@karpathy@N8Programs Another really awesome resource for neural implementations of simple algorithms is "The Art of Transformer Programming" (https://t.co/EgRKFGautv). It has a couple of really cool examples manually setting weights for addition, search and sort.
Our generative UI project is finally launching today! Here is a Japanese visual novel it built from my image:
https://t.co/JIJBCjRFFK
It should roll out within the next few days as a feature in Gemini called "dynamic view"
@actualrealyorth@waxhn@fly51fly It shouldn't matter as it is all symmetrical. Whichever head you choose, during training the transformer would learn to use for deciding which tokens to mask.
@waxhn@fly51fly It also has the advantage of adding no new parameters to the transformer (since a head is re-used).
You don't have to assume that the relationship is the same across heads. During training with selective attention the transformer learns to use head0 to decide what tokens to mask
@waxhn@fly51fly We need to have some score matrix to use for masking. We tried several different variants (including averaging across heads, and using a completely separate set of parameters for generating this matrix - see appendix A.2), but eventually selecting an arbitrary head worked best.
@danielbigham@karpathy We call this "alpha" in our paper (https://t.co/aKbCqEJCqu). It depends of course on the specific tasks and models, you can see various values in tables 2 and 3. For many cases (e.g. t5xxl vs t5s) we saw values around 60%-75%. This can be much higher for some easier tasks.
@dvirsky אנחנו מבקרים כל שנה, וכן יש לנו תקציב שאנחנו שמים בצד כל שנה. עד עכשיו הצלחנו קצת לחסוך כי הנסיעות האלה היו נסיעות עבודה מבחינתי, אבל עדיין היינו צריכים 4 כרטיסי טיסה לשאר המשפחה. את השהות עשינו אצל ההורים.