You will flesh out the AI verification mechanism “Networking Equipment Interconnect Limits”, which imposes bandwidth restrictions on external communication with a pod of chips. You will dig into the security and efficacy of the proposal as well as implementation details.
About the project
Previous work from myself (https://arxiv.org/abs/2506.15867, p. 120) and others (https://www.rand.org/pubs/working\_papers/WRA3056-1.html, “fixed set”) has discussed “Networking Equipment Interconnect Limits”, a potential mechanism for verification of international AI agreements. In effect, this mechanism would restrict the size of clusters in AI data centers by severely limiting the external communication of some pod of AI chips. I think this mechanism is especially exciting among verification approaches, and I am excited to support others who want to work on it (but don’t have enough time myself). This project would involve researching the effectiveness of the approach and identifying key implementation details, such as the pod size, external bandwidth limits, and whether existing hardware could be sufficient. The intended output would be a paper exploring this mechanism in considerable detail, understanding key security assumptions (and how likely these are to hold), key efficacy assumptions (and how likely these are to hold), and positing implementation details if this mechanism was needed soon. For some participants, the project could involve prototyping the mechanism. Participants would do deep dives on topics such as decentralized/distributed training, hardware infrastructure used in frontier AI development and deployment, and hardware side-channel attacks.
Theory of change
In the future, we might have international agreements about AI, including potentially an international halt on AI development or deployment. Because countries don’t trust each other, such agreements would need verification measures to ensure everybody is following the rules. Unfortunately, most of the mechanisms we would like to use for verification have not yet been developed, and some of them have development timelines of many years. That’s a problem if we end up wanting to do an international agreement earlier. This project would push forward one particular verification mechanism that might be especially useful in the near-term, as it is not technologically complex. Having this idea be fleshed out in more detail would both increase the likelihood of it being adopted (and increase the viability of having international agreements at all) and increase the likelihood of the mechanism being effective. See this research agenda for motivation around a halt on dangerous AI (https://arxiv.org/abs/2505.04592). See this report on approaches to verification and why this one is promising (https://arxiv.org/abs/2506.15867).
Your role
Mentees will lead the research and carry out the majority of the work. They will assess what methodologies to use (with advisement), assess the quality of evidence they come across, and contribute the bulk of the intellectual work on the project.
Prerequisites
Proficiency with Python and PyTorch.
Interest in pursuing AI governance or policy professionally.
Experience reading AI papers (e.g., at least 12 total hours spent reading papers on arxiv, at least 10 papers).
Completion of AI Safety Fundamentals or an equivalent intro to AI safety/alignment course, could be an AI governance course.
Comfortability pursuing self-directed research.
Location preference
No preference
Application question(s)
For both questions, you are allowed to use AI assistance, but please clearly state how you used AI for each question (e.g., ChatGPT helped understand background idea X, Gemini helped edit the first draft I wrote, Claude did the whole thing and I just hit submit :) Explaining how you used AI does not count in the time toward each question. Do not discuss the questions with other humans.
-
(15 min) Using the Llama 3.1 paper (https://arxiv.org/abs/2407.21783), please spend 10 minutes trying to figure out how many days the 405B model was trained for (the pre-training phase). Then spend 5 minutes explaining how you arrived at your final answer (e.g., what evidence or calculations you used, how confident you are). Your answer could include a best guess point estimate, a range, or something else, depending on your confidence. You are allowed to consult sources other than the Llama paper. Do not spend more than 20 minutes total. 300 words max.
-
(30 min) Google recently announced their new TPU, Ironwood (https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/). Read the announcement post with a skeptical eye. Then write up a summary of what readers need to know about the chip and any ways in which the blog post presented information in a misleading way. I mainly care about the content you write rather than the language—no need to write something verbose. So you're basically writing a readers guide to that blog post, giving readers the summary and pointing to any places where Google presented things in a misleading way. You are allowed to consult sources other than the blog post.
-
If your resume/CV or your other answers say that you have written some paper or done some project previously, please provide a link. If I cannot read your previous work, I will disregard it in my application review. For multiple-contributor projects, it is also useful for you to explain what parts you are responsible for.
About the mentor
Aaron Scher
Machine Intelligence Research Institute (MIRI)
Aaron’s research at MIRI is focused on International Coordination on AI, with an emphasis on Verification Mechanisms. He earned his Bachelor’s degree in psychology from Pitzer College in 2022 and quickly transitioned to working in AI safety. After a year of up-skilling and helping grow the field, Aaron started doing independent AI alignment research with the MATS program in summer 2023. He later managed 4 research teams through SPAR, the Supervised Program in Alignment Research working on sycophancy and interpretability. Aaron joined the Technical Governance Team in July 2024.