|
|
|
|
|
|
by dilap
1272 days ago
|
|
|
Publishing research is pretty normal for traditional companies. The pledge to make patents available is a small step towards open. But I think the obvious and natural thing for a company in the business of training machine learning models that claims to be open to do is make the models themselves available. OpenAI does rougly the opposite: not only do you not have acccess to the underlying models, even the API access you are given is to a model itself deliberately trained to avoid answering certain classes of queries. To me that's the opposite of open; it's closed, restricted, and centrally controlled. (Very impressive results tho!) |
|
|
Published research is "open" to the extent that it is transparent but it is not "open" to the extent that it can be used and accessed by people. Unless you are an AI researcher, half these papers (to be generous) might as well not exist.
My argument is from that perspective (ability for the average person to use it), academic research only gives the illusion of openness.
Not only that but the training data is often -- but not always -- omitted from academic research. So reproducing the exact results they did is often out of reach without a significant investment in building your own collection of training data.
For example: Facebook and Google have both announced similar technology to OpenAI yet neither is usable out of the box (or at all for practical purposes) where as OpenAI despite being "closed" I can get started in 5 minutes.
Take by contrast to both of those, Stable Diffusion. Which I think is miles ahead of DALL-E... their code and their pre-trained weights are very easy to use as well as being open.