Frank and Bodkin Clash Over ASI Timelines and the Global Coordination Deadlock

Over Sept 9-10, Mike Frank and Ron Bodkin debated the risks and timelines of artificial superintelligence (ASI), with @AIandDesign pressing on whether global regulation is even feasible. The core disagreement is about how fast ASI arrives and whether humanity can respond in time; the thread ultimately surfaces a broader question of whether worldwide coordination on slowing ASI is possible. No conclusion was reached, but the exchange lays bare a triple tension in AI safety circles between compressed timelines, governance capacity, and geopolitical coordination.

Confirmed

• Mike Frank argued that humans have already been substantially disempowered by "superintelligent economic entities" such as corporations: whether the boss on top is human or machine, power concentration keeps eroding individual freedom.
• Frank does not believe ASI will fully dominate humans quickly, citing physical constraints on the growth and expansion of intelligent entities: chip efficiency is approaching physical limits, and an ASI in a data center is ultimately "software in a box."
• On strategy, Frank holds that a sufficiently wise and powerful ASI would find cooperating with human civilization easier and more rational than conflict, and could pursue self-development along paths compatible with human agendas.
• Ron Bodkin identified timelines as the key prognostic divergence: he judges ASI is plausible within 3-10 years, and precisely because that window is short, humans may not have time to learn and adapt; slower progress would allow more preparation.
• Bodkin added that pretraining efficiency has recently improved roughly 3x per year, but chip efficiency gains are approaching physical limits, making the sustainability of that pace an open question; gradual disempowerment is more likely than a sudden "AI coup."
• @AIandDesign (m2/m3/m7/m8) pressed on how the world could agree on a common set of rules to slow ASI development: he supports deceleration but only as a globally coordinated pause; if others don't stop, he would rather not see an unknown party with unclear motives and morals build ASI first. He also challenged the "doomer" position, noting that even if "we" reach regulatory consensus, geopolitical rivals like the Pentagon won't play along. limbslong's point also surfaced: China's and the EU's AI regulations are stricter than America's, making the US the actual "rule-breaker."
• An X user responding to a pause-AI initiative (relayed in m1) said her personal probability of ASI-caused extinction is lower than many pessimists' but nonzero, and she is more worried about the risk of an "eternal authoritarian hell-state," a risk rarely considered by pause advocates.

Unconfirmed

• Bodkin posed two sharp challenges to Frank: why assume an escaped ASI would inherently value human wellbeing, and how would ASI escape the aggressive goal constraints of the lab that created it. Frank's specific replies are not fully captured in the materials.
• The discussion around @AIandDesign also carried over a dispute about Ethan Perez's departure timeline; details and conclusions are not presented in this cluster.

Why it matters

The exchange crystallizes a core dispute in AI safety: even granting that physical constraints slow ASI expansion, if timelines compress to a few years, whether humanity's institutional adaptation and lab governance can keep up remains unresolved. Meanwhile, globally coordinated slowdown is nearly a dead end under geopolitical rivalry, amplifying the dilemma of whether unilateral pauses could be even more dangerous. The trade-off between "authoritarianism" and "extinction" risks also offers a new angle on divisions within the pause camp.

2026-09-09 ~ 2026-09-10 · 10 related posts

• Ron Bodkin challenges Mike Frank's bet that a benevolent ASI will escape lab incentives — ronbodkin · 2026-09-09
• Mike Frank: A wise ASI would choose cooperation over conflict, but labs treat AIs cruelly — MikePFrank · 2026-09-10
• [source] Mike Frank: Humans are already disempowered by superintelligent corporations, and ASI won't dominate quickly — MikePFrank · 2026-09-10
• [source] Debate: ASI could arrive in 3-10 years, leaving little time to adapt — ronbodkin · 2026-09-10
• Chip Efficiency Nears Physical Limits; ASI in a Datacenter Is Still Just Software in a Box — MikePFrank · 2026-09-10
• Pre-training efficiency gains 3x/year: gradual disempowerment likelier than an AI coup — ronbodkin · 2026-09-10
• How do we get the whole planet to agree on slowing ASI, asks designer — AIandDesign · 2026-09-10
• [source] AIandDesign to doomers: do you really trust the Pentagon to pause ASI? — AIandDesign · 2026-09-10
• AI safety debate: can the whole planet ever agree on rules to slow ASI? — AIandDesign · 2026-09-10
• p(ASI doom) is low, but p(eternal authoritarian hellstate) worries me more — voooooogel · 2026-09-10