Two commentators use “overhang” differently to question how much attention the debate should give to future frontier models.
What they said
Wharton professor Ethan Mollick argues the real problem isn’t a future AI model that hasn’t been built yet. It’s the AI already sitting unused. “The capability overhang, the gap between what these models can do and what almost anyone is doing with them, is an opportunity because most people don’t bring their own advantages to AI, and those who do get much more out of it,” he writes. His evidence: “the current capabilities of existing models are barely being used, and are often not even well understood.”
Independent analyst Ben Thompson uses the same word differently. He’s responding to Anthropic CEO Dario Amodei’s public position that AI labs should deliberately slow down, a stance Amodei calls the need to “Pace the Frontier,” using “frontier” to mean the most advanced models a lab is currently capable of building. Thompson lists five overhangs he says actually drive that call, including a capability overhang (models are already good enough that customers don’t need the newest ones) and a safety overhang. On safety specifically, he argues pacing works against its own goal: “pacing the frontier actually increases the window in which bad things can happen.” He is referring to attackers using capable models while defenders still lack reliable automated protection. LLMs, large language models, are the technology behind chatbots like ChatGPT and Claude; “aligned” means trained to follow safety rules, the point being that attackers already have access to that same safety-trained technology. He goes further on motive, writing that Anthropic’s safety argument “just so happen[s] to align with their need for more time to build a moat,” using a term for a competitive advantage rivals can’t easily cross.
Who they are
Mollick teaches at the Wharton School and writes the newsletter One Useful Thing, where he publicly tests and writes about AI tools for a general audience. Thompson founded and has written Stratechery, an independent, subscriber-funded technology analysis publication, full-time since 2014. Both write from outside the frontier labs. Their essays address different decisions: how people use current AI, and how companies and governments should handle future development.
Where they agree, and where they don’t
Both writers argue that existing capability deserves attention alongside the debate over future models. Mollick’s target is builders and everyday users who assume they need a smarter model before AI becomes useful to them. Thompson’s target is a specific policy stance from one company’s CEO. Nothing in either essay suggests they were written in response to the other. They share a practical concern: slowing future development would not remove the capability already deployed.
They diverge on tone and certainty about the underlying safety debate itself. Mollick treats it as real and unresolved, writing plainly, “I don’t think there are bright lines we can point to and say AI will never cross them.” Thompson criticizes the proposal’s competitive effects but explicitly says he believes it is motivated by safety concerns. He argues that genuine safety concerns can coincide with a lab’s need for more time to build a competitive advantage.
What Mollick gets right, and where it’s incomplete
Mollick’s core claim, that most people and companies are nowhere near the ceiling of what current AI models can do, is a claim about adoption, supported in his essay by examples of projects he built with current models. His four traits for working well with AI, deep expertise in a field, wide knowledge across fields, taste in judging outputs, and willingness to experiment, describe a real skill gap, not just a tooling gap. What the essay doesn’t fully address is why that gap persists. If the opportunity is this available, the barrier is likely organizational (time, training, incentive structures) as much as individual, and Mollick’s framing puts most of the responsibility on individual initiative.
What Thompson gets right, and where it’s incomplete
Thompson’s cybersecurity point is concrete and testable: if attackers already have access to capable models, a voluntary slowdown by one lab doesn’t remove that capability from the world. It just changes who has the newest version of it. That’s a real gap in the “pace the frontier” argument as stated. His case also depends on future models improving defense faster than they improve offense. That remains a forecast. He acknowledges safety as a motive, so the commercial overlap should not be treated as proof of bad faith.
What it means for builders
The practical takeaway doesn’t require picking a side in the safety debate. If Mollick is right that most available capability sits unused, the fastest way to get value right now is auditing what your current tools can already do that you haven’t tried, not waiting on a roadmap for the next model. If Thompson is right that slower frontier labs don’t reduce what’s accessible to bad actors, that’s a reason to treat AI-assisted attacks as a present risk to defend against, not a future one to wait out. Those immediate decisions remain useful without settling the separate policy question of how quickly frontier development should proceed.
Updated October 13, 2026: Thompson explicitly accepts safety concerns as a motive for pacing, while arguing that the proposal also serves commercial interests. Both essays discuss capability already available; neither dismisses the importance of future models.
End of article