[New Report] The Legal and Ethical Implications of Monosemanticity in LLMs, IPLR-IG-008
From Abhivardhan, our Chairperson
Glad to share an ambitious flurry of reports we may publish this September 2024.
Let's begin with this infographic report on "The Legal and Ethical Implications of Monosemanticity in LLMs, IPLR-IG-008" by Indic Pacific Legal Research LLP. I must admit it was a pleasure to develop this report with Samyak Deshpande, Sanvi Zadoo and Alisha Garg, former interns at the Indian Society of Artificial Intelligence and Law.
You can access this report at https://indopacific.app/product/monosemanticity-llms-iplr-ig-008/
This report was inspired by two major developments in the global artificial intelligence ecosystem - one in generative artificial intelligence research, and another in the domain of international artificial intelligence safety.
1️⃣ The first development, was Anthropic’s release of an ambitious research paper in mid-2024, "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3," which addresses the topic of complexity of large language models (LLMs) and how the extraction of monosemantic neurons takes place.
2️⃣ The second development, was the International Scientific Report on the Safety of Advanced AI (interim report) released by the Government of United Kingdom and other stakeholders who were part of the AI Seoul Summit, May 2024, as a follow-up to the Bletchley Summit (on AI Safety) held in 2023.
3️⃣ This report focuses on these two developments, and examines Anthropic’s work on monosemanticity, with technical, economic & legal-ethical perspectives.
4️⃣ The report also focuses on the evolution of neurosymbolic artificial intelligence and proposes some pre-regulatory ethical considerations that may be possible to be developed around this emerging class of AI technology.
5️⃣ Best part: nowhere the Anthropic paper hypes about their findings on the AI risk mapping they tried to point out, which is good. Hence, our recommendations are subtle in a certain way.
