Open-Source AI & Open Models Reading List
Interconnects Nathan Lambert ● Covered by 3 sources
A new reading list maps the open-model debate, from strategy to safety. It also shows how China, distillation, and regulation have pulled the fight into the center.
Based on reporting by Interconnects, Nathan Lambert — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Interconnects has put together a running reading list on open models, last updated on 11 Sep. 2026, and it’s meant to be more than a pile of links. The pitch is simple: if someone wants to understand where open models stand now, this is supposed to be the shortest path to getting oriented without pretending the story is settled.
The list is organized around a few arguments that have defined the last few years. There’s the business case for releasing models, the idea that open and closed systems sit on a gradient rather than a clean binary, and the claim that open models will matter less as direct substitutes for the frontier than as tools for custom enterprise workflows. That part also comes with a blunt reminder: open models have been trailing closed ones on performance, and adoption has not followed the same curve for both.
Then comes the geopolitics. The list says the U.S. needs to invest in open models as basic research and innovation infrastructure, while China has kept a leading position in open models over time. It points readers to pieces on Chinese open-source history, structural advantages, how Chinese labs describe their own model-building, and why their releases have stayed so competitive. The examples are not abstract either. Western companies have already been drawn toward Chinese models, in some cases to cut costs, and lawmakers have started probing that use.
A lot of the urgency sits in the technical section, especially around distillation. The list calls it the single most eventful debate around open models in 2026, and the stakes are obvious: how much does training on another model’s outputs help, what can be extracted through APIs, and how much of the current gap is really closing because of that transfer. Interconnects says the open-closed gap has narrowed to roughly 4-6 months, with leading open models coming from Chinese labs since around 2024.
There’s also a clear safety thread running through the reading list. It includes work on marginal risk, cybersecurity, and the argument that frontier open weights need a serious safety path rather than hand-waving. The list doesn’t pretend this is a solved policy question. It reads more like a map of the arguments people will keep fighting over, with the uncomfortable possibility that the next big open-model advance may also be the next regulatory headache.
Open models are no longer the polite, underfunded side project of AI. They’re where business strategy, U.S.-China competition, and model extraction fears all meet, which is exactly why the debate gets so noisy.
My take — AI-written commentary, not fact-checked reporting
The useful move now is to stop talking about open models as if they’re a purity test. They’re becoming industrial plumbing, and industrial plumbing gets regulated, copied, and abused. Anyone still pretending the question is simply open versus closed is missing the part where China, costs, and distillation have already made that binary look quaint.
Read more about this at: Interconnects