"debatable"

Posted Jul 24, 2024 19:02 UTC (Wed) by Heretic_Blacksheep (guest, #169992)
In reply to: "debatable" by josh
Parent article: Zuckerberg: Open Source AI Is the Path Forward

Having skimmed the license, it reads like another "source available but don't compete with us" license, so no... neither open source as defined by the OSI nor free software as defined by the 4 freedoms. But it's also not a typical "you see nothing, you pay for the privilege of using" proprietary license. It's a BSL derivative.

So Zuck either doesn't know what open source really is, or he's being deliberately deceptive. Perhaps both. The former doesn't exclude the later. Given Zuck's & Meta's public history, it's easy to conclude this is a deliberate obfuscation and he personally doesn't care what "open source" is, only that it's another buzzword he can use. FOSS washing?

The source is available to view, and the models can be audited, but you can't exercise all the freedoms associated with FOSS. For some, that's enough. But any businesses really should beware of those wishy-washy non-compete clauses.

"debatable"

Posted Jul 24, 2024 19:27 UTC (Wed) by Cyberax (✭ supporter ✭, #52523) [Link] (1 responses)

The source code for LLMs is pretty much a sidenote. All the interesting stuff is in the weights and the training data.

You can run LLMs with about 1000 lines of pure C code without any external dependencies.

"debatable"

Posted Jul 31, 2024 7:12 UTC (Wed) by cpitrat (subscriber, #116459) [Link]

Compared to a regular program:
- the source code of the model is comparable to the source code of the assembly
- the weights are comparable to the executable/machine code
- the training data is comparable to the source code (and you may consider it generated source code, in which case the scripts scraping the web / downloading databases to generate it would be source code generators)
- the training procedure is comparable to the build system configuration (e.g makefiles)

An appropriate open source/free license applicable to machine learning models would have a vocabulary that encodes these equivalences.