zuzatm

One notable difference from what one would expect from a LLM-RL paper is the use of test-time RL. I guess when you have a very strong verification, you can specialize your network to solve only your problem. Curious if this can be also be applied in natural language reasoning.

It's coming! Check again in 12 hours, I believe it should be up then!

That said in the livestream it is, starting today.

I was an intern in FB in 2010 and I distinctly remember getting this email. My full-time buddies, some of which included very senior engineers who often work with Mark, mentioned the got the shivers when they saw the title and thought they just got booted :)

I'd argue that the results of this paper are not meant to be implemented, hence no code required.

Minor comment: the improvement over the state-of-the-art is in the second order term of the exponent.

Their formula express det. with B_n = exp( n ln n - n lnln n - O(n)) terms.

The old formula had exp(n ln n - O(n)) terms.

It's a nice bound, but won't change much for non-theoretical results.

I did. It a beautiful and very educational website! :)

I don't know why, but I find watch mechanics super fascinating.

Few favorites for people with spare time (and lots of money if one wants to buy these):

Tourbillon: a watch complication where the main mechanism is mounted on a rotating wheel so that gravity doesn't affect the watch in a single direction.

Spring drive: A very recent mechanism that produces a perfect sweeping motion of the seconds hand. It combines both mechanical power (no battery) with a quartz crystal in a super cool way.

Minute repeater: a mechanical way for a blind person to tell the time.