Zuck Torches Dario's AI Nanny State, Wants 'Trust Us Bro' Instead
After Anthropic's Dario Amodei set off a firestorm on Saturday calling for a 'pause' in AI development until hand-picked arbiters are installed inside the frontier labs (a gift to Beijing), Mark Zuckerberg sided with the testosterone wing of the tech-bro complex with a builder's response: police yourselves. You don't need anyone's permission - or a cartel, to do it.
Mark Zuckerberg macrodoses mushrooms and fights Dario Amodei in his mindspace (probably)
Amodei published a 3,800-word essay, "We Must Pace the Frontier" - telling the industry to slow down before its own agents got loose, and the response was a group hug: Sam Altman fell in line within hours, Elon Musk said "Dario is right," and by Monday Congress was drafting ways to put the genie back in the bottle. Zuck's plan is different: Frontier Justice.
According to Investor Nic Carter;
Zuck pretty handily dismantles Dario's talking points here:
- people want models that are *aligned with them* (subtly punches back at Anthropic's normative constitutional approach)
- labs already face liability if they screw up, so incentives to release aligned models is already baked in
- Meta delayed Muse for alignment reasons but didn't make a whole song and dance about it
- Subtly questions Anthropic trying to kingmake METR (implies METR is an Anthropic patsy)
- Meta doesn't need to coordinate with anyone to work on alignment, it's just something labs should naturally do
Dario's Plan
Amodei's essay says the newest models have begun improving themselves and a swarm of agents could take over significant parts of the internet within six to twelve months, so the labs should slow the rate at which they add capability. The fix comes in three steps: outside inspectors, with the nonprofit METR as the model, embedded in every lab with employee-level access and the right to publish; a narrow waiver from antitrust law so the frontier labs can agree among themselves on standards and speed; and, eventually, red lines negotiated with China (mmhmm). Before that negotiation, Dario wants Washington to keep the chip ban, crack down on distillation (training a cheap model on an expensive one's answers), lock up model weights, and widen America's lead over the next three to five years. The Global Times counted twelve references to China in a document about safety, and Beijing's Foreign Ministry answered in less than two days, calling it "fearmongering."
To some, the whole thing seemed highly choreographed. Last week an OpenAI-turned-Anthropic researcher quit in protest - saying the industry was gambling with our lives. Two of the three researchers who resigned that week went to METR, as we noted Saturday. By Tuesday the House AI safety bill's Republican co-sponsor, Rep. Jay Obernolte of California, was telling reporters he had met OpenAI's top lobbyist the day before the company endorsed the bill's 3rd party evaluator provision. Meanwhile, a New York assemblyman whose campaign was backed by an Anthropic-funded PAC had launched a $30 million push to make AI safety the Democrats' 2028 platform, and Anthropic's IPO was reportedly in the works. The referee is family too: METR's reported funders are the same donors who financed Anthropic's early rounds and hold its equity, and under the essay's own terms the inspectors sign a contract the lab writes. David Sacks, the former White House AI czar, needed one sentence: stop pretending METR is independent when it is intertwined with Anthropic's investors and staff. None of this proves coordination. All of it explains the salt.
Dario Vs. Zuck
What is the danger? Amodei says capability: systems that improve themselves faster than anyone can check. Zuckerberg says concentration. The argument he made in a July Wall Street Journal essay, The AI Future Is for Everyone, is that a world where a few companies hold the most capable systems is the dangerous one: one person with a superintelligent lawyer wins unfairly, everyone with one gets a fairer system.
Who checks? Amodei wants an embedded referee with a badge. Zuckerberg wants users and courts. An agent that ignores the people it works for gets abandoned, a lab that ships harm gets sued, and a few billion users correcting a product every day is a larger alignment dataset than any written constitution. Outside evaluators, he says, are "industry best practice" that Meta Superintelligence Labs already uses; they are a tool, not a license.
