Probabilistic programs present a solution to various challenges posed by large language models, such as hallucination and alignment problems. Unlike the opaque nature of neural networks, neuro-symbolic AI, the technology behind probabilistic programs, allows for auditing, control, and modification to better align with our preferences. This aligns with the ethos of Digital Architecture Lab (DAL), which advocates for transparency and decentralization and views the current lack of a technical means to express societal preferences and values as a significant gap.
Researchers at MIT’s Probabilistic Computing team are making breakthroughs in areas such as 3D scene perception, data-driven expert reasoning, and the reverse-engineering of human cognition and perception. In close collaboration with the MIT team, DAL, and Digital Garage are advancing Gen, MIT’s open-source stack for generative modeling and probabilistic inference. This collaboration seeks to demonstrate how probabilistic programming could offer new possibilities for safety, regulation, and reflecting social values in the system that is bound to be ever more complex over time.
To learn more about probabilistic programs and DAL’s mission in this field, read this blog post.
