Comment on Voxtral Transcribe 2: transcribes at the speed of sound

• Tetsuo@jlai.lu ⁨1⁩ ⁨day⁩ ago

  But a modeled isn’t “compiled”.

  The weights are fully readable. Every single one of them.

  In a binary you have to use special software to get to the source code. The weights are the source and can be freely used to create a model with. The weights are used “as is” no transformation has to be done like when reversing a closed source binary that could also use obfuscation to make it more difficult.

  That’s why I would like to insist that open weights are not like a binary as they are usable as is essentially. When I use a model like that my computer is no executing the weights like instructions. They are aptly named “weights” for a reason and are a mere reflection of what the model learned through training.

  ---

  Disclaimer, below this is the explanation between a closed source binary and open weights by an LLM :

  A closed-source binary is like getting a sealed, opaque machine. You push a button, it gives an output. You have no idea how it works inside.

  Open weights are like getting the complete, labeled blueprints (architecture) and the exact, measured specifications for every single component—every gear, wire, and circuit (weights). You don’t have the designer’s engineering notebooks (training data/recipe) or the factory that built it (training compute), but you can absolutely build the machine yourself, measure every part, swap out components, and try to improve it.

  source

  • XLE@piefed.social ⁨1⁩ ⁨day⁩ ago

    The source for creating the model is closed, locked, heavily guarded as a corporate secret. But unlike code for software, this data might be illegally or unethically gained, and it might even mandate Mistral to publish it.

    You can “read” the assembly language of a freeware EXE program just as easily as you can “read” the open model of a closed source LLM blob: not very easily. That’s why companies freak out over potential hidden training data: not even they understand their compiled outputs.

    source

    • Tetsuo@jlai.lu ⁨1⁩ ⁨day⁩ ago

      I agree completely that it is scary we use extensively something that we truly dont understand. But even with the cleanest and most open source and open weights model ever, this statement would be just as true.

      With parameters in the billions all interacting with each others it’s extremely hard to analyse. LLM are inherently hard to understand fully not really because of the opaqueness of the training data but rather the raw number of parameters involved in inference.

      I would certainly be more appreciative of an open weights model were we destroy the environment (I have no choice in that) but at least we get the resulting matrix of parameters rather than OpenAI shit were we have absolutely nothing.

      IMO a model like Deepseek is most certainly a clone of someone else model that published their open weights. An open weights model enabled someone else to build on top of whatever controversial base was used. So at least it can be used by a wider audience and the original trainers don’t have their say in what you do as long as you respect the licence.

      I think the above process has pretty much nothing to do about the struggle of some linux dev losing hair trying to make sense of a Microsoft obfuscated binary…

      I guess many will not share that “lesser evil” point of view on open weights model but it’s not like anybody is gonna stop the AI conglomerate to do their shit in the complete dark and under no regulation whatsoever.

      source