Previously, we’ve written about stochastic computing and why we believe verifiability is necessary:
Since we can’t really know what the AI model will do based on what we give it, we at minimum should know which model we’re asking to do something for us. Be it generate a text summary, generate a video, or perform actions like buy a plane ticket.
While our post focuses on how to provide assurance for which model is performing our tasks, this video highlights the challenges related to building and supporting products that rely on stochastic computing primitives (like LLMs and other AI models).
it is interesting to hear the product leaders’ speak about the uncertainty around AI models. Even within the companies building these models, they have limited understanding of a range of things:
What the models capabilities are
When new capabilities could be available for general use
How to plan for new capabilities within a complete product
How different use cases perform based on current capabilities
How to evaluate for different, latent capabilities within an AI model
This discussion provides a good glimpse into the future of software, products, services (and eventually, physical products & manufacturing) that are built around AI.