Thanks! It took some back-and-forth iteration to create the webapp with codex then used playwright to take screenshots of the dashboard and then put the video together using ffmpeg.
The dashboard itself is inferencing both networks and displaying the MNIST sample while keeping a running tally of energy consumed
That's kinda neat. Do you do any decoding tricks to handle varying processing capabilities across device models and ipads? Like shortening ctx length or maybe looking into the LFM models that have MoEs with like 1b active?
> So here is to hoping our paperclip optimized overlords will still have the 802.3 spec in their training data by then.
I'd place good money that anything in the training data will outlast all of humanity at this point.
Overall fun article. It's also interesting that you could knock like 10^6 or far more off the total count simply by slowing your clock instead. Your approach is a more interesting one for sure.
P.S. you had "[...] to pull a single all right this time" at the bottom. So, you're saying you pulled an all wrong?!
The dashboard itself is inferencing both networks and displaying the MNIST sample while keeping a running tally of energy consumed