Conversation

Notices

1. Embed this notice
   Josh Gay (joshuagay@metasocial.com)'s status on Saturday, 06-Jul-2024 08:24:17 JST Josh Gay

   If people are on the fence on whether a given AI model use data or object code, then lets just use real example AI-programs and put it through the rigor of real GPLv3 style compliance on real programs.

   I did this with my own projects (which now feel ancient being over 20 years old) and this convinced me beyond a doubt that training data needs to be provided to comply with the letter of the GPL. It is robot controller based on an AI model. Corresponding source includes training data.

   • Embed this notice
     Alexandre Oliva (lxo@gnusocial.net)'s status on Saturday, 06-Jul-2024 08:24:16 JST Alexandre Oliva

     in reply to
     if someone gets the model but not the training data, which one/s of the four essential freedoms appear to be missing in your assessment?
   • Embed this notice
     Alexandre Oliva (lxo@gnusocial.net)'s status on Saturday, 06-Jul-2024 13:06:59 JST Alexandre Oliva

     in reply to
     I understand the scenario you describe, but I don't see how it relates with the question.
     say I have the model but not the blobs used to train it (and neither the BigNumber library source nor object code)
     ISTM that I can start from that model and (1) study it to learn what it does, and train it further so as to adapt it to do what I wish; (2) make copies of the model and distribute it; (3) improve the model and distribute the improvements; and (0) run the model for any purpose.
     so, despite not having the training data, ISTM that I have the four essential freedoms nevertheless.
     do you not agree with this assessment?
     note I'm not saying that having the training data (and free software used to create it) is not desirable, just that AFAICT it's not strictly necessary to have the essential freedoms
     GNU Too likes this.
   • Embed this notice
     Josh Gay (joshuagay@metasocial.com)'s status on Saturday, 06-Jul-2024 13:07:00 JST Josh Gay

     in reply to

     • Alexandre Oliva

     @lxo Let's say I have created a cool BigNumber library that performs math functions really well.

     So I take my library and generate a lot of small object code blobs. Labels are the function and parameter calls associated with the object code blobs (fft, cosine, lambda, etc).

     I train a model with a feed forward neural network so that the function call labels can be used to predict blobs.

     My free software programs replace GNU's big number library with this AI model.

   • Embed this notice
     Alexandre Oliva (lxo@gnusocial.net)'s status on Saturday, 06-Jul-2024 15:31:12 JST Alexandre Oliva

     in reply to
     I agree the model is not transparent (I think that's what you meant, rather than opaque)
     I agree it's hardly possible to grasp what the values in a model stand for
     what I don't really see is that the training set offers better insights into the working of the model, especially when it comes to massive sets.
     one might try to come up with theories, but in the end, whatever "understanding" one might infer from the training set is either equivalent and isomorphic to the model, or a misrepresentation of the model.
     the former case would suggest that the model is its own source code, hard to grasp as it is
     the latter case wouldn't really give you an understanding of the model
     
     as for retraining, my limited knowledge of the subject matter suggests that additional training can bring the model to work as desired, regardless of the training that led to the initial model. it may be challenging to gain confidence that the further-trained model has no traces of responses one would like to suppress, but that is also the case in retraining from scratch, given the wild inferences and conclusions that such models may make from irrelevant details in the training sets. I'm curious as to any solid evidence you might have that might challenge this understanding
     thanks for engaging!
   • Embed this notice
     Josh Gay (joshuagay@metasocial.com)'s status on Saturday, 06-Jul-2024 15:31:13 JST Josh Gay

     in reply to

     • Alexandre Oliva

     @lxo

     I strongly disagree. It is not opaque enough.

     You do not have the freedoms to:

     Study and understand: A model's behavior heavily relies on its training data. Without it, fully grasping why the model makes specific predictions is difficult (nearly impossible for sophisticated models).

     Improve and adapt: Effective improvements often require retraining with modified or additional data. Without the original training data, replicating or enhancing the model's capabilities is challenging.

   • Embed this notice
     Alexandre Oliva (lxo@gnusocial.net)'s status on Sunday, 07-Jul-2024 18:09:15 JST Alexandre Oliva

     in reply to
     it's quite a leap to assume I'm taking a position. I'm a philosopher asking questions of someone who's quite familiar with the essential freedoms we've identified to form a position. if I had take a position, you wouldn't be seeing so many questions, you'd be seeing a lot more assertions.
     I had more questions, but I sense the interaction has made you uncomfortable, so I'll leave you alone. thanks for your kindness.
     as for why your initial statement surprised me and led me to explore the discrepancy between it and my thoughts, my reasoning had been that we don't need to look at the entire learning history of a person to assess what the person knows of a subject, nor to teach that person something new. my understanding is that our brains are still a lot more complex than model-based generators, so it seems very surprising to me that techniques that work on more complex systems wouldn't be enough for simpler ones. computing theory taught me that you can generally use higher-complexity solutions to solve lower-complexity problems, even if inefficiently, but lower-complexity solutions just don't work for higher-complexity problems.
   • Embed this notice
     Josh Gay (joshuagay@metasocial.com)'s status on Sunday, 07-Jul-2024 18:09:16 JST Josh Gay

     in reply to

     • Alexandre Oliva

     @lxo one other question. Why are you taking a position when you clearly have absolutely no idea how any of this works? This whole exchange left me feeling gross like after talking to an elected official that is still always campaigning. Just entirely disengenuis.