I've been collecting what top researchers say about AI benchmarks and models — from Twitter/X, YouTube interviews, and podcasts. Here are the most interesting takes from the past week.
From the Lex Fridman interview (March 23, 2026):
"I think it's now. I think we've achieved AGI. It is not out of the question that a Claude was able to create a web service, some interesting little app that all of a sudden a few billion people started using."
"The next scaling law is the agentic scaling law. It's like multiplying AI — we could spin off agents as fast as you want."
From the No Priors podcast:
"I don't think I've typed like a line of code probably since December."
He uses Claude Code exclusively. When the person who built Tesla's Autopilot stops writing code and lets AI do it, that's a signal.
On Meta's Muse Spark release:
"The new model from Meta is already looking like a disappointment: overoptimized for public benchmark numbers at the detriment of everything else."
This from the creator of ARC-AGI — the one benchmark specifically designed to resist gaming.
Fuli Luo leads Xiaomi's MiMo and previously worked at DeepSeek. On Anthropic cutting off third-party harnesses (1,748 likes, 720K views):
"Anthropic cut off third-party harnesses from using Claude subscriptions — not surprising. Three days ago, MiMo launched its Token Plan — a design I spent real time on."
The next frontier isn't model quality — it's who can make inference efficient enough to deploy agents at scale.
"In different hands, Mythos would be an unprecedented cyberweapon. I am not sure how we deal with this, except to note a narrow window where we know only 3 companies could be at this level of capability."
"Anyone has access to mythos and can let the rest of us plebs know what it feels like"
1,667 likes. Even the CTO of HuggingFace is waiting in line.
All quotes with sources: [benchmark.space/voices](https://benchmark.space/voices)
Full timestamped transcripts: [benchmark.space/transcripts](https://benchmark.space/transcripts)
The benchmark data behind these takes: [benchmark.space](https://benchmark.space)
48 researcher quotes, 15 full transcripts, updated every 2-4 hours. All from [benchmark.space](https://benchmark.space).