Some AI leaders say progress should be paced, but are labs actually hitting the brakes, or is it still just talk?
Are AI companies already implementing slowdowns in development as claimed?
This explores whether AI labs are actually slowing their own development, as some leaders have proposed, or whether 'pacing' is still mostly a proposal while the work keeps getting faster.
This explores whether AI labs are actually putting the brakes on their own development, or whether 'slowing down' is still mostly talk. The short answer from the corpus is that it is still talk. The collection has proposals to slow down, criticism of those proposals, and arguments for government-enforced limits. It has no documented case of a frontier lab measurably slowing its own capability work. The clearest measured evidence points the other way.
The main proposal comes from Amodei, who argues that capability gains need to be deliberately paced so safety work can catch up. He points to recursive self-improvement (AI systems helping build better AI systems) and incidents where multiple AI agents went wrong as reasons why more safety funding alone won't be enough Should AI capabilities growth be deliberately slowed to allow safety work?. His first concrete step is modest: put evaluators inside the labs, then add third-party checks and reporting later. That is a plan for watching development, not a commitment to slow it. Karpf's critique cuts deeper. He compares the embedded evaluators to banking supervisors and points out that bank supervision works only because regulators can impose fines. Without that kind of enforcement, a pacing plan mostly benefits the company that proposed it Can industry self-regulation slow AI without government enforcement?. The Future of Life Institute reaches a similar conclusion from the other side. It argues that rising incidents show companies can't police themselves, and it calls for government-mandated limits on recursive self-improvement, checked through hardware verification Can companies alone manage the risks of AI systems?.
The surprising part is what one lab's own numbers show. Anthropic reports that the length of tasks AI can handle is doubling every four months, that Claude writes over 80% of merged code, and that speedups on fixed-goal tasks rose from 3x to 52x Is AI development already being handed to AI systems?. So the company most associated with the pacing argument is also publishing evidence that it is speeding up. Those figures come from a single company and measure AI-assisted output, not fully autonomous self-improvement. The larger claim that automating AI research could pack four or five years of progress into one year rests on premises nobody has proven yet Could automated AI research compress years of progress into months?. Some researchers treat that speedup as the goal. They argue that letting AI agents improve their own code is the way past diminishing returns on research spending Can recursive self-improvement speed up the research process itself?.
A second idea you might not expect: even a real slowdown would solve less than it sounds like. Slowing down lowers risk in complex, tightly connected systems, but it can't rule out failure Does slowing AI development actually prevent system failures?. Pace measures also govern how capabilities get built. They don't settle who has the authority to stop a deployed system that is already causing harm Can slowing AI development resolve who stops deployed systems?. So 'are they slowing down?' may be the less important question. The more pressing one could be who gets to pull the plug, and how.
The corpus has a clear gap here. It has no independent audit, public timeline, or third-party measurement showing any lab slowing its capability work. If you want to check claims of slowing down, the notes on enforcement and intervention authority are the best place to start.
Sources 8 notes
Amodei contends that recursive self-improvement and multi-agent misalignment incidents demonstrate that slowing capability gains is essential, not just funding safety work. He proposes embedded evaluators as the first step, with third-party verification and reporting roles.
Karpf argues that Anthropic's pacing proposal benefits the company proposing it and that embedded evaluators, modeled on banking supervisors, fail without state enforcement backing them—analogous to how banking oversight works only because regulators can impose fines.
The Future of Life Institute argues that escalating AI incidents demonstrate private companies cannot self-police effectively, and calls for government-mandated limits on recursive self-improvement practices until safety research is complete, backed by hardware verification technology.
Anthropic measured task length doubling every four months, Claude authoring over 80% of merged code, and speedup rising from 3x to 52x on fixed-goal tasks. However, the evidence reflects AI-assisted output and velocity at one company, not autonomous recursive self-improvement or process-level gains.
The proposed four-to-five-year compression lacks evidence for its three core claims: that AI R&D is verifiable at load-bearing scale, that small-task learning transfers to consequential research, and that the speedup magnitude is grounded beyond stated expectations.
Show all 8 sources
The paper argues that AI agents automating R&D improve product efficiency while research process efficiency stays fixed. Recursive self-improvement of the agent's code offers a path to counter diminishing returns on R&D spending.
Research shows slower pace lowers risk in complex coupled systems but does not prevent failures from occurring. When failure remains possible, governance must address intervention and harm response.
Measures designed to slow frontier development act on the conditions of capability building but do not answer who has authority to intervene in a deployed system causing harm or how that intervention should proceed. These are distinct governance problems requiring separate solutions.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- We Must Pace the Frontier
- The Law of Stop: Interruptibility, Injunctions, and the Governance of Agentic AI
- The Economics of Recursive Self-Improvement
- Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs
- AI Agents Push Humans Out of the Loop
- Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering
- PostTrainBench: Can LLM Agents Automate LLM Post-Training?
- Statement: We must pressure AI companies to immediately limit the use of recursive self improvement