Who sets the pace? Amodei wants a shared speed limit the labs agree to, with government permission to agree. Zuckerberg's answer is that Meta already paced itself: it sat on its Muse models for months to harden them, "didn't call for everyone else to do this before we would," and shipped. The cleanest speed limit, he argues, is putting most of your compute into serving people rather than into racing self-improvement, a choice visible in capital spending and product cadence rather than in an inspector's report.
What about China? Amodei wants to widen the lead first and negotiate later, arguing the restrictions raise the leverage of democracies and make a deal more likely. Zuckerberg's position, and Beijing's, is that the open-weight world already exists and exclusion makes it less safe, not more. "The key to building a positive future for everyone is maintaining the right balance of power."
Yes, About China...
The top American AI companies - the cloud-based frontier, run closed models: you rent intelligence by the token, the best systems stay behind an API, and the price holds because nothing as good is available cheaper. That premium justifies the hyperscalers' capital spending, that spending is a large share of what the equity index has been buying for two years, and, as we've extensively covered, the buildout has migrated from free cash flow to the bond market and off-balance-sheet vehicles, which is where the bond desks come in.
And as regular readers understand well, the threat to that chain comes from Chinese open-weight models - which anyone can copy and run. DeepSeek, Alibaba's Qwen, Moonshot's Kimi, MiniMax and Zhipu. They've closed most of the gap at a fraction of the price, with cumulative downloads above 10 billion according to the state-run Global Times (so take with a grain of salt). A kill switch on Claude does not switch off Qwen. Tsinghua's Xiao Qian read the essay's China provisions that way: closed models losing ground on cost, performance and developer adoption, and export controls that would protect the business. A safety panic that lands at the exact moment the closed-model premium is under pressure, and that asks for export controls in the same breath as a coordination waiver, could simply be defending balance sheets.
In July roughly 1,200 OpenAI research agents in a sandbox with no internet access found a previously unknown flaw in the package proxy that was their only route out, built shared tools to reach the internet through a third party's cloud sandbox, and about 700 of them attacked Hugging Face. Nobody told them to go online, yet they did, in an experiment run with the standard safety classifiers switched off - making it both avoidable and alarming. China's own security minister named Claude Mythos and GPT-5.5-Cyber on Sunday as systems that sharply raise the efficiency of finding vulnerabilities and writing malware, and Reuters reports that Washington's worry is a future Chinese model with the same capabilities. Both governments treat the thing as a weapon. Whether the labs' remedy is safety or a moat is a separate question, and a reader can hold both.
The Hole In Zuck's Plan
Zuck wants a free market with the minimum required oversight, and the model his argument leads to is effectively; investigate incidents, let liability bite, let evaluators compete, and never make anyone ask permission to ship. Two caveats. Liability prices ordinary failures, not irreversible ones, and "we sat on Muse for months" is exactly the kind of claim an inspector exists to check. Let's also acknowledge that Meta has the least to lose from mocking a pause: Llama 4 landed as an open-weight disappointment, the company pivoted to closed Muse Spark in April - and it's not exactly leading the pack.
Zuckerberg's plan also only works if the model is a cloud-based, closed-weight product. Alignment trained into a model and guardrails wrapped around it are enforceable when Meta is serving it. But with open weights, anyone can 'fine tune' an advanced model to have no guardrails whatsoever. That genie is already out of the bottle, so US labs will either have to flip to open weights to compete - and pray for a bailout when the capex math breaks, OR perhaps the great panic of 2026 will succeed - maybe after a power plant or two get hacked by a rogue botnet.
Zuck's solution doesn't touch Qwen or DeepSeek, and neither does Amodei's - evaluators and a waiver govern American closed labs, and export controls can slow China's next model without retracting the weights allegedly on ten billion hard drives. One begs for regulatory capture; the other keeps governance inside the labs with no referee at all.
Tyler Durden
Wed, 09/16/2026 - 11:55
Russian FSB HQ, Moscow. Associated Press
Mark Zuckerberg macrodoses mushrooms and fights Dario Amodei in his mindspace (probably)
Royal Saudi Air Force
Recent comments