Uploaded October 2026 | Updated November 2026, 2 days ago
Will Brown of Prime Intellect makes the case that continual learning is about to get boring, in a good way: it is the software development life cycle, applied to models. He traces how agent workflows grow from terminal experiments into production software, why evals should capture the vibe check, and which post-training methods he reaches for, from SFT and on-policy and context distillation to online RL, along with why he is skeptical of jumping straight to OPSD. He closes on build versus buy for models, what quant finance shows about running many specialized models, and why post-training will become as routine as CI/CD.
Chapters:
0:00 Why continual learning is about to get boring
1:03 The sci-fi version of learning, and the boring one
2:21 Tools versus software in how we use agents today
3:32 Production agents start to look like software
6:44 Evals that capture the vibe check
9:28 SFT, on-policy distillation, and context distillation
10:31 Why OPSD is not magic
10:55 RL as the button that makes your metrics go up
12:32 Going environment native with a dev and a prod
14:04 Build, buy, or fork your model
17:19 The model development life cycle
18:46 Post-training will get as routine as CI/CD
Resources:
→ Prime Intellect: primeintellect.ai
→ On-Policy Self-Distillation (OPSD) paper: arxiv.org/abs/2601.18734
Recorded at LangChain Presents: The Learning Loop, an evening on continual learning at LangChain's San Francisco office.
Will Brown of Prime Intellect makes the case that continual learning is about to get boring, in a good way: it is the software development life cycle, applied to models. He traces how agent workflows grow from terminal experiments into production software, why evals should capture the vibe check, and which post-training methods he reaches for, from SFT and on-policy and context distillation to online RL, along with why he is skeptical of jumping straight to OPSD. He closes on build versus buy for models, what quant finance shows about running many specialized models, and why post-training will become as routine as CI/CD.
Chapters:
0:00 Why continual learning is about to get boring
1:03 The sci-fi version of learning, and the boring one
2:21 Tools versus software in how we use agents today
3:32 Production agents start to look like software
6:44 Evals that capture the vibe check
9:28 SFT, on-policy distillation, and context distillation
10:31 Why OPSD is not magic
10:55 RL as the button that makes your metrics go up
12:32 Going environment native with a dev and a prod
14:04 Build, buy, or fork your model
17:19 The model development life cycle
18:46 Post-training will get as routine as CI/CD
Resources:
→ Prime Intellect: primeintellect.ai
→ On-Policy Self-Distillation (OPSD) paper: arxiv.org/abs/2601.18734
Recorded at LangChain Presents: The Learning Loop, an evening on continual learning at LangChain's San Francisco office.










