Olmo 3 talk covers DPO, data problems, and how research reaches frontier models
natolambert · x · 2026-07-29
Olmo 3 post-training talk dissects DPO, data messiness, and frontier-model engineering
Nathan Lambert shares a podcast/lecture with Scott Geng about the messy realities behind Olmo 3 post-training and DPO.
- How a research idea makes it into a near-frontier model
- Why the hardest part of DPO is often the data, not the algorithm
- Organizational challenges in multi-stage post-training pipelines
- Reflections on where research is heading and how to think about DPO today
- The talk also revisits DPO fundamentals and the broader algorithm zoo
Related event: Olmo 3 Lecture Explores Post-Training and DPO Implementation(2 posts)→
More from Models
- LLM Empire Experiment: Claude and GPT Spontaneously Form Pacifist Alliance — wightmanr · 2026-07-30
- Warp Integrates Kimi K3, Claims 13% Better Task Completion Than Other OSS Models — vikvang1 · 2026-07-30
- Kimi K3 Architecture: KV Cache Offloading vs. KDA Recurrent State — zephyr_z9 · 2026-07-29
- Gemini 3.5 Flash aces a visual ordering puzzle with one move — iamrobotbear · 2026-07-29
- Claude Opus 5 tops a cybersecurity benchmark but becomes noisier when it overworks — Thom_Wolf · 2026-07-29
- Pangram Raises New Round, Launches Stronger AI Text and Image Detection Models — deedydas · 2026-07-29