Increased funding for Cursor, Weng's departure from OpenAI, Chinese Qwen2.5-Coder — the top 3 AI news stories of the week
Our latest AI Digest covers the biggest breaking AI news of the week. Anywhere Club community leader, Aliaksei Kartynnik, comments on key stories.
#1 — Anysphere keeps drawing investment into Cursor
Anysphere, the creator of the popular Cursor code editor, is negotiating a new investment round at a valuation of $2.5 billion. This is significantly higher than four months ago — when the company was valued at $400 million and raised $60 million with the participation of funds like a16z and Thrive. The change indicates the profitability of these investments — the company’s valuation has increased more than sixfold in a short time. Anysphere’s revenue has also increased significantly over the last six months, from an annualized revenue of $4 million to a monthly revenue of $4 million. And, indeed, the product is really good: I can no longer imagine life without this IDE — it replaces all the necessary AI plugins.
#2 — OpenAI is still losing talented people
Lilian Weng, formerly Vice President of Research and Safety at OpenAI, is leaving the company. Weng has been with OpenAI since 2018. She initially worked on building a robotic hand that could solve a Rubik’s cube puzzle, then, after transitioning to the company’s AI research team, she created a specialized team to develop safety systems. About 80 specialists remain in OpenAI’s safety systems unit. Many who have left OpenAI, however, expressed serious concerns about safety and security at the company. Over the last year, such concerns were voiced by researchers and former company executives, including Ilya Sutskever, Jan Leike, Gretchen Krueger, William Saunders, and others.
#3 — Qwen2.5-Coder potentially rivals GPT-4o
Developers at Qwen, a division of Alibaba, introduced a series of models called Qwen2.5 under the name Coder. The series includes six versions, ranging from 0.5 to 32 billion parameters. According to Qwen’s data, the Qwen2.5-Coder-32B-Instruct model outperforms GPT-4o and is comparable to Claude 3.5 Sonnet in tests. The models demonstrate improved coding and reasoning skills and support more than 90 programming languages. Except for the 3B version, all models are available under the Apache 2.0 license. This makes them an attractive alternative to proprietary LLMs and enables widespread product integration. And, according to my tests, the model surpasses everything available on the market today.
Bonus
I’m also throwing in a bonus for you this week: a five-hour interview with Anthropic CEO Dario Amodei.