Recurrent Looped Transformer

(yifanzhang-pro.github.io)

14 points | by MayCXC 3 hours ago

4 comments

  • vatsachak 1 hour ago

    Paper written by AI

    • dankai 2 hours ago

      Is this just a theory about an architecture or are there actually some benchmarks/results to substantiate it?

      • jal278 2 hours ago

        appears it's just a theory, sort of surprised this lifted up so far on hn

        • fc417fc802 1 hour ago

          A fairly obvious theory at that, unless I've critically misunderstood what's being described. As with so many obvious ideas I've always assumed that the reason I haven't come across it in the wild is because it doesn't work (or is comparatively inefficient, or tends to blow up during the training run, or etc).

          > Realized reasoning gains, hardware efficiency, and RL scaling remain to be established.

          Yeah so the first entry in that list is - if I may be so bold - typically what you'd start with at a small scale _before_ writing up and publishing your "genius" idea. This is the usual crank with delusions of grandeur presenting something straightforward that he hasn't tested as though it were a working breakthrough.

          Ironically the cost of testing such theories has fallen to an all time low given the capabilities of coding models. I wouldn't be surprised if a frontier model could one shot a test of this.

      • jal278 2 hours ago

        not sure w/o context why this is important -- no results/implementation & i believe there are prior combinations of transformers/RNNs. but perhaps I'm missing the relevance/insight

        • lstodd 1 hour ago
          • fc417fc802 1 hour ago

            I can't believe that actually works? This is far more interesting than the half baked ML idea.

            Edit: Thanks to your link I've now learned about the zigzag transformer which is yet another design that somehow feels like cheating reality. https://en.wikipedia.org/wiki/Zigzag_transformer