15 comments

  • kamranjon5 hours ago
    There&#x27;s a really interesting trend of labs using &quot;proprietary&quot; methods to convert existing models to a compressed ternary format.<p>PrismML actually targeted the same Qwen 8b model and got it down to 1.75gb here: <a href="https:&#x2F;&#x2F;prismml.com&#x2F;news&#x2F;ternary-bonsai" rel="nofollow">https:&#x2F;&#x2F;prismml.com&#x2F;news&#x2F;ternary-bonsai</a><p>I wonder how proprietary it all is though, since the BitNet b1.58 paper has been out for a couple years now: <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2402.17764" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2402.17764</a><p>From the wikipedia on 1.58 bit llms: &quot;BitNet derives its performance from being trained natively in 1.58 bit instead of being quantized from a full-precision model after training. Still, training is an expensive process, and it would be desirable to be able to somehow convert an existing model to 1.58 bits. In 2024, HuggingFace reported a way to gradually ramp up the 1.58-bit quantization in fine-tuning an existing model down to 1.58 bits.&quot;<p>The section from huggingface is here: <a href="https:&#x2F;&#x2F;huggingface.co&#x2F;blog&#x2F;1_58_llm_extreme_quantization#fine-tuning-in-158bit" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;blog&#x2F;1_58_llm_extreme_quantization#fi...</a><p>I just wonder how many of these labs are basically following the huggingface recipe here and possibly tweaking it and releasing models without huge training costs.
    • Twirrim5 hours ago
      Independent testing of prismml suggest quite a capability drop off outside of their cherry picked benchmarks. I&#x27;ll be curious to see what this model achieves though.
      • kamranjon5 hours ago
        Unfortunately Fermion Research appears to entirely AI generate all of their content here, even for the research section: <a href="https:&#x2F;&#x2F;www.fermionresearch.com&#x2F;research&#x2F;neutrino-8b&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.fermionresearch.com&#x2F;research&#x2F;neutrino-8b&#x2F;</a><p>&quot;Neutrino-1 8B was trained natively in its shipping format. There is no full-precision product model that was rounded afterward: the ternary representation is the medium the weights learned in, and the training methods that hold this quality at this depth are the lab’s unpublished work. The findings below are the part that travels.&quot;<p>This statement seems misleading at best.<p>Both the model page and the release page are basically unintelligible - I don&#x27;t have a ton of faith in the work here, at least PrismML write coherent releases for their models.<p>Edit: Another beautiful piece of prose here, I almost wonder if they used the 8b model to generate the content for this release...<p>&quot;Across the 6.95B coded weights, 62.63% sit at zero and the remainder splits 18.68% plus to 18.69% minus: sign-balanced to a hundredth of a point with no constraint asking for it.&quot;
        • LtdJorge2 hours ago
          I guess it&#x27;s saying how many of the weights are -1, 0 or +1.
          • jdiff1 hour ago
            It is, but why? And what&#x27;s with the bizarre way of phrasing that? Why the bizarre observation that, indeed, nobody asked for it?
            • mapt8 minutes ago
              &gt; sign-balanced to a hundredth of a point with no constraint asking for it.&quot;<p>This isn&#x27;t their model, this is (probably?) ChatGPT doing a brag &#x2F; promo deck authorial voice. It routinely uses half a dozen sentence constructions that are relatively uncommon in normal or technical speech. Eccentric. Persuasive. Trying too hard. Restating its point in a promotional way that doesn&#x27;t sound natural, leading into a sentence where it hyperbolically sells you on having done the impossible.<p>Normal persuasive speech uses these constructions, especially public speaking doing a VC pitch or an Ancient Aliens, but it would be fucking strange if a person started chaining them in normal conversation and using little else, trying to insinuate competence.<p>You start to recognize it pretty quickly on Youtube.
            • rcxdude46 minutes ago
              Also, wouldn&#x27;t normalisation tend to produce a mean of zero in the weights, anyhow?
      • dofm3 hours ago
        I really had high hopes for the larger Ternary Bonsai and it feels like there is scope to improve, but I get the sense (albeit a naïve, probably not fully informed sense) that improvement can perhaps only come by training directly into ternary.
        • kamranjon2 hours ago
          I’ve actually been really impressed with the 27b model they recently released - amazing performance approaching 40 tok&#x2F;s on m4 max and I didn’t run into any quality issues in the small set of tasks I tried. Haven’t gone full coding with it yet but suspect it’s better than say a 9b or 12b model.
          • embedding-shape40 minutes ago
            &gt; suspect it’s better than say a 9b or 12b model<p>Whaaat, a 27b model might be better than 9b or 12b model? What would make you do such an outrageous claim?
