Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Please don't call model weights source code. Code is something you can (usually) read and understand, if anything weights are closer to a very obfuscated compiled binary - although even that can be analyzed by a good enough reverse engineer.

"Open Source" models are the latest in a long series of attempts to take advantage of FOSS's reputation without actually releasing something that adheres to the principles.



To run a language model you need both the model weights and the source code of the implementation!


I would think truly open source means also including the initial training data. i.e. Everything needed to build[/tweak/customize] the thing from the ground up.


I agree, but I'm responding to someone confusing source code and model weights…


llama.cpp is open source!


Neat! But if we're sticking to the source code analogy, the matrix multiplier probably maps closer to a CPU or an interpreter - I wouldn't say every Java program is open source because you have an open-source JVM.


does that make every image closed source because you need a viewer to view them?


At least if you use the Free Software Foundation's definition of free software, one requirement is access to the unobfuscated source code. If every copy of a free-as-in-speech program binary's source code were deleted from existence by a god then the program would become proprietary.

I would say that every image using a proprietary file format is proprietary. If the creator of such an image intends for the image to be otherwise free as in speech (any receiver can inspect, modify, redistribute with and without modification for any purpose), then the image can be made free if the image creator converts the image to a free format.


Yes, which is why if you're serious you will exclusively use the terminal to protect your computer from even the passing chance of being exposed to a so called "image file".

In all seriousness, compilation and model training are lossy processes and erase a lot of the context needed to understand the output (and with model training we don't fully understand it even with access to the training data). Images aren't necessarily derivative of anything, so the analogy breaks down here.


often images are not only lossily compressed, but they are a composite of many images/layers/image transformations. these are lost when the image is flattened and then distributed.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: