Introducing the new Gemini 2.5 Pro preview, with a +24 LMArena Elo score over its predecessor! It leads on tough coding (AIME, AIDER), science (GPQA), and reasoning (HLE). Style/structure have improved thanks to your feedback. Learn more: https://
goo.gle/4kKynYo
MACHINE LEARNING
-
Google Launches Gemini 2.5 Pro Preview with Major Performance Gains
By
–
-
How LLMs Work: Beyond Traditional Programming Misconceptions
By
–
LLMs aren’t “programmed” in the traditional sense. Many still think AI models are hand-crafted by engineers—just like symbolic AI. That’s not how modern large language models (LLMs) work. This misunderstanding leads to two big problems: 1. Underestimating what LLMs can
-
Closing the 100,000 Year Robot Data Gap IEEE Talk
By
–
Looking fwd to presenting this @IEEEorg @ieeeras online talk tomorrow at 7am PT: "How to Close the 100,000 Year Robot Data Gap":
-
Gemini 2.5 Pro update achieves SOTA performance with thinking budgets
By
–
Introducing our latest update to Gemini 2.5 Pro (06-05), which we expect to become our long term stable release. At a glance: – SOTA on HLE, Aider, and GPQA
– Now supports thinking budgets
– Same cost, on pareto frontier
– Closes gap on 03-25 regressions -
Fine-tuning GPT-2: When LLMs Aren’t Necessary for ML
By
–
After playing with fine-tuning gpt-2 for some simple classification tasks, my honest take is that for most *supervised* ML projects you really, really, really don’t need the LLMs. For simple image/text/signal problems use a pre-trained nn, such as one of the deep NNs trained on
-
Tesla Optimus Robot Trained in Simulation Transfers to Real World Zero-Shot
By
–
Optimus n’a jamais été entraîné dans le monde réel.
— VISION IA (@vision_ia) 5 juin 2025
Tous ses mouvements viennent de la matrice, un monde équivalent à la simulation.
Puis transférés en zero-shot dans le robot physique.
Bienvenue dans l’ère où les machines apprennent dans leur propre matrice.
Et ça… https://t.co/EsJk26XT5y pic.twitter.com/hsDs4vVOvJOptimus n’a jamais été entraîné dans le monde réel. Tous ses mouvements viennent de la matrice, un monde équivalent à la simulation. Puis transférés en zero-shot dans le robot physique. Bienvenue dans l’ère où les machines apprennent dans leur propre matrice. Et ça
-
Annoy powered virtual friend in production at Hugging Face
By
–
Annoy was pretty cool we had it in prod at @huggingface in the early days, powered our virtual friend
-
Pretraining: The Beautiful Art of Learning to Compress Knowledge
By
–
pretraining, the act of learning-to-compress the entirety of human knowledge and thereby creating a general-purpose model that can simulate any natural process, is really a beautiful learning paradigm. everything is just going to get messier & more complicated from here
-
Switch2 Day impacts Hugging Face team productivity
By
–
Switch2 Day has taken its toll on the @huggingface team’s productivity