Burning through my Anthropic credits while testing Claude Code...

Tested it out on an IRL Github frontend issue. Seems to work pretty well (but it's really expensive)

Feels like Sora is having the same moment that Gemini had after release. Very technically impressive, but Sora hasn't delivered on the expectations/hype.

It only seems to work well for a few types of clips (ex. news, landscape shots). Physics/interactions between objects are notably low-quality.

this ai startup boom is showing that silicon valley (the show) has aged really well.

life imitates art.

the science of LMs should be fully open✨

today @akshitab.bsky.social @natolambert.bsky.social and I are giving our #neurips2024 tutorial on language model development.

everything from data, training, adaptation. published or not, no secrets 🫡

tues, 12/10, 9:30am PT ☕️

neurips.cc/virtual/2024...

on today's episode of weird math i learned: you can view complex numbers as (a+bi) in C, or as [a,b] in R2. so you can see functions as either a map from C->C or from R2->R2.

the weird part: your function can be differentiable when viewing it as R2->R2, but not when viewing it as C->C.

It's hard to communicate our experiences and ideas through words. There are so many layers between what I think, what I say, and what you understand.

"Just because we’re using the same words doesn’t mean we mean the same things.... But let's try anyway."

nathanieldrew.substack.com/p/how-to-hav...

tip-of-the-tongue search is an awesome use case for LLMs, and web search makes it way more powerful.

now that LLMs have been in our hands for ~2 years, i feel like it takes a lot to impress me. i'm kind of surprised it took multiple prompts to get this right.

What is the silliest scenario you can think of where Bayes' rule provides a decent probability estimate?

SVD is one of the most beautiful parts of linear algebra:

1. View a matrix as a sum of rank-1 matrices
2. Read orthonormal bases for the 4 fundamental subspaces
3. Get a low-rank approximation, etc.

Sadly higher order tensors are much harder to factorize for NN weight compression...

Our human brains are so energy-efficient compared to today's AI training/inference. To be fair, we've also had billions of years for intelligence to evolve.

Regardless, I wonder if neural networks and digital computing (as opposed to analog) are the wrong approach to intelligence.

TIL: Concurrent programming in Rust can be sketchy—there's no defined memory model.

Suppose you run a program on 2 machines with different ISAs. Outputs obtained on one machine may be *impossible* on the other. There are no guarantees about the order of memory reads/writes.

Felt the itch to build something over the past week or so.

Decided to start working on my own RSS feed reader that curates your content - no complicated algorithms!

github.com/TanayB11/pap...

i think we’re conditioned to care about scaling our impact. it's the whole point of EA/utilitarianism.

in theory you should get more fulfillment from this, but i feel like you get more fulfillment from changing *someones* world instead of changing *the* world.

stevenscrawls.com/care-doesnt-...

did some competitive programming today. it's amazing how much you can enjoy doing difficult things after you stop caring about the outcome.

"I am always sincere, but never serious." - Alan Watts