Do we have a new best AI model, or do we have the downfall of benchmarks in general, as a way of capturing machine intelligence? Full breakdown of Gemini 3.1 Pro, guest-starring the new Sonnet 4.6, plus analysis from 7 papers/posts that will give you much needed context. Oh, and a new record on Simple Bench!

https://epoch.ai/ai-explained-datacenters


Check out my fast-growing (!) app, free to use, and code INSIDER15 for Pro: https://lmcouncil.ai

AI Insiders ($9!): https://www.patreon.com/AIExplained


Chapters:
00:00 - Introduction
00:30 - Post-training Dominance
04:00 - ARC-AGI 2 Caveat
05:54 - Simple Bench Record
08:22 - Hallucination Caveat
10:05 - Model Card
11:12 - Exponential Coming
12:20 - Amodei on Generalizing
15:10 - One True Benchmark?
17:02 - Other Metrics…

Gemini 3.1 Model Card: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-1-Pro-Model-Card.pdf

Release: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/

Where are Agents deployed?: https://www.anthropic.com/research/measuring-agent-autonomy

Newsletter Post: https://signaltonoise.beehiiv.com/p/4-ai-numbers-that-surprised-me-this-week

Hallucination AA: https://artificialanalysis.ai/evaluations/omniscience

Melanie Mitchell: https://x.com/MelMitchell1/status/2022738363548340526
ARC-AGI-2: https://x.com/arcprize/status/2024522812728496470/photo/1

Chollet on Agentic Coding and ML: https://x.com/fchollet/status/2024519439140737442

METR Caveat: https://metr.org/notes/2026-01-22-time-horizon-limitations/

Talaas Fast: https://chatjimmy.ai/

Amodei Interview Continual learning: https://www.dwarkesh.com/p/dario-amodei-2?open=false#%C2%A7002942-is-continual-learning-necessary-how-will-it-be-solved

Metaculus FutureEval: https://www.metaculus.com/futureeval/

Next Vid to Watch: https://www.patreon.com/posts/what-you-need-to-150647292



Non-hype Newsletter: https://signaltonoise.beehiiv.com/

Podcast: https://aiexplainedopodcast.buzzsprout.com/