Can governments quickly and cheaply slow AI training?

2026-03-07 · Source: AI Alignment Forum · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Advanced, short

Summary

This analysis explores the effectiveness of various inference-verification methods designed to slow AI training, particularly in scenarios where powerful AI development might need to be decelerated for safety or oversight. The author, having investigated the question for a week, concludes that current prototypes of inference-verification are likely ineffective. Standard measures, which restrict communication between servers, may not significantly impede training, especially for reinforcement learning (RL) models that require minimal communication. Developers could allocate 95% of compute to RL rollouts under verification and use 5% covertly for training updates. However, more aggressive measures, such as proof of work/memory accounting for over 95% of computation, frequent memory wipes, and output re-computation reducing covert channel capacity below 0.01%, could potentially buy at least one year of delay. The author emphasizes that de-risking these three aggressive measures should be the primary focus of verification research, noting that their rapid implementation feasibility remains an open question, though recent thoughts suggest it is feasible and prototypes are expected soon.

Key takeaway

For AI scientists and policy makers considering AI development slowdowns, current inference-verification methods are insufficient. You should prioritize research and prototyping of aggressive measures like robust proof of work/memory, frequent memory wipes, and highly accurate output re-computation. Without these, training will likely continue unimpeded, potentially missing critical windows for safety and oversight development.

Key insights

Current inference-verification prototypes are likely ineffective, but aggressive measures could significantly slow AI training.

Principles

• AI training involves extensive communication.
• RL training requires minimal communication.
• Inference verification aims to restrict communication.

Method

Proposed methods include unplugging inter-rack cables, limiting bandwidth, periodically erasing clusters, and recomputing sample outputs to verify inference generations.

In practice

• Implement proof of work/memory for >95% compute.
• Perform memory wipes every few minutes.
• Reduce covert channel capacity below 0.01%.

Topics

• Inference Verification
• AI Safety
• AI Training
• Reinforcement Learning
• Proof of Work/Memory

Best for: AI Scientist, AI Engineer, Machine Learning Engineer, Research Scientist

Related on AIssential

• Rethinking Pre-Training for Agentic AI with Aakanksha Chowdhery - #759 — Agentic AI requires fundamental pre-training shifts, not just post-training, to enable multi-step reasoning and dynamic tool use.
• Zero knowledge verification for frontier AI training is possible — Zero-knowledge proofs can technically verify frontier AI training, enabling enforceable governance beyond self-reporting.
• DeepMind’s New AI Found A Strange New Way To Think — Reliable AI systems emerge from unreliable components via iterative refinement and robust judging loops.
• AI is deteriorating in realtime — AI models risk "model collapse" and increased hallucinations when recursively trained on synthetic data.
• Teaching Reasoning Models When to Stop: Data, Rewards, and Self-Aware Decoding — Efficient CoT reasoning requires balancing length reduction with accuracy, often by leveraging model confidence.
• not much happened today — AI agent development is shifting from interactive chat to robust, verifiable, and automated production systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Alignment Forum.