Anastasios Angelopoulos

Nathan Labenz × Prakash Narayanan × Anastasios Angelopoulos × Debo Olaosebikan × Samvit Ramadurgam × Michael Freeman × Sagar Vaze

More than 50% of Arena conversations lasting at least 20 turns showed at least one misaligned behavior, while $40 post-cache “Thank you” traces drove Max to route across frontier models, cheaper models, and caching.Mistral Large 4 trained on about 3,800 GPUs using 10²⁵ FLOPs, showing efficiency under compute scarcity, but Prakash argued sovereign backing may give Mistral another chance after a missed model cycle.

Harry Stebbings × Anastasios Angelopoulos

Kimi K3’s wins over every American model on meaningful tasks challenge the distillation-only narrative and strengthen the case for enterprise AI sovereignty built on proprietary data.Anastasios expects a multi-hundred-billion or trillion-dollar American-first open-source company, but predicts two-thirds of at least 75 neo-labs will fail or be acquired; regulation, data economics and infrastructure debt remain decisive risks.

Anastasios Angelopoulos

LMArena’s moat is organic scale—roughly 250 million conversations and mid-tens of millions monthly—rather than a static benchmark catalog.Its $100 million raise funds free inference, hiring, and platform upgrades, while no-pay-to-play rankings and expansion into occupational, multimodal, and agent evaluations shape the next catalyst.