You just need to pick your hyperparameters (random seed) right

Batman breaks a gun

I think the problem is on the consumption side. It’s hard for me to opt for a blog post over ultra-low activation energy doomscrolling, even though I know I’ll enjoy it a lot more. Don’t think I’m alone here, unfortunately.

Training a MoE router over fast food chains

Huh oh well. It’s probably easy to make more right?

No problem. Btw my random seed is a uint1e9

Coke Zero just tastes like standard Coke (derogatory)

I think people find the results intuitively implausible because, after all, the most expensive places tend to have been developed a lot. And it’s probably true that rent in manhattan would be lower if we hadn’t put a city there.

Allow me to explain:

Another disappointment. Maybe next year.

Happy Ides of March to all who celebrate! 🎉 🎈🥳🔪👑🎊

if I’m not disciplined about time management they can actually be productivity negative, by allowing me to do marginal exploratory work quickly enough to seem worthwhile, but not quickly enough to actually be worthwhile.

Trump governing to the left of Biden on climate policy

I’m in week three of debugging why my reward is not climbing with a new base checkpoint and would like to please be added to this list

Little guy would burn up

LLMs can’t be conscious because it’s too hot inside an H100 for the homunculus to survive

The synthesis here is that everyone who’s sure they know the answer, credentials or not, is a crank

Karl Friston devastated

How do you know that

Well obviously I wouldn’t write a wrong regex, and yet it’s not working, hence non-determinism

If I say a bat is a bird, and you say it’s not, you may be right, but I think it’s pretty reasonable for me to ask you to define “bird”.

Worse, they’re green now.

I feel like arxiv.org/pdf/2006.10726 maybe works for a similar reason

There are enough vaguely similar results that stabilizing first & second order activation statistics happens in the brain during sensory adaptation, and improves test time activation in neural networks, that this working doesn’t surprise me. Which is not to say I understand it

You should call your mom more

It doesn’t matter where you think the battle lines should be drawn. Our only real choice as political actors is to pick a side.

Compost bin

I’ve invented a machine that eats your vegetables for you, making kids everywhere more free

You’d probably still want to RL for thinking at least, but I bet you could avoid needing to train a good RM if you could generate per-sample high quality responses to check against.

Bar soap can already do all of these things

50 lb bag of rice for $35 dollars

Seriously. Who has $35 to spend every time they want a bowl of rice?

Elon who?