Keep pulling the thread on Oriol Vinyals.
Oriol Vinyals predicts that complex, hand-coded scaffolding systems built around models will eventually be written on-the-fly by the model itself.
Reasoning capabilities developed from training on narrow domains like coding and math have been shown to generalize to unrelated, complex topics like taxes and international relocation.
Oriol Vinyals believes the AI industry has not yet seen the equivalent of the "GPT moment" for video and images.
A core, unsolved quest in machine learning is to train a model on all video and image data without text and have it extract the same level of understanding that language models achieve from text.
Oriol Vinyals believes the ability for AI models to genuinely innovate, particularly in scientific fields like machine learning, is a key capability that currently lacks a clear research path.
Oriol Vinyals believes that based on the expectations of seven years ago, current AI models would likely have been declared as achieving Artificial General Intelligence (AGI).
A key challenge in training models on unlabeled visual data is linking abstract concepts to what is seen in an image without explicit language annotations.
World models like Google's Omni could provide a simulation dimension that enables systems to predict outcomes before acting in the physical world, with applications in self-driving cars and robotics.
A significant gap for using world models in robotics is the lack of precision for fine motor control, such as grasping, because the models lack data for modalities like touch and force.
Google released consumer agents named Spark at its I/O conference.
A promising mechanism for agent memory is to have the agent write its thoughts and knowledge into an external, modifiable file system.
Serving models with personalized weights for each user's memory would be practically difficult, making non-parametric memory systems like external files more convenient.