Research
Iliad researches the most pressing open questions in technical AI alignment, with staff researchers and Fellows collaborating closely across our research streams.
Our research areas
We cover a wide spectrum of theoretical alignment research.
01
Agent foundations and abstractions
02
Debate
03
Heuristic arguments
04
Geometry vs. structure in ReLU networks
05
Learning theory (SLT and non-SLT)
Connected to our teaching
Nearly every module of the Iliad Intensive has a research team working on it.
Many of our researchers first join us through the Intensive and the Fellowship.
For an overview of the topics in our orbit, browse the Iliad Intensive curriculum.
Our goals
Grow the field, keeping the bar high
We want more people working on these research streams without lowering the bar for rigor. The math and math-adjacent community is facing a crisis — we want to steer this unbounded energy toward alignment-relevant problems.
Encourage more cohesive and diverse research
Areas like debate and heuristic arguments benefit from cross-pollination with other fields, such as learning theory. We also want to diversify our bets, for example by investing in learning theory beyond SLT.
Prepare for the agent swarms
We are building infrastructure for close collaboration between researchers and (swarms of) agents: a shared Zulip and a shared research repository that both humans and agents can work from.
Build a curated research database
In the interest of moving fast, we need to move away from conference papers as the main unit of research communication (this system is clearly Goodharted). We want to cultivate a shared knowledge base that can grow quickly. Important information can be shared through papers, blog posts, and Lean repos.
Research with Iliad
The Iliad Fellowship is the main way researchers join our work.
→ Explore the Fellowship