Frontier AI models are built from thousands of small decisions: data sourcing, filtering, mixtures, curricula, scaling experiments, optimizer choices, kernels, evals, failed runs, and protocols for deciding what gets scaled.
This is process knowledge.
Many users agreed that frontier AI models depend on detailed process knowledge rather than any single breakthrough idea, citing personal experience with papers that fell short for lacking such details.
No Digg Deeper questions have been answered for this story yet.
I wrote about why more of that process knowledge should become public, and why Marin is organized around open development of frontier AI.
https://openathena.ai/blog/open-development-of-frontier-ai/
This process knowledge is the real secret sauce of modern AI labs.
And it rarely enters the scientific record.
Open weights reveal the model. A technical report may gesture at the recipe. But the process that produced the model usually remains private.
Marin is our attempt to build in that direction: not just releasing open models, but making the process of developing frontier AI more legible, reusable, and public.
This is understandable. Businesses manufacture knowledge for business ends.
But that means transparency is often tactical: a way to establish credibility, attract attention, recruit talent, or capture mindshare.
Once those goals are achieved, disclosure shrinks.
Not every detail can be open.
But the default should move toward legibility, reproducibility, and public accumulation rather than private inheritance.
The result is that academia and the public are increasingly locked out of understanding how frontier AI is actually built.
If open labs and academia are going to contribute meaningfully to AI, more of this process knowledge has to become public.
@dlwh I think the blog post date mignt be wrong (its dated in the future)
@dlwh This is my pet theory why almost none of the optimizers invented in the last 10 years have stuck, the actual formula for the update that you can slap your name on is a tiny part of what makes things work, not really useful without accompanying process
@dlwh Couldn’t agree more. I often see folks thinking that there’s this one big idea behind a powerful model but in reality, most of the time, it’s about getting many small things right rather than inventing a new big thing.
@RishiBommasani i write from the future. :-) Thanks, will get it fixed
@yaroslavvb 100%. I almost called out optimizers specifically. I'm on one such paper and the lack of process knowledge is the main reason it didn't hold. But I think you see it in data mixing/filtering work too, and of course arch. really everywhere.
@dlwh This is my pet theory why almost none of the optimizers invented in the last 10 years have stuck, the actual formula for the update that you can slap your name on is a tiny part of what makes things work, not really useful without accompanying process