AI Research Night — Case Studies in Pretraining
with ElevenLabs, Geodesic Research & PostHog · at the Shoreditch Treehouse
If you've pretrained a model before you know you need data worth learning from, compute, and a serious answer to "why not just call an API?" This is an evening with three teams who have answers — at different scales — followed by an open discussion and plenty of canapés to keep us going.
Talks (15 min each + Q&A)
Pretraining a GPT on a laptop — Angelos Perivolaropoulos, (Research Engineer at ElevenLabs) - What intuition from 10M parameters does (and doesn't) transfer upward.
A from-scratch encoder for session replay data — Nico Waltz, (AI Research Engineer at PostHog) - A self-supervised encoder trained directly on DOM recordings, built to replace a video-rasterizer + VLM pipeline in production.
Alignment pretraining — Cameron Tice (Co-Founder & Executive Director, Geodesic Research) - Lessons from training 6.9B-parameter LLMs on 500B tokens across four different data conditions to test how AI discourse in pretraining corpora affects alignment.
Who this is for
Active PhD students and AI researchers with hands-on experience applying their work — people who have trained models. Registration is approved manually.
Agenda
18:30 — Doors, drinks
19:00 — Talks
20:10 — Open discussion
20:40 — Dinner in the treehouse