Beyond Fine-Tuning: Frontiers in LLM Post Training, Hosted by Assembled

​Join us for a candid, forward-looking conversation on what happens after you pre-train a large language model—tailored for founders and builders who have to ship this year, not someday.

Our panel of hands-on researchers will unpack where post-training is headed in the next 6–12 months, how reward-hacking headaches are evolving, what “thought disclosure” might really look like, and which open problems Big Labs are scrambling to solve—so startups can place smarter bets today.

Expect technical depth, plenty of audience Q&A, and time to network over good snacks and drinks.

Panelists
Arushi Somani – Research Scientist at Anthropic
Arushi is a researcher at Anthropic, where she works on reinforcement learning and alignment for frontier Claude models. Arushi was a core contributor to Claude 3.7 and Claude 4, and her research focus on measuring and improving the faithfulness of reasoning models as well as supervision-free post-training methods. Previously, she worked on pre-training large multimodal models at Adept and deep neural networks for program synthesis at the Berkeley Sky Lab.

Bowen Baker – Research Scientist at OpenAI
Bowen leads the team working on chain-of-thought interpretability at OpenAI, where his work centers on topics such as detecting misbehavior in frontier reasoning models. His previous work at OpenAI has spanned topics including multi-agent autocurricula, state estimation from vision, attention based network architectures for reinforcement learning, and most recently multi-agent social dilemma games.

Jacob van Gogh – Research Scientist at Amazon AGI SF Lab
Jacob builds planning and agentic tool-use systems for Amazon’s newly formed AGI SF Lab, contributing to the lab’s “Nova” foundation-model program. Jacob has also worked on LLM planning and tool use at Adept. Previously he spent four years at Lyft, where he led the real-time online-learning initiative for dispatch and pricing.

Moderator: Joe Gershenson – Head of Engineering, Assembled
Joe leads engineering at Assembled, turning cutting-edge LLM research into production-grade support agents used by hundreds of enterprises. Before Assembled he directed research engineering at Adept, ran ML infrastructure at Stripe, and co-founded a YC-backed predictive-modeling startup.

Agenda
5:30 PM
Check-in, snacks & drinks
6:00 – 6:45 PM
Panel discussion
6:45 – 7:00 PM
Audience Q&A
7:00 – 8:00 PM
Networking

Spaces are limited! Please register early to secure your spot.

​​Questions? Contact melissa@assembledhq.com

​​Assembled is committed to creating inclusive tech events. If you require any accommodations to fully participate, please let us know when you register.

​​This event may be photographed for promotional purposes. By attending, you consent to potentially appearing in event documentation.