English Wikipedia is 29.4B characters, with ~1.5 bits of information per character. So that means, at 3.6 bits per parameter, a 12B model could memorize all english Wikipedia, it seems wild!
MACHINE LEARNING
-
SignSGD and Signum: Sign-Based Gradient Optimization Methods
By
–
When we derived SignSGD and Signum, it was surprising how just keeping the sign of stochastic gradients is sufficient for convergence and we could prove that theoretically. In practice, Signum and derivatives like Lion work well even at large scale. It is interesting to see Adam
-
Long Context Capabilities in AI Models Beyond 192k Tokens
By
–
We like long context. Go beyond 192k plz ; )
-
Company Recognized in Gartner Magic Quadrant Data Science ML Platforms
By
–
We wouldn’t be here without the trust and support of our customers and partners. 🤝
— DataRobot (@DataRobot) 6 juin 2025
Thank you for being part of our journey to the 2025 Gartner® Magic Quadrant™ for Data Science and Machine Learning Platforms.
Explore the full report: https://t.co/IjzzoJ3wLr pic.twitter.com/K9x6gphgTsWe wouldn’t be here without the trust and support of our customers and partners. Thank you for being part of our journey to the 2025 Gartner® Magic Quadrant™ for Data Science and Machine Learning Platforms. Explore the full report: https://
bit.ly/4mHwdL9 -
Groq LoRA Fine-Tuning: Adapt Models Without Retraining
By
–
Your model was good. Until it wasn’t.
— Groq Inc (@GroqInc) 6 juin 2025
Performance stalled. Costs climbed.
Your team hit the limit. Again.
Groq’s LoRA fine-tuning changes that:
✅ No retraining
✅ No downtime
✅ No compromise
Now you can adapt in real time without starting over.
How it works in the comments… pic.twitter.com/1aD8tKOHhjYour model was good. Until it wasn’t.
Performance stalled. Costs climbed.
Your team hit the limit. Again. Groq’s LoRA fine-tuning changes that: No retraining No downtime No compromise Now you can adapt in real time without starting over. How it works in the comments -
ScreenSuite: Comprehensive Evaluation Suite for GUI Agents
By
–
Today we release ScreenSuite, the most comprehensive evaluation suite for GUI agents (aka Computer Use agents). We packed 13 benchmarks, and 3 different environments, to evaluate the full range of agentic capabilities for vision models. And it turns out, @Alibaba_Qwen models are
-
Attention optimization over truth in reinforcement learning systems
By
–
It's because the objective is not truth but attention and they get RL'd by it, so they are a lot more optimal than you give them credit for.
-
Computer Vision Papers Feature Superior Visual Quality
By
–
computer vision papers have the best figures. this is a fact of life.
-
Boltz-2: AI Model Accelerates Protein-Binding Drug Discovery
By
–
MIT CSAIL, @AIHealthMIT, & @RecursionPharma have released a protein-binding affinity model called "Boltz-2" that could expand the role of AI in drug discovery.
— MIT CSAIL (@MIT_CSAIL) 6 juin 2025
The system predicts how molecules interact — & how tightly they bind — in record time: https://t.co/Vvq7Nc4yIH pic.twitter.com/4Lamu0XnDXMIT CSAIL, @AIHealthMIT
, & @RecursionPharma have released a protein-binding affinity model called "Boltz-2" that could expand the role of AI in drug discovery. The system predicts how molecules interact — & how tightly they bind — in record time: https://
bit.ly/43NRlXr -
Tesla Optimus Robot Now Folds Laundry and Cleans Tables
By
–
Il y a un an, les robots humanoïdes ne savaient presque rien faire.
— VISION IA (@vision_ia) 6 juin 2025
Aujourd’hui, Optimus de Tesla plie du linge, jette les poubelles, nettoie une table.
Le transfert sim→réel est presque résolu.
Les gens ne réalisent pas encore à quelle vitesse ça va.
Extrait : https://t.co/8av1IWoGsF pic.twitter.com/CBEgPVcmRfIl y a un an, les robots humanoïdes ne savaient presque rien faire. Aujourd’hui, Optimus de Tesla plie du linge, jette les poubelles, nettoie une table. Le transfert sim→réel est presque résolu. Les gens ne réalisent pas encore à quelle vitesse ça va. Extrait :