Jeremy Howard: Nobody at Google DeepMind Realized How Good Gemini Neon Was

AccBalanced · x · 2026-11-22

Jeremy Howard highlights that seemingly no one at Google DeepMind realized how good Gemini Neon was — at launch Google repeatedly described it as merely "approaching" frontier models. Perplexity CEO Arav Srinivas praised Neon (reasoning at High) as rare in being both intelligent and fast, and migrated some teammates to it. Howard's takeaway: evals are a very imperfect measure of model utility.

Original post →

More from Models

Models channel →