Constrain frontier capability scaling by confidence in safety and coordinate slowdowns when needed, while continuing alignment and defensive research. [1][2]
Initial statement. Constrain frontier capability scaling by confidence in safety and coordinate slowdowns when needed, while continuing alignment and defensive research.
Source [1] · Authored statement
An alien mind
Attributed to: Jakub Pachocki · OpenAI
Published
“Scaling AI systems has to be constrained by our confidence in safety.”
Pacing RSI, opening sentence of paragraph beginning Scaling AI systems · View original source ↗
Source context. Authored by Jakub Pachocki with September6 date. Explicitly argues against unchecked maximum-speed collective race; alignment/defense continue. Description of rapid advances and possible defensive benefits is context, not an acceleration endorsement.
Scope
Frontier capabilities and recursive self-improvement, including increasingly capable model training.
Conditions
Safety confidence, alignment/monitoring, third-party evaluation, coordinated common safety standards and voluntary slowdowns until shared bars exist.
Stated rationale
Authored by Jakub Pachocki with September6 date. Explicitly argues against unchecked maximum-speed collective race; alignment/defense continue. Description of rapid advances and possible defensive benefits is context, not an acceleration endorsement.
Source details
primary · Original publisher or author; verified independently of scout output.
Discovered 3 Oct 2026.
Published 6 Sep 2026.
Accessed 3 Oct 2026.
Source [2] · Supporting original statement
What if automating AI R&D triggers an intelligence explosion?
Attributed to: Jakub Pachocki · arXiv / authors
Published
“First, policymakers should develop ways to pace and constrain scale-ups of automated AI R&D.”
PDF page7, Steering and constraining an intelligence explosion, first numbered policy component · View original source ↗
Source context. Name appears on first-page author roster; paper explicitly says views are authors own, not employers. arXivv1 datesSeptember28. This is coauthored support for tools and managed capability growth, not a standalone demand for an immediate all-lab halt. Benefits and safety R&D still valued.
Scope
Scale-ups of automated AI R&D and the risk of software-driven intelligence explosion.
Conditions
Governments should consider safety/input conditions, capability growth limits per period, verification of agreements, incident responses and targeted workload pauses; weigh abuse and costs of delayed progress.
Stated rationale
Name appears on first-page author roster; paper explicitly says views are authors own, not employers. arXivv1 datesSeptember28. This is coauthored support for tools and managed capability growth, not a standalone demand for an immediate all-lab halt. Benefits and safety R&D still valued.
Source details
primary · Original publisher or author; verified independently of scout output.
Discovered 3 Oct 2026.
Published 28 Sep 2026.
Accessed 3 Oct 2026.
Editorial assessment. Research proposal only; independent adversarial review and explicit human approval required. Contrary/scope evidence: Stronger aligned systems may defend against dangerous AI and scientific progress can help safety, but essay rejects treating greatly accelerated AI research as the right collective answer; calls for managing RSI risk and coordinated slowing as needed.