Going Beyond World Models and VLAs (generalistai.com)

Generalist has announced the development of GEN-1, a groundbreaking foundation model for physical interaction, which has been trained from scratch using about 99% of its parameters. This represents a significant shift away from traditional fine-tuned models, suggesting that with sufficient data and computational power, training from the ground up can yield superior results. GEN-1 is not merely a refinement of existing vision-language models or world models; it aims to achieve ambitious goals like zero-shot robotics, enabling robots to carry out tasks they've never encountered before with high success rates and minimal task-specific data. This approach emphasizes goal-driven research over method-driven research, allowing for greater agility in adapting various techniques to meet specific objectives. Generalist believes that by focusing on concrete milestones—like achieving high performance with minimal data requirements—they can push the boundaries of what is achievable in robotics. As constraints evolve, particularly in terms of data availability, models like GEN-1 are positioned to redefine the landscape of AI, opening pathways toward physical AGI and more sophisticated robotic capabilities. The insights from GEN-1 are expected to influence both academic research and practical applications in the industry.