPEN SOURCE MODELS AND PUTTING BETTER TRAINING GIVE ME A THE DISCORD IN MY BIO KEEP UP WITH SPACE

Transcript

Developers just caught a company hiding something inside of their new model. The company Cursor released a new coding AI model called Composer 2. It had frontier level coding intelligence. Benchmark charts comparing it to models from companies like Anthropic and OpenAI. They then used those charts and numbers to justify raising more money for the company. But then developers started digging around in the API. Found something very interesting. The internal model ID literally reference Kimi K 2.5, which is an open source AI model built by the Chinese startup Moonshot AI. The new model might actually be Kimi K 2.5 with reinforcement learning added on top. But the crazy part. The discovery happened a day after launch. Developers noticed the model ID ran tokenizer test and started posting the evidence online. The controversy exploded so fast that Cursor eventually admitted that the model was built on top of Kimi's open source base. They said, "Yep, Composer 2 started from an open-source base!" But he said, "Only 1/4 of the compute spent on the final model came from the base, the rest is from our training." As a result, he said Composer 2's performance on various benchmarks is "very different" from Kimi's. Robinson also insisted that Cursor's use of Kimi's was consistent with the terms of its license, a point the Kimi account on X repeated in a subsequent post congratulating Cursor, where it said Cursor used Kimi "as part of an authorized commercial partnership" with Fireworks AI. Only about a compute actually came from the original and the rest from their own training. But the whole situation sparked a bigger debate in the AI world. Are companies actually building new AI models, or are they just taking open source models and putting better training give me a follow. Join the Discord in the link in my bio if you want to keep up with this AI space.