And, ID, in the final analysis when you say "if it scores better it scores better" you are begging the question.
That's less than not interesting. It's a distraction. Here's another way to talk about the situation that isn't a distraction: "If it predicts better, it predicts better." And now think about what it means to "predict better". What is the essential function of a lossless compressor? On Wed, Oct 7, 2020 at 9:56 AM James Bowery <[email protected]> wrote: > How many times have I stated in this group that big language modelers > bragging about their parameters counts worse than cottoncandyfluffpuffery? > Reducing the parameter counts by a factor of 1000 and achieving comparable > performance on just about _any_ plausible language benchmark is damning of > such hypercottoncandyfluffpuffery. > > On Wed, Oct 7, 2020 at 4:08 AM <[email protected]> wrote: > >> James, if theirs is trained on enwik9 and if GPT was too, then we can >> compare. Remember, if both models reached a training level where both >> really need not eat more data, you'd see the smarter one is better no >> matter how much data the other eats. So let's say GPT scores better than >> theirs, it's ok if it has a big model, because the other can't reach its >> accuracy even if eats to get a as big model as GPT's. Big model is only a >> problem if can't run it. That requires optimization before moving on. One >> day when we reach the seemingly best AI, we can then start worrying about >> model size and speed as its the only gain left. >> >> And if I'm wrong, then my first thought on this thread is correct. >> *Artificial General Intelligence List <https://agi.topicbox.com/latest>* >> / AGI / see discussions <https://agi.topicbox.com/groups/agi> + >> participants <https://agi.topicbox.com/groups/agi/members> + delivery >> options <https://agi.topicbox.com/groups/agi/subscription> Permalink >> <https://agi.topicbox.com/groups/agi/T6761a13445e5864b-Mc32deb96ba6f64b000912289> >> ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T6761a13445e5864b-Maded5f0b68705bc8420f1524 Delivery options: https://agi.topicbox.com/groups/agi/subscription
