* Motion design and creative studio depicting the 
present and the future of technology.

ping@sensor.graphics
Instagram
About




Current research

Hardware Evolution
[]


Spatial AI>    Spatial AI is an exploration that imagines a combination of two emerging technologies: artificial intelligence tools and spatial computing. Recontextualizing neural networks and integrating them into living spaces, the project invites to imagine AI as a structure that understands physical reality and becomes a new, constantly present layer in it.
>    With spectacular advancements over the last few years, AI polarized the discussion around it. It’s positioned as ‘magic’ and a universal good by some and feared as a catalyzer of many doomsday scenarios by others — and attracting the most attention from the people, these extreme opinions tend to obscure the current state of the technology and its possibilities and challenges.



>    At their core, present-day AI technologies, namely LLMs and diffusion models, work as powerful statistic machines that use dozens of thousands of parameters to generate the response to the user prompt. They navigate swathes of data to produce the most probable outcome — but despite their impressive capacities to decipher user’s input and transform it into something new, at the moment, they lack the understanding of the concepts that would be similar to the human comprehension. 
>    The modern-state AI is not connected to the physical reality in the way that we are — all of the concepts exist for it as a series of data bits, or tokens, that are connected to each other rather than to anything in the real world. Because of this, AI can not evaluate its responses to see if there is any sense in them. Something obvious for humans turns out to be impossible for an artificial mind.



       
>    At their current condition, AI technologies are able to find patterns and organize existing information, empowering human innovation but not able to innovate themselves. There are various efforts to overcome this barrier and some first results. In research of LLMs, DeepMind were able to add an evaluation component to their model searching for solutions for a mathematical problem, and as the result, it was able to come up with innovative approaches to the question.
>    There are also initiatives that aim to escape weaknesses of current models by adding elements of logic and semiotic knowledge to them. An approach called ‘neuro-symbolic AI’ brings together statistical methods of neural networks and symbolic systems that contain logical relationship between phenomena, i.e knowledge graphs. 





>    Purely symbolic approach was popular until the 1990s, but it fell out of fashion when its limitations became apparent. Symbolic systems’ scalability and ability to act in uncertain situations were extremely limited, as they relied only on the data that was added manually. As a part of hybrid approach, symbolic AI made a return in the 2010s and is now applied in companies like IBM.

>    Other efforts are focused on adding spatial awareness to AI. Models can be placed into virtual simulations and learn the dependencies in a fully simulated environment, or they can be put into a robot that they can control so that they can interact with the real physical location to understand it. One of the most famous proponents of this approach is Dr. Fei-Fei Li, a ‘godmother of AI,’ whose research was at the foundation of making computer vision possible and who considers spatial understanding to be a key to general reasoning. 




>    In this short film, we reflected on how AI is learning to understand and navigate the three-dimensionality of the physical world. AI models are presented as installations placed into rooms, and their abstract representation is based on structure and algorithms of data processing within neural structures.

>    Taking AI out of the black box and embodying it, we highlight the beauty of neural connections that bring about thinking and reveal how information is constantly flowing around us.



>    In this speculative scenario, AI is imagined as an object integrated into the architecture that is transparent in its behavior. Processing data, it considers its surroundings and context of interaction with a user, for whom this spatially oriented approach could be another step in the intuitively comprehensible and multimodal experience of the technology. With XR, this representation could be made possible, bringing more knowledge and understanding to the end user.

>    This spatially defined approach could also have locality of AIs in its foundation. Deploying custom decentralized neural networks would give users more control over how they use the technology and allow them to adapt it to their needs. In this vision, AI is ‘disenchanted’ but close and accommodating — becoming an everyday technology rather than a mystical force.




Credits
Design & Animation: Sensor.Graphics 
Research & Curation: Laura Herman, Maggie Mustaklem
Year: 2024





Sensor.Graphics© 2024—09—P02