In controlled experiments, MIT CSAIL researchers discover simulations of reality developing deep within LLMs, indicating an understanding of language beyond simple mimicry.
New algorithm helps robots practice skills like sweeping and placing objects, potentially helping them improve at important tasks in houses, hospitals, and factories.
Neural network controllers provide complex robots with stability guarantees, paving the way for the safer deployment of autonomous vehicles and industrial machines.
LLMs trained primarily on text can generate complex visual concepts through code with self-correction. Researchers used these illustrations to train an image-free computer vision system to recognize real photos.
“Alchemist” system adjusts the material attributes of specific objects within images to potentially modify video game models to fit different environments, fine-tune VFX, and diversify robotic training.
Three neurosymbolic methods help language models find better abstractions within natural language, then use those representations to execute complex tasks.
A multimodal system uses models trained on language, vision, and action data to help robots develop and execute plans for household, construction, and manufacturing tasks.
“Lightning” system connects photons to the electronic components of computers using a novel abstraction, creating the first photonic computing prototype to serve real-time machine-learning inference requests.
