Matteo Pellegrini

@[email protected]

[Matteo Pellegrini @pmatteo1998]

@DavidLiptonWI @Bluebearmonkey @Tyeankk467

[Matteo Pellegrini @pmatteo1998]

@DavidLiptonWI @Bluebearmonkey @Tyeankk467 I don't think that you're getting what "sphere of influence" actually means
https://t.co/WRYGhRfG2J

[Matteo Pellegrini @pmatteo1998]

@DavidLiptonWI @Bluebearmonkey @Tyeankk467 That is completely obvious because not being in the US sphere of influence and not wanting to have colonies (unlike the us ) it's forced to produce even the lower end goods like food . on the contrary all the others Asian countries don't have such problems.

[antirez @antirez]

Btw I believe we have a mostly wrong framing of what could be done in Europe. Italy's Leonardo supercomputer datacenter alone plus Swiss National Supercomputing Centre has more than enough compute to train a very large LLM. It's not something impossible, also there is not magic recipe: it's just scaling, every smart team with the GPUs is doing it. People that fatally believe it is not something within reach are wrong.

[Matteo Pellegrini @pmatteo1998]

@Bluebearmonkey @zephyr_z9 Chinese have better/cheaper inference, better caching and low cost TPUs and a strong FOSS ecosystem (see DwarfStar etc.)
https://t.co/E4oJu3oln6

[Matteo Pellegrini @pmatteo1998]

@f_baldi_ Il problema del nucleare privato è che ha molta più concentrazione di mercato (dato costi e certificazioni) e molta meno scala delle batterie, soprattutto quelle grid scale che stanno venendo sviluppate e su cui stiamo investendo 0 in termini di R&D.

[Matteo Pellegrini @pmatteo1998]

@f_baldi_ No il nucleare deve competere con altre alternative a cominciare dalle batterie.

[Matteo Pellegrini @pmatteo1998]

@Leonardobecchet Si ma lo stato non offre meccanismi per fare bidding sugli spazi questo significa concentrazione del mercato e grandi profitti per le compagnie energetiche

[Matteo Pellegrini @pmatteo1998]

@mujhehatao @teortaxesTex No i'm Just trying to model Google's thinking, I don't have a fav in this race.

[Matteo Pellegrini @pmatteo1998]

@teortaxesTex Because in their mind are playing some long term game I think that they are handling this kind of things differently. And that reconnects to my assertion that they're trying to win the economics game \w OAI. 2/2

[Matteo Pellegrini @pmatteo1998]

@teortaxesTex > if they were comfortably ahead they'd have no problem releasing 3.1-final or 3.2-preview
idk if fixing RL is a priority for Google management and if iteration speed is a thing achievable within Google.
1/2

[Matteo Pellegrini @pmatteo1998]

@teortaxesTex I'm looking at the base model, not current scores (there's a clear structural time gap). Both xhigh 3.1Pro act as an upper bound. Google solved base efficiency—just look at 3.0 vs 3.1 on CritPt: performance jumped, but token usage barely moved. The RL is their only missing piece.

[Matteo Pellegrini @pmatteo1998]

@teortaxesTex Idk we'll see

[Matteo Pellegrini @pmatteo1998]

@teortaxesTex Matching Pro against medium is basically betting they don't have comparable RL. If you look at xhigh, 3.1 Pro is already more token-efficient. Plus, 3.1 Flash hasn't even dropped yet btw

[Matteo Pellegrini @pmatteo1998]

@teortaxesTex 1) 3.1 pro is already more token efficent
2) RL is cheaper for Google + better envs + genie (hopefully)
2.1) token efficiency is (primarily) a metter of rl
3) new tpus are cheaper + energy efficiency (imagine the politics \w the upcoming energy crisis)

[Matteo Pellegrini @pmatteo1998]

@gianlucac1bis @giovamartinelli Quasi
https://t.co/eisFYHjBfd

[Matteo Pellegrini @pmatteo1998]

@Ander_Bruckes @emmevilla Fonte:
https://t.co/X76NpVUy9i

[Matteo Pellegrini @pmatteo1998]

@Ander_Bruckes @emmevilla No è incompleto qua perché
1) c'è stata una distruzione della domanda di gas
2) assume che le importazioni LNG vengano scambiate 100% su TTF.

è chiaro che se hai una così imponente distruzione della domanda i prezzi scendono ma non ci sta niente di cui rallegrarsi.

[Matteo Pellegrini @pmatteo1998]

@teortaxesTex AOC quasi-commie?