Open Models Are More Than Weights: How Do Distillation, Risk, and Access Count?

Nathan Lambert gave a guest lecture in Harvard CS 288IR: AI Safety, on the topic of Open Language Models. Zach Mueller reposted this video.

The video thumbnail is a screenshot of a slide: on the left, the title is Open Language Models, and below it are Nathan Lambert, Guest lecture, CS 288IR: AI Safety @ Harvard University, dated October 22, 2026. On the right, someone is standing by a window, with autumn outside.

Open models are often reduced to a single action: releasing the weights. The lecture breaks the question down more finely, with distillation, risk, and access rights all on the list. Nathan Lambert mentioned that he revisits these questions about every six months to take stock of the state of the field.

The lecture covers questions including: what an open model is, why we should care, how competitive the best open models are today, how open models are built, what role distillation plays, whether people are actually using open models, and where things go next.

Chapters:

• 00:00 Introduction
• 01:36 What is an open model?
• 08:47 Why should we care about open models?
• 18:52 How competitive are the best open models today?
• 25:28 How are open models built? The role of distillation
• 34:52 Are people actually using open models?
• 41:47 What comes next with open models?
• 47:32 Conclusions

For readers without a technical background, open models usually refer to language models whose weights can be downloaded and run locally, but the boundaries are not clean-cut. For readers with a development background, distillation, risk management, and changes in access methods are the parts of the open-model ecosystem that are easier to skip over and that also have a greater impact on subsequent choices.

At the end, Nathan Lambert asks the audience: what questions remain unanswered? Original post link: https://x.com/TheZachMueller/status/2114698539687137730