Juan Maroñas

@sarvjeet_ghotra @artetxem @nvidia I received two from that program. 17'18'

Remember when NVIDIA was giving free GPUs to everyone in academia? They don't do that anymore, do they? I'm afraid most companies releasing open source LLMs are doing the same. The only difference is they are giving free software instead of free hardware (pun intended).

@svpino I don't agree. Many things can be implementes in many ways. Comments allow to express this way without requiring others to know what each single method from a library do. Also allow to Split Up parts of an implementation, organizing stuff.

Many people write comments because they are incapable of writing good, clean code. The vast majority of comments I read are unnecessary. They are a liability. I follow these principles: 1. Your code is not a project management system. Don't write "TODO" notes in code. 2. Good, clear code is self-explanatory. If you find yourself writing too many comments, you are doing it wrong. 3. If you need to rewrite a comment when you change the code, your comment is telling the wrong story. 4. Never write a comment to explain a code that's hard to understand. Rewrite the code instead.

@pmddomingos I prefer to live in a country where healthcare is free and the probability of my son being shot at school is non-existent

In the last 30 years, America has produced seven trillion-dollar companies, while Europe has not produced a single one worth more than $100B.

Happy to share work accepted at #ICML2023 proceedings.mlr.press/v202/maronas23… How can we create C non stationary and dependent processes in an efficient way

Spent time reviewing for #ICML2023. Author wrote changes to incorporate. Paper gets accepted with camera ready showing no changes. Won't review anymore for a conference unless I have submitted something. A-R discussion is more a mafia interchange that something for science.

@Serthand Pues yo me he operado 5 veces, entre ellas del hombro, pagando 50 euros al mes.

Siempre nos reímos de esto, pero en serio hay gente adulta por ahí que cree que pagando 50€ al mes les da para cubrir una hospitalización, una intervención o lo que sea.

@PatrickKidger @verdverm Also joining kernels avoid memory transfers and I think It uses internal multiprocessor registers in a more controlled way. Overall very nice tricks.

JAX has a JIT compiler, so that Python isn't used at runtime. This removes a lot of the overhead when manipulating many small objects. (And is part of why I really like JAX for scientific computing.) It also performs algebraic simplification to try and simplify the computation graph where possible. Specifically wrt allocation - JAX (and XLA under-the-hood) statically allocates memory in advance. But since you mention sympy - you might like github.com/google/sympy2j…, which allows you to interface sympy and JAX together! Does that answer your question?

:D🎉🎉 (From someone on HN.)

@artetxem 🤑

We live in a time where AI papers deliberately omit essential details like model size and training tokens... which end up being leaked and published by mainstream media 🤦 cnbc.com/2023/05/16/goo…

@PatrickKidger I've seen similar behavior on adaptative gradient methods when training GPs.

Double descent? Pfft, in *real industry* we get 9-fold descent! (seriously what the hell is this) (this is my actual training run)

Un periodista preguntó anoche a Giannis Antetokounmpo sobre si la temporada de los Milwaukee Bucks ha sido un fracaso o no Su respuesta es magnífica

@PasionBasketNBA Si, hacer pasos

Volvió Ja Morant y lo hizo como mejor sabe hacer 💥✈️

@PasionBasketNBA Con 400 pasos

La acrobática canasta de Jimmy Butler

@PatrickKidger @lizbwood North Madrid - Valencia.

@lizbwood Generally: pick north-south based on weather preferences; pick east-west based on timezones you have to be available for meetings. E.g. Greece is warm but EU-timezone-only. Portugal is warm and possible for US working hours. Personal favourites are London or Amsterdam.

Not a real idea yet, but for the remote ML team, I was thinking of getting us all together for a summer of working together in-person. Those with experience as #digitalnomads, what would be a good place to host an ML house in the summer for people from US and EU? US has EU ID.

@forocoches Marketing

De estas asignaturas , cuál de ellas te entra un escalofrío nada más verla forocoches.com/foro/showthrea…

Finally found a real motivation to use chatGPT beyond creating fakes stories about my hommies.

@IbaiLlanos Pues por el culo te la hinco

Cuando vaya a México quiero comer el mejor taco del país así que decidme dónde está que me comeré 25.

@RobertoParPal Hay que entender que el rol de Gary en twitter es dar x el culo jaja

No hay más ciego que el que no quiere ver. Es como cuando criticaban a los hermanos Wright y se reían de sus primeros intentos de volar con un artilugio…

@tdietterich @emilymbender And also allows updating the paper if you find any mistake or want to clarify something, which in my opinion is ita best feature

@emilymbender Arxiv is an archive, not a venue. It also requires supplemental materials to be merged into one pdf instead of separate files like the neurips web site. It is also more searchable

#NeurIPS Is doing something really wrong if everyone is tweeting links to the arXiv versions of their papers there. It's supposed to be this super prestigious conference yet no one wants their (ahem) archival version of their work to be associated with it? #NeurIPS2022 >>

Tiene cojones que la serie del inmortal cuente la vida del jefe de los miami y que lo llamen jose..

@RobertoParPal Imagínate como estará el Salvador (creo que era este pais) que movió parte de su reserva a cryptos

Nunca pongas tus huevos en la misma cesta… me parece increíble el nivel de atontolamiento que ha producido el crypto hype