Most active commenters

    ←back to thread

    Google Titans architecture, helping AI have long-term memory

    (research.google)
    584 points Alifatisk | 13 comments | 07 Dec 25 12:23 UTC | HN request time: 0.653s | source | bottom
    Show context
    okdood64 ◴[07 Dec 25 14:05 UTC] No.46181759[source]▶
    From the blog:

    https://arxiv.org/abs/2501.00663

    https://arxiv.org/pdf/2504.13173

    Is there any other company that's openly publishing their research on AI at this level? Google should get a lot of credit for this.

    replies(12): >>46181829 #>>46182057 #>>46182168 #>>46182358 #>>46182633 #>>46183087 #>>46183462 #>>46183546 #>>46183827 #>>46184875 #>>46186114 #>>46189989 #
    mapmeld ◴[07 Dec 25 15:02 UTC] No.46182168[source]▶
    Well it's cool that they released a paper, but at this point it's been 11 months and you can't download a Titans-architecture model code or weights anywhere. That would put a lot of companies up ahead of them (Meta's Llama, Qwen, DeepSeek). Closest you can get is an unofficial implementation of the paper https://github.com/lucidrains/titans-pytorch
    replies(7): >>46182351 #>>46182946 #>>46184154 #>>46185017 #>>46186942 #>>46187280 #>>46188385 #
    1. alyxya ◴[07 Dec 25 16:34 UTC] No.46182946[source]▶
    The hardest part about making a new architecture is that even if it is just better than transformers in every way, it’s very difficult to both prove a significant improvement at scale and gain traction. Until google puts in a lot of resources into training a scaled up version of this architecture, I believe there’s plenty of low hanging fruit with improving existing architectures such that it’ll always take the back seat.
    replies(5): >>46183227 #>>46184404 #>>46184696 #>>46186138 #>>46186853 #
    2. UltraSane ◴[07 Dec 25 17:11 UTC] No.46183227[source]▶
    Yes. The path dependence for current attention based LLMs is enormous.
    replies(1): >>46184174 #
    3. patapong ◴[07 Dec 25 19:10 UTC] No.46184174[source]▶
    At the same time, there is now a ton of data for training models to act as useful assistants, and benchmarks to compare different assistant models. The wide availability and ease of obtaining new RLHF training data will make it more feasible to build models on new architectures I think.
    4. p1esk ◴[07 Dec 25 19:36 UTC] No.46184404[source]▶
    Until google puts in a lot of resources into training a scaled up version of this architecture

    If Google is not willing to scale it up, then why would anyone else?

    replies(1): >>46187379 #
    5. tyre ◴[07 Dec 25 20:11 UTC] No.46184696[source]▶
    Google is large enough, well-funded enough, and the opportunity is great enough to run experiments.

    You don't necessarily have to prove it out on large foundation models first. Can it beat out a 32b parameter model, for example?

    replies(1): >>46185008 #
    6. swatcoder ◴[07 Dec 25 20:48 UTC] No.46185008[source]▶
    Do you think there might be an approval process to navigate when experiments costs might run seven or eight digits and months of reserved resources?

    While they do have lots of money and many people, they don't have infinite money and specifically only have so much hot infrastructure to spread around. You'd expect they have to gradually build up the case that a large scale experiment is likely enough to yield a big enough advantage over what's already claiming those resources.

    replies(2): >>46189610 #>>46191181 #
    7. nickpsecurity ◴[07 Dec 25 22:52 UTC] No.46186138[source]▶
    But, it's companies like Google that made tools like Jax and TPU's saying we can throw together models with cheap, easy scaling. Their paper's math is probably harder to put together than an alpha-level prototype which they need anyway.

    So, I think they could default on doing it for small demonstrators.

    8. m101 ◴[08 Dec 25 00:15 UTC] No.46186853[source]▶
    Prove it beats models of different architectures trained under identical limited resources?
    9. 8note ◴[08 Dec 25 01:44 UTC] No.46187379[source]▶
    chatgpt is an example on why.
    replies(1): >>46193265 #
    10. dpe82 ◴[08 Dec 25 08:01 UTC] No.46189610{3}[source]▶
    I would imagine they do not want their researchers unnecessarily wasting time fighting for resources - within reason. And at Google, "within reason" can be pretty big.
    replies(1): >>46190731 #
    11. howdareme ◴[08 Dec 25 10:34 UTC] No.46190731{4}[source]▶
    I mean looking antigravity, jules & gemini cli, they have have no problem with their developers fighting for resources
    12. nl ◴[08 Dec 25 11:45 UTC] No.46191181{3}[source]▶
    I mean you'd think so, but...

    > In fact, the UL2 20B model (at Google) was trained by leaving the job running accidentally for a month.

    https://www.yitay.net/blog/training-great-llms-entirely-from...

    13. falcor84 ◴[08 Dec 25 15:23 UTC] No.46193265{3}[source]▶
    You think that this might be another ChatGPT/Docker/Hadoop case, where Google comes up with the technology but doesn't care to productize it?