Seb Krier Critiques Obsolete AI Regulation Frameworks
Why It Matters
Current legislative trends focusing on foundation models may fail to address risks emerging from how AI is actually deployed via software wrappers and tools. This debate could shift the regulatory burden from model developers to end-use application developers.
Key Points
- AI innovation is shifting from base model training to post-training, tool use, and inference scaling.
- Safety mitigations are increasingly implemented at the software scaffold level rather than inside the model weights.
- Horizontal AI regulations are criticized for being 'unsexy' but necessary in their complexity compared to single-intervention laws.
- Effective regulation requires domain-specific expertise and a focus on outcomes rather than entity-based restrictions.
Tech policy expert Seb Krier has argued that current AI legislative proposals are becoming obsolete due to a fundamental misunderstanding of the technology's evolution. Krier contends that recent advancements in AI capabilities stem primarily from post-training, scaffolding, and inference compute scaling rather than just core model development. Consequently, many proposed safety mitigations targeting the model level are ineffective, as safeguards like filters and classifiers are increasingly applied at the software scaffold level. The critique suggests that horizontal, 'one-size-fits-all' regulations fail to account for the complex chain of actors involved in modern AI implementation. Krier advocates for outcome-based regulation and domain-specific oversight, suggesting that the current focus on a few large labs ignores the distributed nature of AI innovation and risk management.
Imagine passing a law for car safety that only talks about the engine, while the brakes and seatbelts are actually part of the car's body. That is how Seb Krier describes current AI laws. He argues that most of the 'brains' and safety features of AI today don't happen when the model is built, but when it is put into specific apps and tools. Because of this, trying to fix AI risks by only regulating the big labs is like trying to stop cybercrime by tweaking a processor chip—it just doesn't work. We need laws that look at how the tech is used in the real world, not just how it was made.
Sides
Critics
Argues that current AI bills are obsolete because they focus on model-level mitigations instead of deployment scaffolds and specific outcomes.
Defenders
Proponents of the 'entity-based' and horizontal approach to risk management as seen in the EU AI Act.
Noise Level
Forecast
Legislators may face increasing pressure to pivot from broad 'AI Acts' toward sector-specific updates in cybersecurity, healthcare, and finance. We will likely see a growing divide between 'model-level' transparency requirements and 'application-level' liability frameworks.
Based on current signals. Events may develop differently.
Timeline
Krier publishes critique of AI regulation
Seb Krier outlines why current draft laws are behind the curve regarding post-training and scaffolding innovation.
Join the Discussion
Discuss this story
Community comments coming in a future update
Be the first to share your perspective. Subscribe to comment.