Mathéo LD

/goal push the rock up the mountain

[Mathéo LD @matheold]

@han_junlin @amilabs congrats, what excites you the most?

[Mathéo LD @matheold]

@wjqkoko good luck

[Mathéo LD @matheold]

or in a less romantic way: forward deployed engineer

[Mathéo LD @matheold]

mathematicians' job will turn into being prophets

[Mathéo LD @matheold]

understand the messages of "god" and spread them in a n understandable/usable way to the population

[Mathéo LD @matheold]

@maharshii finding good non euclidian optimizers

[Mathéo LD @matheold]

@jbhuang0604 maybe one on the evolution of attention & attention algorithms ? (flash attention, MLA, Residual, etc...)

Or maybe on building a good dataset (huggingface guide is quite good for fineweb)

Or on harness design / autoresearch / RSI

[Mathéo LD @matheold]

translation: not giving me money is a terrible allocation of resources

[morgan — @morqon]

sam: local inference is a terrible allocation of scarce resources

“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre

but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time

you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment

somehow many people's intuition is the opposite of that”

[Mathéo LD @matheold]

@zacharynado I'm sure you've bookmarked this tweet for today

[Mathéo LD @matheold]

@MustafaShukor1 Hey, interested! Send a DM

[Mathéo LD @matheold]

https://t.co/KD0bXTmLWY

[Mathéo LD @matheold]

Day 4: Implemented PPO from scratch. It was a PITA!

A lot of parameter tuning, and the hardest part is handling episode termination / truncation because of how gymnasium handles it.

Fortunately Fable 5 was able to debug me

[Mathéo LD @matheold]

Today we're implementing PPO from scratch

https://t.co/mvu1bOM7JU

[Mathéo LD @matheold]

https://t.co/ZbCU6Ywrel

[Mathéo LD @matheold]

If you're interested in my Deep RL submissions: https://t.co/5DxhE9OwJw

[Mathéo LD @matheold]

Day 2/? of learning world models and RL

In the morning finished David Silver's course on RL. It's not that good to be honest because it doesn't force into really trying the algorithms for different problems so you don't really build the intuition of what algos are best for what scenario and I kept confusing different algorithms.

e.g. TD(1) refers to TD(\lambda=1) and not TD(n=1) fucking notation.

Then I started the @huggingface Deep RL course by @ThomasSimonini which is way better in contrast because it follows the fastai learning philosophy which I found really effective as long as you are curious to learn the theory by yourself (having the theory from Silver helps). Finished the first 4 Units and the optional one on @OptunaAutoML for hyperparameter tuning which is a great tool (just the dashboard feels like bootstrap in 2010).

[Mathéo LD @matheold]

https://t.co/SHffgLIeVU

[Mathéo LD @matheold]

https://t.co/KhQNHdKSnu

[Mathéo LD @matheold]

Day 1/? of learning world models and RL

Today I finished the first part of David Silver's course on RL (MDPs, solving with DP, prediction and control) tested on the test and finished the Easy21 assignment (a different blackjack).

See you tomorrow at 8 AM CET

[Mathéo LD @matheold]

https://t.co/6K0rA4H3MF