Sakana AI has released a new Japanese visual language model, "Llama-3-EvoVLM-JP-v2," newly constructed through evolutionary model merging. The constructed model features the new capability to perform question-and-answer tasks in Japanese across multiple images. Blog →
RESEARCH
-
Adam-atan2: Removing Epsilon Hyperparameter from Adam Optimizer
By
–
Adam-atan2 is a neat one line code change to remove the epsilon hyperparamter from Adam by changing the divide to an atan2(). This feels like a generally useful trick for many places you might be worried about divide by zero and numeric precision issues.
-
Black Forest Team Launches FLUX: New Image Generation Models
By
–
¡NUEVO EQUIPO y MODELOS de IMAGEN! Se han juntado bajo el nombre de Black Forest Team varios investigadores de muchas de las tecnologías tras los modelos actuales y han creado un nuevo equipo, y publicado nuevos modelos: FLUX Y pinta bien!
-
Gemini 1.5 New Experimental Model Tops LLM Arena Rankings
By
–
¿Lo de que de repente haya un nuevo modelo experimental de Gemini 1.5 que se haya colocado primero en el LLM arena cómo se gestiona? Mirando de forma disgregada cada ranking parece ser que hace muy bien los deberes de dialogar y en código o matemáticas está a la par. Pero wow
-
AGI Scaling Limits Don’t Mean AI Won’t Be Transformative
By
–
Many people have been recognizing that AGI won't be coming from mere scaling of current tech, and that generative AI has been severely overhyped. This is all true. But those people often conclude, "therefore AI is not going to be transformative — this is a nothingburger".
-
MLCommons AlgoPerf Training Algorithms Competition Results
By
–
Congratulations to everyone who submitted to the
@MLCommons AlgoPerf training algorithms competition! We were delighted to provide compute resources for evaluating so many exciting submissions. -
Meta Wins AlgoPerf Competition in Both Tuning Tracks
By
–
The #AlgoPerf competition was designed to find better training algorithms to speed up neural network training across a diverse set of workloads. We’re proud to share that teams from Meta took first place across both external tuning and self-tuning tracks!
-
Schedule Free approach wins self-tuning track competition
By
–
More on the Schedule Free approach that won the self-tuning track.
-
Statistical Equivalence: Indistinguishability of Randomness and Complexity
By
–
The statistical equivalence principle: randomness and complexity are indistinguishable.