Trying to decide whether the big AI labs should slow down or keep sprinting? Here’s the fast way to think it through. I just watched a breakdown from this savvy creator (Matthew Berman) on Dario Amodei’s essay “We Must Pace the Frontier,” and he does something I really appreciate: he reads the whole thing line by line and flags exactly where he agrees, where he hesitates, and where he flat-out disagrees.
Quick context first. An Anthropic researcher resigned, posted that neither OpenAI nor Anthropic is acting responsibly, and racked up 170 million views. Dario answered with the essay. Within hours, Sam Altman and Elon Musk both said they agree. A few months ago these guys were feuding. Now they’re nodding along together, which is wild.
The criteria the creator uses
He judges every proposal on three questions:
- Does it actually reduce the risk of losing control of AI (recursive self-improvement is his top worry)?
- Does it keep open source alive?
- Does it avoid concentrating power in a handful of companies?
The comparison: Dario’s three-step plan
- Embedded evaluators. Third-party safety teams get employee-level access inside each lab. Slack, unreleased models, everything.
✅ Pros: Anthropic committed unilaterally, OpenAI and xAI followed. The creator is “100% bought in.”
❌ Cons: Only works if evaluators are truly independent and don’t become a hurdle for small startups. - Democratic coordination. Frontier labs set shared safety standards, backed by regulation and an antitrust waiver so competitors can talk.
Pros: Standards beat chaos.
Cons: This is where the creator gets uneasy. He calls it “subtly regulatory capture.” A ten-person startup can’t afford capability certifications the way Anthropic can. And “we want to be regulated, just don’t call us a cartel” rubs him the wrong way. - Global coordination with China. Four levels, from banning bioweapon use up to a full pause.
Pros: Levels one and two seem feasible.
Cons: The US leads, so China has little reason to agree to slow at the same rate. China’s state press already called the essay a “cold war tactic.”
The risk the essay skips
Here’s the title of the video, and the creator’s core point: Dario never mentions concentration of power. If only a few licensed companies can build frontier intelligence, they get enormous leverage over the economy and governments. Satya Nadella made the same argument, saying the ecosystem needs both closed and open models and can’t be “controlled by a handful of entities.”
The recommendation
The creator’s stance is basically Altman’s middle path: pace, don’t pause. Take the embedded evaluators. Harden your systems (OpenAI already paused an RL run after the Hugging Face incident to do this). Skip the antitrust exemption. Leave open source alone. As Musk put it, “software is math,” you can ban it but you can’t stop it.
How to apply this to your own reading
- When a CEO asks for regulation, check who it slows down. Rahm Emanuel claimed no industry ever asks to be regulated. He got community-noted with railroads, crypto, and drones as examples.
- Separate “safety” from “licensing.” Evaluators are safety. Capability checkpoints for every builder are licensing.
- Watch the IPO angle. David Sacks argues pacing is just good business after a public cyber incident. Liability exposure, not altruism.
I came away nervous about recursive self-improvement but way clearer on which fixes actually help. The full video walks through LeCun, Gurley, Sacks, and Bernie’s 20-year-prison proposal too. Worth the watch if you want the whole map!