Lirus

[Lirus @Lirus_9]

@AniketVashisht8 Congratulations!

[Lirus @Lirus_9]

@o_v_shake Teaching is a two-way street: the teacher finds joy, the student finds passion or results. Only the latter is worth paying for.

[Lirus @Lirus_9]

@o_v_shake @cerebras Congratulations buddy!

[Lirus @Lirus_9]

@sidgairo18 I have been good sir!

[Lirus @Lirus_9]

@sidgairo18 @iclr_conf @MoritzBoehle @FrancescoLocat8 Congratulations 🎉

[Lirus @Lirus_9]

@AniketVashisht8 Wohooo
Congratulations 🎉

[Lirus @Lirus_9]

@Iameghnaa Congratulations Meghna 🎉

[Lirus @Lirus_9]

@arvidkahl Take a look at crawl4ai

[Lirus @Lirus_9]

@vishaal_urao Congratulations Vishaal 🎉
4 Neurips is crazy

[Lirus @Lirus_9]

@paraschopra 3/3
In hindsight, it seems like Sohl might've overfitted to his proxy and missed applying Occam's razor or some regularization. Given this is published in 2022, i think this probably played some role.

[Lirus @Lirus_9]

@paraschopra 2/3
In 2020, Ho et al. improved on Sohl's work with DDPM by relaxing some of the constraints. That made training much more stable and led to the amazing results we see now with diffusion models.

[Lirus @Lirus_9]

@paraschopra 1/3
Interesting! Sohl-Dickstein's 2015 work on diffusion models laid the foundation for a lot of what we see today. But his final formulation of variational inference approach struggled with stability in training.

[Lirus @Lirus_9]

@_sakshams_ @AIatMeta Congratulations Saksham!

[Lirus @Lirus_9]

@hsrhackerhouse AI

[Lirus @Lirus_9]

@NirantK @OpenAIDevs Sadly these probabilities are not really calibrated well 🫤. Usually you will end up with 100% probability for classification tasks.

[Lirus @Lirus_9]

@o_v_shake Come on maiti, it's high time to drop the course/book.

[Lirus @Lirus_9]

@o_v_shake I have seen that there's a lot of diversity in the ml-system design interviews.

Some resources that i feel are good to brush through.

- https://t.co/7pJir4R39I
- Exponent ml-sys design playlist
- Chip Huyen

[Lirus @Lirus_9]

@o_v_shake I thought I was good at probability until I came across 3.

[Lirus @Lirus_9]

@o_v_shake At a micro-level that seems to be case but I am not sure about the macro trends. Furthermore, I think the top tech's still usually prefer MTech's or a good pedigree.