Most active commenters

    ←back to thread

    Google Titans architecture, helping AI have long-term memory

    (research.google)
    584 points Alifatisk | 11 comments | 07 Dec 25 12:23 UTC | HN request time: 0.416s | source | bottom
    Show context
    kgeist ◴[07 Dec 25 14:57 UTC] No.46182122[source]▶
    >The model uses this internal error signal (the gradient) as a mathematical equivalent of saying, "This is unexpected and important!" This allows the Titans architecture to selectively update its long-term memory only with the most novel and context-breaking information

    So one can break a model by consistently feeding it with random, highly improbable junk? Everything would be registered as a surprise and get stored, impacting future interactions

    replies(6): >>46182150 #>>46182410 #>>46182651 #>>46183200 #>>46183413 #>>46193429 #
    1. bethekidyouwant ◴[07 Dec 25 16:00 UTC] No.46182651[source]▶
    In what world can you not always break the response of an AI by feeding it a bunch of random junk?
    replies(3): >>46182745 #>>46182845 #>>46186503 #
    2. CooCooCaCha ◴[07 Dec 25 16:12 UTC] No.46182745[source]▶
    I mean ideally AI would be resilient to junk, don't you think?
    replies(2): >>46182820 #>>46184144 #
    3. vlovich123 ◴[07 Dec 25 16:19 UTC] No.46182820[source]▶
    Humans are pretty vulnerable to junk so I’m not sure.
    4. kgeist ◴[07 Dec 25 16:22 UTC] No.46182845[source]▶
    I mean, currently LLMs are stateless and you can get rid of all the poisoned data by just starting a new conversation (context). And OP introduces "long-term memory" where junk will accumulate with time
    replies(2): >>46182912 #>>46184405 #
    5. dmix ◴[07 Dec 25 16:31 UTC] No.46182912[source]▶
    In something like Cursor if it messes something up your can click 'undo'. I'd imagine a small snapshot would only persisted to the memory if you keep it's output and even then it's mostly just a summary.

    There's probably lots of small signals of "the user is happy with the output" plus the longer the history the more it will converge on the middle of being what you want. Including when the user says "don't do [x]" which override past stuff.

    6. amarant ◴[07 Dec 25 19:05 UTC] No.46184144[source]▶
    Ideally, you'd run your own instance of this, I think.

    I can see a product where you purchase a model that has basic training, and then, using the features outlined in the paper, it learns on the fly from your usage.

    I can also see there being a secondary market for specially trained models, long-term memory filled with some specific skill, done in some specific way. To make a silly example, imagine buying a licence to Torvald's OS coding assistant, ready to insult your prs before you even commit them!(And possibly help you write code in Torvald's style too)

    This would of course require Linus to use the model enough for it to learn,I won't comment on the likelihood of that happening: it's just a silly example after all

    7. soerxpso ◴[07 Dec 25 19:36 UTC] No.46184405[source]▶
    I believe you're misunderstanding what the OP means about "long-term" memory. From what I can tell, it's not actively modifying the weights of the underlying model, it just "remembers" things from a high number of tokens into the past of its context. The point is that this allows it to remember something it read ~200 pages ago in a very long context window, not that it can remember something from one session into another clean session.
    replies(1): >>46186798 #
    8. xnx ◴[07 Dec 25 23:31 UTC] No.46186503[source]▶
    Indeed. In what world can you not break any tool when deliberately misusing it?
    replies(1): >>46196667 #
    9. AlexCoventry ◴[08 Dec 25 00:09 UTC] No.46186798{3}[source]▶
    This model has fast weights, which actually are modified during inference.
    replies(1): >>46187397 #
    10. energy123 ◴[08 Dec 25 01:48 UTC] No.46187397{4}[source]▶
    Marketplace for fast weights inbound
    11. lacoolj ◴[08 Dec 25 19:42 UTC] No.46196667[source]▶
    BRB getting an anvil