AI: Events
How to Make a Large Language Model Smaller Without Losing Quality
Technical context • Development
The TorchAO developers have expanded their toolkit for quantization-aware training, now supporting new architectures, modes, and tasks.
NXP and Hugging Face explain how to train robotic artificial intelligence on custom data and run it on a low-power embedded device.
AI: Events
How AI Learns to Improve Its Own Code: An Experiment in Self-Optimization
Technical context • Research
AMD researchers have demonstrated how an AI agent can iteratively optimize high-performance code without human intervention.
Helion, a DSL for writing fast ML kernels, has gained a new automatic tuning mechanism based on Bayesian optimization that saves developers' time.
AI: Events
Instant Neural Network Updates: How Doc-to-LoRA and Text-to-LoRA Are Changing the Game
Research
Sakana AI has proposed a method to instantly update the knowledge of language models without costly retraining – by generating adapters directly from text.
Inception Labs has released Mercury 2, a new generation of diffusion language models that generate text in a fundamentally different way than the AI assistants we are accustomed to.