Thiago Lira

ML Research Scientist @ Articul8
interested in NLP

[Thiago Lira @L_Thiago]

@nicolasenne @TheAbsol Show me the Ledyba

[Thiago Lira @L_Thiago]

GPT 5.6-Sol is incredible at inspecting N64 MIPS decompiled code! By using N64Recomp I was able to add widescreen, 240fps and AA support for a 1997 game

https://t.co/UCx18RMeS0

[Thiago Lira @L_Thiago]

@agpatriota https://t.co/FJbxtvS4MO Esse é o último paper que eu li sobre o tema. Bem interessante. Parece que já tiveram várias respostas discordando dele

[Thiago Lira @L_Thiago]

@agpatriota Desde ontem no anúncio deles, eles estão fazendo uma live por dia com alguma coisa nova na linha de produtos

[Thiago Lira @L_Thiago]

@yacineMTB I used it everyday for transcribing japanese text from manga for studying and out of nowhere it became useless at this task, it still transcribes but it is terrible now

[Thiago Lira @L_Thiago]

And Dorothea:

[...] four aluminum towers
rise from its walls flanking seven gates with spring-
operated drawbridges that span the moat whose
water feeds four green canals which cross the city,
dividing it into nine quarters, each with three
hundred houses [...]

[Thiago Lira @L_Thiago]

Using @runwayml Gen-3 model to generate aerial shots of fantastical cities from Italo Calvino's Invisible Cities!

Here is Diomira:

[...] a city with sixty silver
domes, bronze statues of all the gods, streets paved
with lead, a crystal theater [...]

[Thiago Lira @L_Thiago]

@itsandrewgao @WilliamGao1729 @akashnet We could probably bootstrap some couple million more tokens with the ASR model and youtube

[Thiago Lira @L_Thiago]

@itsandrewgao @WilliamGao1729 @akashnet Awesome! I did a similar project some months ago (even had the same idea for a name haha)

https://t.co/HA4PFeI62A

I trained this ASR model with latin subtitles + audio from Latin youtubers! Let me know if you want to colaborate on something latin related!

[Thiago Lira @L_Thiago]

@agpatriota Eu uso uma RTX 3090, e um Ryzen 5700x e 32GB de RAM

Acho que a 3090 é o melhor "custo benefício" é caro mas é difícil arranjar 24GB de memória de vídeo por esse preço com outra configuração. Muitos clusters caseiros de Deep Learning são 2 ou 4 3090s

[Thiago Lira @L_Thiago]

@srush_nlp Thank you! I always have fun with your puzzles

[Thiago Lira @L_Thiago]

@EMostaque Incredible!!

[Thiago Lira @L_Thiago]

@EMostaque The prompt:

Al di là di sei fiumi e tre catene di montagne sorge
Zora, città che chi l’ha vista una volta non può piú di-
menticare

[Thiago Lira @L_Thiago]

@EMostaque Good work btw! Looking forward to testing the new model!

[Thiago Lira @L_Thiago]

@EMostaque I'm in the waitlist for SD3 👀👀👀👀

[Thiago Lira @L_Thiago]

@EMostaque I wish I could use the prompts in the original italian, but one thing at a time

[Thiago Lira @L_Thiago]

@EMostaque I always test new text2img models with a random prompt from some description of a city in Invisible Cities (which is my favorite book)

[Thiago Lira @L_Thiago]

@francoisfleuret Im just set warm-up steps to zero so that it starts hot

[Thiago Lira @L_Thiago]

@maxschr90 @huggingface @UpdatingOnRome Soon! (Roma aeterna est)

[Thiago Lira @L_Thiago]

Presenting Cicero! The first model on @huggingface that can transcribe Classical Latin audio! It was trained on 10h of videos of classical Latin from the latinist community on Youtube https://t.co/dGMWWdjB7d

[Thiago Lira @L_Thiago]

The idea is to transcribe lots of videos and make a Latin dataset (that is not just old books) to fine-tune and make an LLM that can talk about modern stuff in Latin too!