Comment on Bill proposed to outlaw downloading Chinese AI models.

• jarfil@beehaw.org ⁨5⁩ ⁨weeks⁩ ago

  Open source requires giving whatever digital information is necessary to build a binary.

  In this case, the “binary” are the network weights, and “whatever is necessary” includes both training data, and training code.

  DeepSeek is sharing:

  • NO training data
  • NO training code
  • instead, PDFs with a description of the process
  • binary weights (a few snapshots)
  • fine-tune code
  • inference code
  • evaluation code
  • integration code

  In other words: a good amount of open source… with a huge binary blob in the middle.

  source

  • teawrecks@sopuli.xyz ⁨5⁩ ⁨weeks⁩ ago

    Is there any good LLM that fits this definition of open source, then? I thought the “training data” for good AI was always just: the entire internet, and they were all ethically dubious that way.

    What is the concern with only having weights? It’s not abritrary code exectution, so there’s no security risk or lack of computing control that are the usual goals of open source in the first place.

    To me the weights are less of a “blob” and more like an approximate solution to an NP-hard problem. Training is traversing the search space, and sharing a model is just saying “hey, this point looks useful, others should check it out”. But maybe that is a blob, since I don’t know how they got there.

    source

    • jarfil@beehaw.org ⁨5⁩ ⁨weeks⁩ ago

      There are several “good” LLMs trained on open datasets like FineWeb, LAION, DataComp, etc. They are still “ethically dubious”, but at least they can be downloaded, analyzed, filtered, and so on. Unfortunately businesses are keeping datasets and training code as a competitive advantage, even "Open"AI stopped publishing them when they saw an opportunity to make money.

      What is the concern with only having weights? It’s not abritrary code exectution

      Unless one plugs it into an agent… which is kind of the use we expect right now.

      Accessing the web, or even web searches, is already equivalent to arbitrary code execution: an LLM could decide to, for example, summarize and compress some context full of trade secrets, then proceed to “search” for it, sending it to wherever it has access to.

      Agents can also be allowed to run local commands… again a use we kind of want now (“hey Google, open my alarms” on a smartphone).

      source

      • p03locke@lemmy.dbzer0.com ⁨5⁩ ⁨weeks⁩ ago

        There are several “good” LLMs trained on open datasets like FineWeb, LAION, DataComp, etc.

        Then use those as training data. You’re too caught up on this exacting definition of open source that you’ll completely ignore the benefits of what this model could provide.

        an LLM could decide to, for example, summarize and compress some context full of trade secrets, then proceed to “search” for it, sending it to wherever it has access to.

        That’s not how LLMs work, and you know it. A model of weights is not a lossless compression algorithm.

        Also, if you’re giving an LLM free reign to all of your session tokens and security passwords, that’s on you.

        source

        • -> View More Comments
      • teawrecks@sopuli.xyz ⁨5⁩ ⁨weeks⁩ ago

        Those security concerns seem completely unrelated to the model, though. You can have a completely open source model that fits all those requirements, and still give it too much unfettered access to important resources with no way of actually knowing what it will do until it tries.

        source

        • -> View More Comments
  • Crotaro@beehaw.org ⁨5⁩ ⁨weeks⁩ ago

    Thanks for the explanation. I don’t understand enough about large language models to give a valuable judgement on this whole Deepseek happening from a technical standpoint. I think it’s excellent to have competition on the market and it feels that the US’ whole “But they’re spying on you and being a national security risk” is a hypocritical outcry when Facebook, OpenAI and the like still exist.

    What do you think about Deepseek? If I understood correctly, it’s being trained on the output of other LLMs, which makes it much more cheap but, to me it seems, also even less trustworthy because now all the actual human training data is missing and instead it’s a bunch of hallucinations, lies and (hopefully more often than not) correctly guessed answers to questions made by humans.

    source

    • jarfil@beehaw.org ⁨5⁩ ⁨weeks⁩ ago

      There are several parts to the “spying” risk:

      • Sending private data to a third part server for the model ton process it… well, you just sent it, game over. Use local models, or machines (hopefully) under your control, or ones you trust (AWS? Azure? GCP?.. maybe).

      • All LMM models are a black box, the only way to make an educated guess about their risk, is to compare the training data and procedure, to the evaluation data of the final model. There is still a risk of hallucinations and deceival, but it can be quantified to some degree.

      • DeepSeek uses a “Mixture of Experts” approach to reduce computational load… which is great, as long as you trust the “Experts” they use. Since the LLM that was released for free, is still a black box, and there is no way to verify which “Experts” were used to train it, there is also no way to know whether some of those “Experts” might or might not be trained to behave in a malicious way under some specific circumstances. It can as easily be a Troyan Horse with little chance of getting detected until it’s too late.

      it’s being trained on the output of other LLMs, which makes it much more cheap but, to me it seems, also even less trustworthy

      The feedback degradation of an LLM happens when it gets fed its own output as part of the training data. We don’t exactly know what training data was used for DeepSeek, but as long as it was generated by some different LLM, there would be little risk of a feedback reinforcement loop.

      Generally speaking, I would run the DeepSeek LLM in an isolated environment, but not trust it to be integrated in any sort of non-sandboxed agent. The downloadable smartphone app, is “safe” as long as you restrict the hell out of it, don’t let it access anything on its own, and don’t feed it anything remotely sensitive.

      source

      • Crotaro@beehaw.org ⁨4⁩ ⁨weeks⁩ ago

        Thank you a lot for the load of information! I just now got to reading it all. I was very skeptical about the fact that it is fed by the output of other LLMs but the way you explain it makes sense to me that it might not be that much of a problem. I guess a super blunt analogy could be “It’s only incest if it’s your children” lol

        source