• 13 Posts
  • 1.88K Comments
Joined 2 years ago
Cake day: July 7th, 2023
  • just_another_person@lemmy.worldtoTechnology@lemmy.world•Meta’s star AI scientist Yann LeCun plans to leave for own startupEnglish
    1·
    5 minutes ago

    From your own linked paper:

    To design a neural long-term memory module, we need a model that can encode the abstraction of the past history into its parameters. An example of this can be LLMs that are shown to be memorizing their training data [98, 96, 61]. Therefore, a simple idea is to train a neural network and expect it to memorize its training data. Memorization, however, has almost always been known as an undesirable phenomena in neural networks as it limits the model generalization [7], causes privacy concerns [98], and so results in poor performance at test time. Moreover, the memorization of the training data might not be helpful at test time, in which the data might be out-of-distribution. We argue that, we need an online meta-model that learns how to memorize/forget the data at test time. In this setup, the model is learning a function that is capable of memorization, but it is not overfitting to the training data, resulting in a better generalization at test time.

    Literally what I just said. This is specifically addressing the problem I mentioned, and goes on further to exacting specificity on why it does not exist in production tools for the general public (it’ll never make money, and it’s slow, honestly). In fact, there is a minor argument later on that developing a separate supporting system negates even referring to the outcome as an LLM, and the supported referenced papers linked at the bottom dig even deeper into the exact thing I mentioned on the limitations of said models used in this way.

  • just_another_person@lemmy.worldtoTechnology@lemmy.world•Meta’s star AI scientist Yann LeCun plans to leave for own startupEnglish
    2·
    32 minutes ago

    It most certainly did not…because it can’t.

    You find me a model that can take multiple disparate pieces of information and combine them into a new idea not fed with a pre-selected pattern, and I’ll eat my hat. The very basis of how these models operates is in complete opposition of you thinking it can spontaneously have a new and novel idea. New…that’s what novel means.

    I can pointlessly link you to papers, blogs from researchers explaining, or just asking one of these things for yourself, but you’re not going to listen, which is on you for intentionally deciding to remain ignorant to how they function.

    Here’s Terrence Kim describing how they set it up using GRPO: https://www.terrencekim.net/2025/10/scaling-llms-for-next-generation-single.html

    And then another researcher describing what actually took place: https://joshuaberkowitz.us/blog/news-1/googles-cell2sentence-c2s-scale-27b-ai-is-accelerating-cancer-therapy-discovery-1498

    So you can obviously see…not novel ideation. They fed it a bunch of trained data, and it correctly used the different pattern alignment to say “If it works this way otherwise, it should work this way with this example.”

    Sure, it’s not something humans had gotten to get, but that’s the entire point of the tool. Good for the progress, certainly, but that’s it’s job. It didn’t come up with some new idea about anything because it works from the data it’s given, and the logic boundaries of the tasks it’s set to run. It’s not doing anything super special here, just very efficiently.

  • just_another_person@lemmy.worldtoTechnology@lemmy.world•Meta’s star AI scientist Yann LeCun plans to leave for own startupEnglish
    3·
    2 hours ago

    🤦🤦🤦 No…it really isn’t:

    Teams at Yale are now exploring the mechanism uncovered here and testing additional AI-generated predictions in other immune contexts.

    Not only is there no validation, they have only begun even looking at it.

    Again: LLMs can’t make novel ideas. This is PR, and because you’re unfamiliar with how any of it works, you assume MAGIC.

    Like every other bullshit PR release of it’s kind, this is simply a model being fed a ton of data and running through thousands of millions of iterative segments testing outcomes of various combinations of things that would take humans years to do. It’s not that it is intelligent or making “discoveries”, it’s just moving really fast.

    You feed it 102 combinations of amino acids, and it’s eventually going to find new chains needed for protein folding. The thing you’re missing there is:

    1. all the logic programmed by humans
    2. The data collected and sanitized by humans
    3. The task groups set by humans
    4. The output validated by humans

    It’s a tool for moving fast though data, a.k.a. A REALLY FAST SORTING MECHANISM

    Nothing at any stage if developed, is novel output, or validated by any models, because…they can’t do that.

  • just_another_person@lemmy.worldtoTechnology@lemmy.world•Meta’s star AI scientist Yann LeCun plans to leave for own startupEnglish
    3·
    3 hours ago

    I sure do. Knowledge, and being in the space for a decade.

    Here’s a fun one: go ask your LLM why it can’t create novel ideas, it’ll tell you right away 🤣🤣🤣🤣

    LLMs have ZERO intentional logic that allow it to even comprehend an idea, let alone craft a new one and create relationships between others.

    I can already tell from your tone you’re mostly driven by bullshit PR hype from people like Sam Altman , and are an “AI” fanboy, so I won’t waste my time arguing with you. You’re in love with human-made logic loops and datasets, bruh. There is not now, nor was there ever, a way for any of it to become some supreme being of ideas and knowledge as you’ve been pitched. It’s super fast sorting from static data. That’s it.

    You’re drunk on Kool-Aid, kiddo.

  • just_another_person@lemmy.worldtoLinux Gaming@lemmy.ml•New to Linux Advice
    2·
    6 hours ago

    Just make it easy on yourself: start with using a Fedora LiveUSB just to try out the different desktops.

    Gnome is more like MacOS KDE is more like Windows

    I’d just try those out first until you’re more familiar with things. All distros are pretty much the same as far as features go under the hood, and you won’t see a massive gaming difference on any of them.

    Sticking with AMD hardware is the right move if you’re gaming. Steer clear of unbranded machines you can’t easily check compatibility on as you may end up with unknown hardware without proper driver compatibility. Some top gaming brands that make an effort to be Linux compatible:

    • MSI
    • Framework
    • System76
    • Dell (only the enterprise stuff)

    You’ll probably only find good deals on the MSI stuff in the upcoming sales, but maybe the Framework refurbished store online will have some desktops. System76 rarely has sales, and are a bit overpriced in IMO.

    If you build your own, just stick with MSI and AMD. You can probably build a solid gaming rig for $750-ish.

  • just_another_person@lemmy.worldtoLinux Gaming@lemmy.world•Steamworks SDK 1.63 has been released – Added libs for linuxarm64 and androidarm64English
    6·
    9 hours ago

    Yes, that’s what I’m saying. Running Android on Linux is one thing, and that only requires Waydroid and has nothing to do with Steam.

    Having an SDK stack specifically for androidarm64 means there are extensions there to hook into the Steam client, meaning Android apps that use Steam platform for…something. This was never announced or discussed.

    It has nothing to do with Frame specifically, because you don’t need an entire SDK entry point for one device that has nothing to do with Android anyway.

    So it must mean that they intend to give the option to hook in to Steam for game devs that already have Android builds and distribution or something. Like Netflix games, Rockstar, Bethesda…etc.