@DimitrisPapail Amazing, exactly what I was saying so much, why wasn't anyone studying 100% correct arithmetic modeling? This gives me trust in transformers! Did you experience an "aha" moment where it just started getting arbitrary length arithmetic?

@sama Does it mean GPT-5 will be a blanket term hiding many models, or do "intelligence levels" just mean more compute? @sama

@BlinkDL_AI It's fine, rwkv 7 is super exciting as well! Do your best, my man! Meanwhile, can you explain how **you** see it, assuming R1 has 37B active parameters per token (~37GB at Q8) and an NVMe RAID0 with 60GB/s, how can one expect more than 60/37 (~1.7) token/s?

RT @htmx_org: lisan al gaib!

@ENDESGA I like that Linus coldly put it into the kernel and forces people to be reasonable about it. As the result, some cool people made great drivers for the kernel! Also, I believe the initial Rust push was partly due to the C crowd treating Rust fans as "impure" on social media

RT @pegobry_en:

@lisyarus Thanks! Will be curious to see how the simulation formula looks like 👀

@lisyarus Cool! Really wanted to implement erosion for a From Dust kind of game ^^ Do you have a piece of code your are willing to share?

@BlinkDL_AI Lets go!

@theo Except Groq doesn't have R1, only distills, which are much more limited ;(

RT @ClementDelangue: Our science team has started working on fully reproducing and open-sourcing R1 including training data, training scrip…

@rustdesk Most orgs I worked in already flag it as "scammer software" and do not allow it on work computers.

@yaazarai Isn't it bucket sort or something like that? Or radix? Considered one of the fastest?

@TynanSylvester Wanna cooperate on a self hosted private solution? 😁

@MLSophist No, they are rock stars! We need more artisans like that, there should be no billionaires, just well paid artisans.

@bee_fumo @HSVSphere @ludwigABAP @yupiop12 That's fair

@idariusdan No one pays designers anymore