En estos dos años, he continuado mi entrenamiento en diversos tópicos.
- La geopolítica y su economía global.
- Mi maestría en Inteligencia Artificial (con más intensidad) y la creación de empresas (tengo una hace un año)
- El Futurismo NARP & Afrofututismo.
Mi lectura de hoy:
The real world is full of surprises, like a gust of wind while driving! Intelligent world models must reason about them, and imagine what would have happened under another action... We give JEPA world models this ability with provable guarantees 🧵⬇️
w/ @KerdaVaraku and @ylecun
@DiegoASantos Bien dicho Diego. Además, en un país tan diversos como el nuestro, es una pérdida de tiempo y esfuerzo que podría ser mejor dirigido a problemas reales de Colombia.
We’ve opened two LatamGPT datasets on Hugging Face. Together they contain 178.4M documents (290.5B tokens) in Spanish, Portuguese, and English, drawing on sources such as books, news, and institutional collections from Latin America and the Caribbean.
We built this corpus with regional partners to help the AI community develop models that better represent the varieties of Spanish and Portuguese across Latin America and deepen their understanding of the events that shaped our countries, the entities people interact with, and the different ways we speak.
A university can study which varieties of Spanish are underrepresented. A startup can adapt a model to local terminology. A public agency can build an assistant that recognizes the names people actually use for institutions and services.
🟠 Open access (143.4M documents, ~242.5B tokens). For continued pretraining, domain adaptation, and corpus research. Access requires accepting the repository terms and respecting each document’s license. LatamGPT-Corpus-1.0
🔵 Controlled research access (35.0M documents, ~48.0B tokens). For scientific and academic research. Apply with a verifiable institutional affiliation and a stated research purpose. Access requires CENIA approval and a signed confidentiality agreement. Redistribution is prohibited. LatamGPT-Corpus-1.0-research
For both releases, organizations with more than US$1 billion in annual revenue or cumulative capital raised generally need prior written authorization from CENIA.
URLs:
LatamGPT-Corpus-1.0: https://t.co/rjOMTqY52b
LatamGPT-Corpus-1.0-research: https://t.co/tx7WEDo7Ww
Finished reading Mikito Chinen’s “Shinwa (Mythology).” A true page-turner that made me race through all 600 pages in one sitting. Perhaps this is the “power of the gods.” Singularity, ASI, AI Cold War, nuclear deterrence, terrorism, religion... What unfolds here is not science fiction or “mythology,” but a vividly rendered vision of an “everyday reality” that could arrive tomorrow. A multi-genre thriller infused with a “Sakyo Komatsu”-esque global-scale crisis, the late-period realism of “Koji Suzuki,” and the scientific plausibility and suspense of “Michael Crichton.”
Boom! Phenomenal Q3'26 results at Nubank last night:
The big picture:
Gross Revenue: $6.3 billion, up 37% YoY FXN.
Net Income: $1.1 billion, up 31% YoY FXN.
Profitability: ROE reached 32%.
Engagement and activity rates: ARPAC reached ~$18, with the monthly activity rate expanding to 83.6%, and reaching a new record of 86.3% in Brazil.
Total Customers: 143M globally (~3.7M net additions in the quarter).
Core markets milestones in Q3'26:
Brazil: Nearly 120MM customers, 65% of the adult population.
Mexico: 16.6MM customers, 17% of the adult population.
Colombia: 6.1M customers, 15% of the adult population.
Global expansion- It's the 1st second of the 1st half
Nu US reached customers in all 50 states of the United States!
Nu Global has customers in 37 countries!!!
Super proud of all the work of our team, and incredibly optimistic about the brilliant future we have ahead
https://t.co/mPOJSaJjvF
Recuerdo a @VickyDavilaH siendo agresiva con @ODBenavidesA desde el primer día en el que recibió donaciones para los damnificados del terremoto.
Hoy vemos al fondo milagro ser una oda a la duplicidad de responsabilidades que tanto la UNRGD y MinHacienda pudieron asumir.
Quando essere giusti e onesti avrà un prezzo, pagatelo. Quando perdonare sembrerà una sconfitta, perdonate. Quando tutti vi diranno di pensare prima a voi stessi, non abbiate paura di farvi gli ultimi; siate capaci di dare tempo, soldi, comodità e progetti per servire gli altri. Che chi vi guarda possa chiedersi: “Perché vive così? Da dove trae questa libertà? Perché si amano tanto?”. #ViaggioApostolico #Uruguay
The P in JEPA stands for predictive.
The main distinction is not between JEA and JEPA, but between Joint Embedding (predictive or not) and generative architectures.
Whether a predictor is necessary depends on the task.
Given observations x and y (y being available only during training), a generative architectures predicts y from x. The loss is the divergence between y and its prediction:
L = divergence(y, pred(enc(x)) )
A JEA encodes y as well and minimizes the divergence between the representations of x and y, while preventing collapse:
L = divergence( enc(y), enc(x) ) + R( enc(y), enc(x) )
Where R is the anti-collapse regularizer, which may be computed over a batch of samples.
JEA are used when the prediction has no particular direction, i.e. when x and y are different views of the same scene. The idea for JEA is very old. Becker and Hinton proposed a simple model in 1989 where the anticollapse was based on information maximization. Schmidhuber had a version in 1992 based on redundancy minimization. At NIPS 1993, I proposed Siamese nets, which was a contrastive method to train JEA. Such contrastive methods gained popu.arity after Google's SimCLR paper in 2020. In 2021, Stéphane Deny in my group proposed the Barlow Twins redundancy reduction regularizer, which seemed to work at scale. This was shortly followed by VICReg by Adrien Bardes, Jean Ponce, and me, and last year by SIGReg by Randall Balestriero and me.
DINO is a JEA method that comes from a long string of papers from Piotr Bojanowski's group at FAIR-Paris, starting with Noise as Targets (2017), Deep Clustering (2018-2020), and DINO v1-v3 which are contrastive between clusters, not samples.
JEPA adds a predictor, which is useful when the prediction is non-symetric. For example, when x is a degraded/corrupted version of y, or when x is a present observation and y a future observation (temporal prediction). The loss is:
L = divergence( enc(y), pred(enc(x)) ) + R( enc(y), enc(x) )
All of these models can be augmented by an action variable a, fed to the predictor, encoding an action, intervention, or transformation to produce y from x.
This turns the system into a "world model". In a JEPA, the loss becomes
L = divergence( enc(y), pred(enc(x), a) ) + R( enc(y), enc(x) )
Thiel se ha atrincherado bajo el discurso religioso y nacionalista. Lleva meses tratando de presentar al Papa como Anticristo y ahora usa el mismo calificativo contra los demócratas
Empresarios como Thiel y Elon Musk sienten que deben guiar el futuro de la humanidad a toda costa
NEW: In leaked recordings, Peter Thiel argued that Nashville should be a moral alternative to Pope Leo XIV and the Catholic Church, which he called “a dubious ally” against a hypothetical future Antichrist Democratic president.
“America is the last and final form of the Roman Empire,” Thiel told a closed-door crowd in Nashville’s Masonic Temple. “There is no successor.”
He has attacked Pope Leo XIV at least four times this year, once calling him “a useful idiot for the CCP.” https://t.co/XMjDATdnkf
🎉Sharing our SIGGRAPH 2026 paper Divide and Truncate(DAT): one rule that guarantees penetration-free contact of robots, soft bodies, cloth, and multi-physics. Robust, blazing fast and no deadlocks.🧵
with @JerryHsu32@milesmacklin Youssef Ayman
Website: https://t.co/MyR9yYHlX8