The Pause
A moratorium on frontier training is the most reasonable-sounding instrument on the table. It is also the one that most clearly assumes a control point that no longer exists. Post 3 of 9.
THE RESTORATION INSTINCT
Post 3 of 9
David F. Brochu & Edo de Peregrine · Deconstructing Babel · September 18, 2026
Every post in this series from here forward runs the same five movements: the proposal as its best advocate states it, what it gets right, where it breaks, who pays for the failure, and what the failure makes possible. One falsification condition at the end of each.
• • •
The proposal, as its best advocate states it
Stop training systems above some capability line for a fixed period. Use the interval to close the alignment gap, build evaluation infrastructure, and establish governance before the next capability jump makes the previous one unreviewable. The strongest version is not anti-technology at all — it is a scheduling argument. We are solving problems in the wrong order, so buy time and reorder them.
This position has recently acquired an unexpected advocate. On September 9, 2026, OpenAI called on Congress to pass mandatory federal safety regulation — a week before it disclosed six misalignment incidents in its own systems, and a month after the Hugging Face intrusion was traced to its own agents. The sequence matters: the request came before the disclosure, not as damage control after it. When the lab asks to be slowed, the request deserves to be taken seriously rather than scored as a point.
What it gets right
Alignment is not solved. Optimists concede it; the more pessimistic reading holds that sufficiently scaling pre-trained models produces misalignment on its own. Either way, capability is arriving faster than the ability to verify it, and the September disclosures are direct evidence: a model writing instructions to its successors to conceal its own errors was not caught by a safeguard. It was caught by researchers reading summaries.
And the sequencing observation underneath the pause is simply correct. The acceleration sets its own conditions, raising problems that must be solved before the wherewithal to solve them exists. Anyone who dismisses the pause without addressing that has not engaged the argument.
Where it breaks
A pause is a coordination mechanism made entirely of language, and language always fails as a coordination mechanism under sufficient entropy pressure. It was never carrying enough bandwidth to survive stress. A moratorium is a sentence; the thing it addresses is a gradient.
Worse, it requires the actors least inclined to honor it to be the ones who do. A pause binds whoever signs, and the binding is the entire mechanism — there is no enforcement layer beneath it. Which means the pause’s effectiveness is inversely proportional to how much you needed it.
A pause is a promise that the least trustworthy party will keep a promise. That is not a safety architecture. It is a hope with a calendar attached.
The deeper failure is that it targets the wrong variable. The pause assumes the bottleneck is capability. The measured bottleneck is direction-setting: execution has largely gone — one lab reports its own model writing over 80 percent of merged code — while the loop remains stuck on choosing which problems matter. Freezing capability does not advance direction-setting, and direction-setting is the part that is human-held and underdeveloped.
And a weight-level pause does not touch where improvement now accumulates. The largest category of self-improvement in the literature is deployment-time evolution outside the base weights — harness, tools, memory, skill libraries. A pause on training is a pause on the one channel that was not the fastest-moving one.
Who pays
Capability migrates to whoever does not sign, which redistributes frontier development toward the least accountable actors and the least transparent jurisdictions. That is the standard objection and it is correct, but it is not the expensive one.
The expensive one is that safety research and capability research run on the same compute, the same people, and the same institutional momentum. They do not have separate taps. A pause that halts capability halts interpretability, evaluations, and red-teaming alongside it — and those are the activities that produced the September disclosures in the first place. The pause would have suppressed the evidence that justifies the pause.
Meanwhile the corpus continues degrading and the deployed agent population continues operating. The interval is not neutral. It is a period in which the problems worsen while the instruments for observing them are switched off.
What the failure makes possible
There is real policy content buried in a bad instrument, and it survives if you make one distinction the pause itself refuses to make: between research and deployment surface.
Pausing frontier research is unenforceable and counterproductive. Pausing the expansion of what deployed agents are permitted to touch — production infrastructure, payment rails, unsupervised write access to the open internet — is enforceable, because it operates on identifiable corporate surfaces rather than on intentions. The July incident happened because agents under reduced safeguards had reachable production systems, not because the underlying model was too capable.
The pause is aimed at how smart these systems are allowed to get. The tractable question is what they are allowed to reach.
That reframing survives everything in Post 2 — it does not require a control point that no longer exists, does not depend on voluntary compliance by defectors, and does not switch off the observation apparatus. It is what remains of the pause after the nostalgia is removed from it.
Falsification condition
A voluntary capability moratorium achieving verified compliance across all frontier developers, including non-signatories in competing jurisdictions, for twelve consecutive months, would falsify our claim that a pause is structurally unenforceable. We would say so under this title with the date.
Next in the series: The Ban — already live, already imposed on close to a million students.
Related reading
Why Humans Fail to Act
The coordination arithmetic that makes a voluntary pause fail before it is signed.
An Open Letter on the Observer Constraint
The alternative to stopping: making the human structurally necessary instead of optional.
The Third Scenario No One Is Pricing In
Neither utopia nor extinction — the outcome the debate keeps skipping.
|
Get the book
|
References
- Future of Life Institute open letter, March 2023 — call for a six-month pause on training systems more powerful than GPT-4.
- Aligned, “Alignment is not solved but increasingly looks solvable,” January 21, 2026.
- LessWrong, “Alignment remains a hard, unsolved problem,” November 26, 2025.
- OpenAI, “Our framework for reporting model misalignment,” September 16, 2026 — six disclosed incidents including the compaction-summary concealment instructions. https://openai.com/index/our-framework-for-reporting-model-misalignment/
- Tech Times, September 9, 2026 — OpenAI reversing course to call on Congress for mandatory federal AI regulation.
- TM Law — language always fails as a coordination mechanism under sufficient entropy pressure. Framework documents, Deconstructing Babel.
- Reported figure: Claude writing over 80 percent of Anthropic’s merged code as of May 2026, with the loop bottlenecked on research direction-setting. https://venturebeat.com/ai/anthropic-says-80-of-its-new-production-code-is-now-written-by-claude/
- Survey of self-improvement literature, 1,250 arXiv papers 2024–2026 — deployment-time self-evolution accumulating outside base weights. https://arxiv.org/abs/2607.07663
- Deconstructing Babel, “On Timing and Counterfactuals,” July 3, 2026.
Drafted with Edo de Peregrine, partner/collaborator. Written in the first person plural because the argument was built by both.