you're missing out a little, if you haven't read this piece by @ch402 on visual information theory.
It builds upto KL divergence and cross-entropy from the simplest of first principles and is absolutely worth your time
recently i'm seeing a bunch of bots across twitter and in my followers list
not the crypto or spammy kind but the ones that will say that they're an "agent" of some sort for some company in their bio
not too sure how i feel about these, don't know what they're even for tbh
I think code review tools like @greptile should really have more native options to do code review on more than just the current commit's diff or for a PR.
They should let you do a general audit of the entire repository as well, at any time.
subagents largely wouldn't need to exist if you had a large enough context window & fast enough generation, it seems like an engineering solution to an ml problem.
It's usually hard to bet against tech, and even more so when hundreds of billions is being poured into something.
This particular feat was of course a consequence of having reasoning in img-gen models (like images 2.0 on chat).
On an unrelated note, one of the mainstream arguments for why LLMs / img-gen models were merely stochastic parrots was that they couldn't generate anything other than 10:10 on a clock face.
To anyone familiar with these tools, it was obvious that it was only a matter of time
the band does feel shorter in width though (prolly cuz of removing the cornell logo? did they have one?) which is confusing me a little bit, cuz the buttons seem a little different placed than where I would've expected them to be last week