Yann LeCun's new startup, Advanced Machine Intelligence Labs, represents a potential blind spot in the AI industry's push to slow development. As major tech companies like OpenAI, Anthropic, and Google DeepMind publicly commit to slowing artificial intelligence (AI) progress, smaller startups founded by veteran researchers are quietly exploring alternative approaches that could circumvent these safety agreements entirely.

What Is the "Pace the Frontier" Movement?

The AI industry's largest players have recently rallied around a concept called "pacing the frontier," an effort to slow the development of cutting-edge AI capabilities while building stronger safety guardrails. Anthropic CEO Dario Amodei proposed the idea in an essay titled "We Must Pace the Frontier," arguing that the most advanced AI systems are progressing faster than humans can safely monitor and control.

Dario Amodei

The proposal gained surprising support from rivals including OpenAI's Sam Altman, SpaceX's Elon Musk, and Google DeepMind's Demis Hassabis. Their backing suggested the industry might finally unite around safety concerns. However, the consensus quickly fractured as executives and researchers grappled with what "slowing down" actually means in practice.

How Could Smaller Labs Outpace the Safety Slowdown?

The real challenge lies in enforcement. Ben Goertzel, founder of SingularityNET and a prominent AI researcher, highlighted a critical flaw in the slowdown strategy: the assumption that only a handful of companies can push the frontier forward may not hold true much longer.

LeCun's Advanced Machine Intelligence Labs and other startups founded by former Google DeepMind chief scientist Jeff Dean are exploring experimental architectures that could challenge the industry's current training practices. These alternative approaches suggest that cutting-edge AI development might not require the massive computing resources that only tech giants can afford.

"I think the assumption that only a few companies can be at the frontier may not look so true a year from now. The smaller the cutting-edge things get, the harder it is to police," said Ben Goertzel.

Ben Goertzel, Founder of SingularityNET

This observation raises an uncomfortable question for the industry: if AI breakthroughs can happen at smaller, nimbler organizations, how can a coordinated slowdown actually work? The major labs may agree to pace themselves, but they cannot control what happens outside their walls.

Why Are the Biggest Players Disagreeing on How to Slow Down?

Even among the companies publicly supporting the slowdown, disagreement is already emerging. Meta CEO Mark Zuckerberg took a notably different stance from OpenAI and Anthropic, arguing that each lab should move at its own pace and take responsibility for its own safety measures.

The core disagreement centers on whether AI safety is a collective problem requiring industry-wide coordination or an individual responsibility. This philosophical divide has practical consequences. If only some labs agree to slow development, others could seize the opportunity to catch up or establish themselves as the new frontier, creating a competitive incentive to break ranks.

Beyond philosophical differences, the industry lacks shared definitions of key concepts. A.J. Bhadelia, who oversees Americas public policy at Cohere, noted that fundamental scientific disagreement persists about what constitutes AI risk and how severe different risks actually are.

"There remains fundamental scientific disagreement in that space. Some of the recent discussions about AI risk have approached the territory of science fiction," said A.J. Bhadelia.

A.J. Bhadelia, Americas Public Policy Lead at Cohere

What Practical Obstacles Stand in the Way of a Real Slowdown?

Even if the major labs agreed on a slowdown, several loopholes could undermine it. Some researchers have proposed limiting the computing power used to train AI models, but today's frontier progress increasingly depends on test-time workloads and software development surrounding AI rather than just raw model training.

Ritwik Gupta, a computer science and AI professor at the University of Maryland, explained the challenge: "I'm not fully convinced that pausing training alone is actually going to meaningfully change the rate of progress of the frontier".

"I'm not fully convinced that pausing training alone is actually going to meaningfully change the rate of progress of the frontier," explained Ritwik Gupta.

Ritwik Gupta, Computer Science and AI Professor at the University of Maryland

The definition problem also creates enforcement challenges. Where exactly does AI development end and general software development begin? This ambiguity makes it nearly impossible to police compliance across the industry.

Steps to Understanding the Slowdown Debate

• Understand the Core Proposal: Anthropic CEO Dario Amodei proposed slowing AI development while dedicating more resources to model alignment, which ensures AI systems behave according to human values and intentions.
• Recognize the Competitive Tension: Even companies that publicly support a slowdown face pressure to move faster than competitors, creating an incentive to break agreements if others do the same.
• Identify the Enforcement Gap: Without clear definitions, shared metrics, and independent oversight, there is no reliable way to verify that labs are actually slowing down or to prevent startups from racing ahead.
• Consider the Startup Wildcard: Smaller companies founded by veteran AI researchers like Yann LeCun may be able to achieve frontier-level breakthroughs without the massive resources of tech giants, making industry-wide coordination nearly impossible.

The slowdown movement also hinges on establishing independent evaluators within AI labs to monitor safety practices. Amodei proposed placing "embedded evaluators" with employee-level access inside frontier labs, but this raises new questions: Who should do the auditing? Who will govern the auditors? What types of evaluations should be administered?

Meta's Zuckerberg suggested a laissez-faire approach, noting that Meta Superintelligence Labs already engages independent evaluators and "other labs can just do this too." However, this decentralized model lacks the consistency and accountability that a coordinated system might provide.

The AI Evaluator Forum, a group of independent evaluation organizations, published a public letter on Friday with over 100 signatories outlining a framework for embedded evaluators. The letter called for diversity in evaluations and operational safeguards to help evaluators maintain independence.

As the industry grapples with these challenges, one thing becomes clear: the gap between publicly supporting a slowdown and actually implementing one is far wider than most executives anticipated. With startups like LeCun's exploring alternative architectures and fundamental disagreements persisting about what constitutes AI risk, the "pace the frontier" movement faces significant obstacles to success.