Tech • AI • Robotics • Game
Pacing the Frontier, Perfect Little Angels, Nico Wittenborn IRL, Home Depot with Acquired
9.1/10TL;DR
A weekend clash over whether to slow frontier AI development exposed growing divisions among major labs, policymakers and investors over safety, competition and U.S.-China strategy.
KEY POINTS
A fast-rising call to “pace the frontier”
The latest debate builds on moves that began in late July, when employees and leaders linked to OpenAI, Anthropic, DeepMind and Meta backed a “pacing the frontier” statement. By August, OpenAI had published a policy note on slowing model development around cyber-risk and said it had paused a major reinforcement-learning run for two weeks. On September 12, Anthropic chief Dario Amodei pushed the idea into the center of public debate with a new essay calling for coordinated restraint.
Amodei’s three-part proposal
The plan has three main elements. First, independent third-party evaluators would get access to leading AI companies to verify safety practices, with METR discussed as one possible evaluator. Second, frontier labs in democratic countries would work with governments to establish shared safety standards. Third, democratic governments would try to coordinate with authoritarian states, including China, on pacing advanced AI development.
Independence and antitrust concerns
The proposal immediately drew questions about whether outside evaluators are truly independent, given overlapping staff and ties between safety groups and major labs. It also raised antitrust issues: formal coordination among competitors on slowing development could require government backing or exemptions under laws such as the Sherman Antitrust Act. Critics argued that what looks like safety coordination could also function like cartel behavior if dominant firms jointly restrict output.
Support from major AI figures
The essay quickly won public support from several top industry figures. Elon Musk endorsed it early on Saturday, followed by Sam Altman, while Demis Hassabis broadly agreed with its direction while suggesting it needed refinement. The alignment was notable because these figures often differ on pace and governance, yet converged on the idea that frontier development may require tighter oversight.
Critics say labs can slow down on their own
Opponents argued that the companies leading the frontier do not need permission to reduce their own pace. David Sacks said that if OpenAI and Anthropic believe unreleased models are dangerous, they should simply act responsibly rather than seek special regulatory structures. He also rejected the idea of suspending antitrust constraints to let a small group of labs coordinate, and questioned giving intertwined evaluators authority over rivals that are not even at the frontier.
Fear of regulatory capture
A central objection is that burdensome safety rules could lock in incumbent labs while making it harder for smaller startups to build AI products. Application-layer companies that do not train frontier models could still face delays, embedded reviewers and compliance costs if regulation spreads too broadly. That, critics warned, would turn safety policy into a barrier to entry rather than a targeted response to high-end model risk.
Trump rejects slowdown arguments
Donald Trump used social media and public remarks to dismiss calls for slowing AI. He argued that “whoever wins AI wins,” said the U.S. is ahead of China, and warned that extensive guardrails would damage a strategic advantage. He also described AI and data centers as potentially the biggest economic development engine in history, while portraying anti-build sentiment as a threat to American competitiveness.
Business incentives remain disputed
Analysts disagree on whether pacing helps or hurts frontier labs commercially. One view is that regulation would favor large incumbents by raising compliance costs for everyone else. Another is that slowing the leaders would compress their margins by allowing rivals such as Google DeepMind, xAI and others to catch up, especially as model pricing is already under pressure. Some investors noted that Anthropic and OpenAI operate as public benefit corporations, giving management more room to prioritize safety over short-term shareholder returns.
The semiconductor bottleneck shapes the debate
The discussion also widened to the physical infrastructure behind AI. Investors highlighted constraints in chips, power, rail access and transmission as real limits on scaling. Some argued that even if frontier model development is paced, the U.S. should still accelerate electricity generation, grid upgrades, nuclear deployment and other energy investments, since cheap power would support broader industrial competitiveness beyond AI alone.
Specialized AI for chip design is gaining ground
Separate industry commentary underscored why semiconductor design remains a hard target for general AI models. Executives in the sector argued that chip development still takes four to six years and can cost hundreds of millions of dollars, with high risk that market needs change before a design ships. They said progress will likely come from domain-specific models trained on proprietary semiconductor datasets rather than from broad consumer LLMs, because chip-design data is scarce, closed and highly precision-dependent.
CONCLUSION
The new fight over AI pacing is no longer a niche safety discussion. It has become a test of how the U.S. balances innovation, market power, legal constraints and strategic competition with China.
Explain this