What inspired us We were looking for a major gap in AI governance. A problem that hasn't yet been picked up by AI regulation, but urgently needs to be solved to make LLMs safe. We found one: the internal use of very powerful models within AI labs itself, mostly for R&D purposes.
How we built it and what we learned Rather than proposing new AI-specific oversight from scratch, we asked whether provisions from an existing, mature, and regulated industry can be borrowed to govern AI labs' internal use of frontier LLMs for R&D. We chose banking, given our collective background in financial services. We wanted to understand how internal models are regulated there, and through a comparative architecture analysis, we identified four mechanisms most transferable to governing AI R&D.
Challenges we ran into We started broad, running a literature review to identify the risks associated with internally deployed models, then mapped those risks against the regulatory landscape to understand to what extent they've already been addressed. We wondered whether to limit our focus to binding law, or to also include voluntary frameworks.
Accomplishments that we're proud of We are proud that we chose to narrow our focus down to a single, constructive research proposal rather than staying broad: "Could any of the regulatory mechanisms built for banks' internal models be useful in mitigating risks from an increasingly automated process of LLM development?"
What's next for Neural Networking We consulted De Nederlandsche Bank (DNB) on three targeted questions related to our analysis. DNB responded substantively and consented to being cited, and — reflecting their interest in the outcome — has since extended an invitation to discuss our findings further.
Built With
- analogy
- architecture
- banking
- comparative
- framework
- literature
- regulatory
- review
Log in or sign up for Devpost to join the conversation.