Descriptions:
Anastasios Angelopoulos, co-founder and CEO of LMArena (the company formerly known as LMSYS Chatbot Arena), joins Latent Space to discuss the platform’s evolution from a Berkeley research project into a $100M-funded company with a reported $1.7 billion valuation — and to address the “Leaderboard Illusion” paper from Cohere researchers that accused Arena of running undisclosed pre-release model evaluations.
Angelopoulos traces Arena’s origin: incubated by a16z’s Anand Rajaraman, who identified the Berkeley team, provided grants, and formed an entity before the founders had committed to starting a company. The pivot from academic project to for-profit was driven by the inference costs of running free, model-neutral head-to-head comparisons at scale — the platform funds all inference at enterprise discount rates and now operates out of a San Francisco office. The rebranding from LMArena to simply Arena reflects ambitions to expand beyond language models to other modalities.
On the Leaderboard Illusion controversy, Angelopoulos walks through Arena’s published rebuttal point by point: the paper alleged undisclosed private pre-release testing created unfair advantages, and cited Meta testing a large number of models before release. Arena’s response identifies factual errors in the paper — including a mischaracterization of open vs. closed source model representation (the actual split was closer to 60/40 open-favorable, not the skew the paper claimed) — and argues the remaining criticisms were corrected in subsequent revisions. The episode highlights growing tensions as AI benchmark politics increasingly involve commercial interests.
📺 Source: Latent Space · Published December 31, 2025
🏷️ Format: Interview
