There is no such thing as Open Source AI. Why? Because Open Source was invented explicitly for software source code… and Neural Net Weights (NNWs) are not software source code — they are unreadable by humans, nor are they debuggable. Furthermore, the fundamental rights of Open Source also don’t translate over to NNWs in any congruent manner. Why is pointing out this fact so important? Actually, it is extremely important given that we are potentially on the precipice of government regulation of computation. If “Open”AI succeeds at regulating intelligence, current Open Source licenses (MIT / Apache) which are sloppily (no current better options) applied to NNWs will not stand on defensible legal grounds. Therefore, we urgently need to standardize as an industry on the following: “Open Weights” licensing frameworks with rights similar to the founding Four Freedoms of Open Source — but purpose built and designed for NNWs. I’ll be working on this with my partner @HeatherMeeker4 (the worlds leading open source IP / licensing attorney) and others. Contributions welcome!
@JosephJacks_ Why not generalize this to open data? Hasn't @openstreetmap done some licensing work in this direction?
@JosephJacks_ Transparency is important. Detailed documentation describing model architecture, training methodology, hyperparameter configurations, etc should be present along with weights.
@JosephJacks_ Weights are not opaque, you can debug them.
@JosephJacks_ Freedom of speech is too precious. Legal protections are not enough. Cryptography is a good start.
@JosephJacks_ Interesting observation! I think this is the paradigm now haha
@JosephJacks_ Interesting observation! I think this is the paradigm now haha
@JosephJacks_ Spare your soul the toll. This is an Open AI take on Open Source lol
Sorry but it seems like a flawed position to me JJ. Many open source licensed projects can be compiled to an artifact that is not debuggable by humans. Heck most humans can’t debug Go let alone a binary. How is this now different for neural nets? As long as the source is available to reproduce the artifact, it isn’t. Plus, open source means free, as in freedom to change the source however you want to as long as you retain a copy of license with the source. It’s not common practice to OSS the artifact as artifacts are not considered source, implicitly. The issue you are looking at here is more the act of open sourcing models without the raw training data? Simple answer. They aren’t open source. Period.
@JosephJacks_ Just abolish copyright, problem solved.