AI Safety Chief Urges Billions for Verified ML Infrastructure
Geoffrey Irving advocates formal verification methods for AI despite imperfect specs and hardware risks.
Geoffrey Irving, Chief Scientist at UK AISI, argues that wide verification of machine learning infrastructure merits billions in compute investment. He cites a RAND report on formal methods for trustworthy AI deployment, noting that while specs will be imperfect and human or hardware elements like Rowhammer remain vulnerabilities, such verification is still highly valuable. Policy expert Seb Krier echoed the view, emphasizing ongoing limitations in verification systems. The discussion centers on securing AI through mathematical formal methods rather than eliminating all risks.
Combined views
782
2 posts, first seen 15h ago
AI Safety Chief Urges Billions for Verified ML Infrastructure
Geoffrey Irving advocates formal verification methods for AI despite imperfect specs and hardware risks.
Geoffrey Irving, Chief Scientist at UK AISI, argues that wide verification of machine learning infrastructure merits billions in compute investment. He cites a RAND report on formal methods for trustworthy AI deployment, noting that while specs will be imperfect and human or hardware elements like Rowhammer remain vulnerabilities, such verification is still highly valuable. Policy expert Seb Krier echoed the view, emphasizing ongoing limitations in verification systems. The discussion centers on securing AI through mathematical formal methods rather than eliminating all risks.
Reactions from ranked influencers
2 posts
This won't solve all the problems! Our specs will be imperfect, any human component to the sandbox will be a weakness, and Rowhammer-ish stuff may bite eventually. But wide verification is very much worth an investment of billions of dollars of tokens. https://www.rand.org/pubs/research_reports/RRA4881-1.html
This won't solve all the problems! Our specs will be imperfect, any human component to the sandbox will be a weakness, and Rowhammer-ish stuff may bite eventually. But wide verification is very much worth an investment of billions of dollars of tokens. https://www.rand.org/pubs/research_reports/RRA4881-1.html