Teaching Everyone to Fish for Tokens
Interconnects Nathan Lambert
Open models may not copy Linux after all. They’re splitting into a costly training race and a longer-tail world of tuning, with Nvidia betting big on both.
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
People keep reaching for Linux as the clean analogy for open AI, but the fit is only partial. The source argues that open-source recipes for language models — the full training data, code, and process — resemble the open-source operating system story more closely than open weights do. Open weights, by contrast, are more like deployable versions of software: useful, portable, and often long-lived, but not the same thing as the recipe that made them possible.
That distinction matters because model weights age fast while still hanging around for years. Many companies are still built on workflows from Llama 3, even as agentic behavior has moved on. The real leverage, though, sits with the recipe. Projects like Olmo and earlier work such as Pythia show that a company can take the process, adjust it, and run it again to produce a fresh set of weights. In the best case, the community pushes improvements back into the next run.
Nvidia clearly sees the business logic. It is pouring money into nearly open-source models, including Nemotron, and releasing all the data it can legally share plus training code. The bet is simple: if more people can make “token machines,” demand for inference grows, and Nvidia sells more chips. The source says Nvidia is reported to be spending $26 billion on this effort, and that puts the whole strategy under a hard deadline.
There are two ways this can go. One is that it works: open-source recipes create so much extra chip demand that the investment pays off. The other is that it doesn’t, and open models drift into a separate lane, optimized for efficiency, modifiability, and specialization rather than head-to-head competition with the biggest closed labs. The source sees that second outcome as the likelier one, with open models serving a long-tail market of enterprise agents, on-prem systems, and repetitive private-data tasks.
That split is already visible in the market. More people are focused on post-training, finetuning models like DeepSeek V4 Flash, Inkling Small, and GLM 5.X through tools such as Tinker, which the source calls the most popular finetuning API today. At the same time, training a base model that can act as a general reasoner is getting more opaque, and the old pretraining/post-training vocabulary may not hold up much longer.
My take — AI-written commentary, not fact-checked reporting
The open-model crowd keeps talking like abundance automatically produces independence, but money still runs the show. Nvidia gets that better than most of the “open” believers: give away enough tooling and the market politely buys more of your hardware. Meta’s version is simpler and meaner — flood the zone with weights, then watch everyone else scramble for margin.
Read more about this at: Interconnects