            • kamranjon27 minutes ago
              Sorry I should have clarified - I meant that a ternary 27b model would outperform a non-quantized or 8 bit quantized 9 or 12b model - which it is generally close to (or much smaller than) in size. So yeah the comparison I was trying to make was between models of equivalent size or models that could run on similarly sized hardware.
              • embedding-shape3 minutes ago
                Ah yeah, that makes a ton more sense :) I mean, what you said earlier also makes sense but was too obvious, now it makes sufficient sense, thanks for explaining :)
        • avadodin2 hours ago
          All you need is Ternary Aware Training and for AI researchers to come up with a backronym for TIT.
  • moinism4 hours ago
    The content on that page is too AI-generated to make sense to me; I don&#x27;t understand what the model is for.
    • naruhodo2 hours ago
      I have the same question.<p>I get that it’s designed to run on a CPU, big GPU or MacBook (although the way that was phrased confused me at first).<p>I’m struggling with what a “decoder-only” model is good for.
    • SwellJoe4 hours ago
      AI doesn&#x27;t want anything, so it doesn&#x27;t care whether it conveys meaning in its writing. And, apparently the developers of this project also don&#x27;t care whether it conveys meaning. They just assume we&#x27;ll wade through the slop? I dunno.
      • iwontberude27 minutes ago
        It has no ability to ground output in causality. It does not understand what causes humans to learn. It’s a correlation and randomness engine.
  • secult3 hours ago
    There is not a single person mentioned on the website, github created 3 days ago, no real contact, everything hidden. Completely anonymous. Domain owner hidden.
  • Havoc3 hours ago
    Can’t say I’m a fan of containers for this. A big chunk of local LLM gains come (imo) from the open modular nature of llama.cpp and friends. Easy to modify. Easy to experiment.<p>Containers are the proprietary binary blob in hardware world equivalent
    • embedding-shape1 hour ago
      What? How are those even related with each other? You can just as easy modify and experiment with llama.cpp in a container as outside of it, they really shouldn&#x27;t impact one another. Containers don&#x27;t suddenly make llama.cpp less &quot;open modular&quot; somehow, and I&#x27;m not sure how you&#x27;d arrive as such conclusion.
  • yborg5 hours ago
    Largely outperformed by Ternary-Bonsai-8B by their own chart, doesn&#x27;t seem clear what their special sauce is here.
  • androiddrew51 minutes ago
    So I&#x27;d like to see a Nemotron3 Ultra converted to a 1.58bit format then have them retrain on the open dataset.
  • codeduck1 hour ago
    So... not a new subatomic particle discovery then.
  • sparse-Matrix1 hour ago
    Unfortunately it crashed out &#x27;no space left on device&#x27; while installing the python demo&#x2F;quickstart.<p>Only problem was there is plenty of space on the device. PLENTY (not quite 750gb).
  • NetOpWibby3 hours ago
    There’s a new announcement every other day wrt models. How do y’all keep track of them all same know what’s decent? Good grief!<p>And if it’s decent today, it’s shit in eight months! I tool hop as much as the next dev but this is a bit much.
    • weikju3 hours ago
      I just tune out. It’s not worth knowing, following every development in the field. If something works now it will probably work in 8 months even if it’s no longer the new hype thing. Who cares.<p>Not using any of it is also a valid option though it doesn’t satisfy your FOMO. But nothing ever will.
    • badatnames2 hours ago
      There&#x27;s no need to, the 50 foot view is simply that many alternatives exist and they mostly fall into 3 meaningful weight classes with comparable performance among each class&#x27;s members: too expensive to use indiscriminately, too big to run at home, and too small for complex work. As for names and faces in between, the overarching conclusion is that we&#x27;re rapidly approaching commodity status and those don&#x27;t really matter much
  • madhu_ghalame4 hours ago
    The real strength of an 8B model is efficiency. It will be interesting to see the balance between performance and inference cost.
  • bmiekre1 hour ago
    This all sounds like middle-out
  • Alien1Being3 hours ago
    AI slop site with AI slop research...<p>Blog populated with incoherent PR material generated by Yet Another AI.<p>Sigh...
    • zoom66282 hours ago
      Can Dang implement a slop rating on submitted pages? Not a block but at least a % likelihood of AI slop content and that could also be tied with a BS rating as well.<p>Could use AI for both which seems hilariously appropriate.
      • codeduck1 hour ago
        Or perhaps a slop flag on submissions that gets added once enough readers have flagged it as such. I&#x27;d love that feature.
  • drbscl2 hours ago
    Slop article, slop site... slop model?
  • runtime_lens3 hours ago
    [flagged]