Lichang Chen

Coding Agent RL & Harness of 🥑 @Meta MSL. | Previously: GenAI unit @GoogleDeepmind | PhD’25 @umdcs BS’20 @zju_china

[Lichang Chen @LichangChen2]

Try Contemplating Mode in https://t.co/awjrUo1ITW! It uses parallel thinking to deliver more well-rounded answers. We’d love your feedback!

I feel very fortunate to have contributed to this project and to keep learning through it. It has been both fun and rewarding to grow alongside one of the best reasoning teams on the planet.

[Lichang Chen @LichangChen2]

Unfortunately, papers & experiences are just the tickets to the interviews😅😅 but in frontier labs, what matters most is the solid engineering (for 90% of the researcher) at this stage where RL scaling comes to the environment/data/harness scaling, and over 99% of the fancy RL algorithms just don’t work.

[Kyle Wong @ewveggies]

6 yr ML PhD, trained Olmo 3, trained Nemotron 3, but still forced to grind Leetcode and Neetcode 75.

Despite all the headlines saying otherwise, Leetcode is clearly not dead.

Somehow knowing dynamic programming is more important than knowing data parallelism for interviews.

[Lichang Chen @LichangChen2]

@chenxi116 @grok It’s like Kobe joined Celtics to me😆

[Lichang Chen @LichangChen2]

@_arohan_ Anthropic also got John Jumper

[Lichang Chen @LichangChen2]

@edchi I think we are on the AGI: Model’s IQ is above the level of general humans.😆 Curious about your definition of AGI.

[Lichang Chen @LichangChen2]

@CoreAutoAI But Bluff should be banned here!

[Lichang Chen @LichangChen2]

Tatsu and Percy are amazingly great at the education. I feel lucky that I became their online “student” last year and learned a lot on the pretraining, which as a post-train researcher hardly had a chance to touch.

[Roan @RohOnChain]

Anthropic pays $750,000+ a year for engineers who can build LLM architectures from scratch. Stanford taught the entire thing in 1 hour lecture & released it for free.

Bookmark & watch this today before someone takes it down.

[Lichang Chen @LichangChen2]

@WeizhuChen congrats!🎉

[Lichang Chen @LichangChen2]

@willccbb I was thinking prime intellect’s equity is more promising/valuable than Anthropic’s.😆

[Lichang Chen @LichangChen2]

interesting blog.

[koko 𝑥𝑠 @kokoxs]

https://t.co/dlaW0RSd0T

[Lichang Chen @LichangChen2]

@_arohan_ @DougDoesOps @DougDoesOps can I get an invite?

[Lichang Chen @LichangChen2]

@Alan_Lau @nomad421 @NSF Congrats, Alan

[Lichang Chen @LichangChen2]

@kevin_x_li Are 15 turns really enough for solving complex SWE tasks? I was thinking my codex always call over 30 turns for my tasks

[Lichang Chen @LichangChen2]

🧐🧐

[Konstantin Dobler @konstantdobler]

@Hesamation Better version without arbitrary institution cutoff, some data cleaning and splitting contribution of each paper among institutions. China + USA dominant ofc, but looks a bit different, doesn't it?

[Lichang Chen @LichangChen2]

Stuart might be right

[Muskonomy @muskonomy]

NEWS: UC Berkeley computer scientist Stuart Russell takes the stand on day 5 of the Elon Musk vs Sam Altman trial, with direct examination by Elon Musk’s lawyer Steven Molo.

Russell told the court the current AI race has a “winner take all” dynamic. He said “whichever company develops AGI first (AI that matches or exceeds human capabilities in every area) will have a significant advantage.”

Because AGI could replace most human work, Russell warned the companies that build it will “quickly come to control perhaps the majority of economic activity on the planet.” He added that even governments could “become subordinate to these companies.”

[Lichang Chen @LichangChen2]

@LijieyYang Congrats!

[Lichang Chen @LichangChen2]

@arena @OpenAI Even worse than Muse Spark? 🥲🥲

[Lichang Chen @LichangChen2]

I like the idea but I would rephrase it as: Coding agent will be the new foundation for the whole tech industry. They can become any roles by writing the code and docs, e.g, if it works on a research project/folder, it becomes automated researcher as long as you give it an excellent harness and make it iterate to improve the code and docs.

[Jacob Effron @jacobeffron]

.@swyx current thesis:

"2025 was coding agents. 2026 is coding agents breaking containment to do everything else."