Can we please talk for a moment about the giant gaping hole in OpenAI's portfolio of offerings?

While the announcements made during Dev Day were an absolute whirlwind of joyous news, there was one thing I desperately wished would come but never did.

Publishable fine-tunings…

Right now you can, of course, do a fine-tuning job and offer a product which uses it, no problem.

But, what if…

You could create a fine tuned model and then… Publish it?

Make it public and share it with the world.

OpenAI could incentivize the process with revenue sharing like they’re planning on doing with custom GPTs, but the math is way simpler—just give devs who create a fine-tune x% of the revenue.

Devs who do a fine-tune could publish their model either for-use-only or as a new quasi-foundational model others could use as a base model upon which to do further fine-tuning (and continue to earn a percentage of downstream revenue[1] from or perhaps a share of the training revenue).

Then, perhaps as part of the publishing agreement would be that OpenAI themselves would then have the option of using those models themselves to augment the inference ability of ChatGPT if, for instance, a particular fine-tune was exceptionally strong in some narrow task.

I imagine there’s probably some issues or concerns about IP rights, but I imagine most of those could be pretty readily handled.

I’m not sure where something like this falls given their governing mission statement, but I think it is a great way to explode the reach, power, and influence of OpenAI’s models…

When I see the ecosystem of Stable Diffusion fine-tunings, I can’t help but wonder what the LLM community could do given the opportunity.



Please also allow fine-tunings on text-embedding-ada-002.


  1. I imagine as a function of the percentage of total training tokens contributed to the final model and relative position in the model lineage ↩︎


That would be awesome! But I think it’s worth noting that by not allowing people to publish their fine-tunes, we’re essentially forced to publish the dataset and hyperparameters instead, which might be better for transparency. :thinking:

1 Like