Reasoning Traces by Together Fund
Continuing our series of research salons. Reasoning Traces is Together's evening for going deep on a single idea, in a small room, with the people actually building and researching it.
WHY ALIGNMENT, WHY NOW ?
Alignment is the problem of making AI systems actually do what we intend — not just what we asked. As models grow more capable, the gap between "capable" and "aligned" becomes the most consequential open problem in the field.
The field has crystallized around four objectives — Robustness, Interpretability, Controllability, Ethicality (RICE) — and split into two big fronts: forward alignment (training systems to be aligned — RLHF, preference modeling, scalable oversight for tasks humans can't easily supervise) and backward alignment (proving they actually are — safety evals, interpretability, governance).
A real question is now on the table: as models start doing tasks we can't effectively oversee, does alignment scale with capability — or quietly fall behind it?
WHO'LL BE IN THE ROOM
Researchers working on alignment, interpretability, and evals
Founders building with frontier models who live with these tradeoffs daily
The sharpest engineers thinking about oversight, red-teaming, and safety
Anyone genuinely interested in how we keep capable systems pointed the right way
Link to relevant pre-read - https://arxiv.org/abs/2310.19852
HOW THE NIGHT RUNS
We open by framing the topic. Then the person deepest in it sets the scene - what's real, what they're seeing, the open questions. From there the floor opens: you jump in, push back, build on it.
It's a discussion, not a lecture. Loose by design; the good stuff happens when a room thinks together, not when someone lectures from a stage.
We have limited seats to ensure rich conversation so RSVP early!