Here's what you'll learn when you read this story:
- While AI generators are far from the biggest contributors to climate change, their carbon footprint is steadily growing as more people use these platforms.
- Designed by scientists at the University of California, Los Angeles (UCLA), the new diffusion model uses laser light as a decoder, significantly reducing computational power and energy.
- Although these systems need to be integrated into existing AI infrastructure, the authors believe that energy efficiency is extremely beneficial for AI-powered wearable systems.
For your average person, using chatgpt doesn't actually produce that much collaboration2. In most estimates, each query produces approximately 2-3 grams of carbon dioxide (depending on the power data center of the energy source), containing the energy used to train the model first. Stretch it into a year, and it is about 11 kilograms, Many The energy industry alone is less than the impact of human carbon.
However, the environmental impact of AI has been growing concern, especially when ChatGpt users consider creating more than 700. A million Images just a week after March 2025. Ultimately, creating AI models that can make text and images as eco-efficient as possible is a top priority as it leads to a lot of carbon dioxide.
A new study, led by scientists at the University of California, Los Angeles (UCLA), details a new type of “optical generation model,” which uses light to generate images during the decoding process and uses only a portion of the computational power that is normally required. The results of the study were published in the journal Nature.
“Our work shows that optics can be utilized to perform optical generation tasks on a large scale,” said UCLA's senior author, Aydogan Ozcan, at a press conference. “By eliminating the need for heavy, iterative digital computations during inference, optical generation models open the door to energy-efficient AI systems that can transform everyday technologies.”
Of course, the system does not completely avoid digital computation. Instead, Ozcan's team trained a shallow digital encoder, along with a diffractive optical (i.e., light-based) decoder, as one complete system. This avoids the thousands of iterative steps typically used in digital decoders, and consumes computational resources. Instead, the new system, according to researchers, will provide images in “snapshots.” The center of this system is a liquid crystal screen called a spatial light modulator (SLM) that engraves static (essentially image information) onto a laser beam. By passing through the second decoded SLM, the system generates an image instantly.
“This is probably the first example of optical neural networks being a computational tool that can produce practical value results, not laboratory toys,” said Alexander Riboski of Oxford University. New Scientist.
To test this new system, the UCLA-led team created both black and white images and the creation of a Vincent van Gogh-style “artwork” (certainly he'll be excited about this). The results were comparable to the advanced diffusion model, but the process used only a portion of the energy compared to the traditional model. Another bonus with a light-based decoder is that it can improve security and privacy by creating ways that content is inaccessible except for the correct decoder.
While it is unlikely that these systems will be widely integrated in the short term, this breakthrough makes sense for wearable systems such as AI glasses, which consume less power.
Thus, while the majority of AI use is the increase in contributors to climate change, this study shows that there is plenty of room for sustainable improvement.
Darren lives in Portland, has a cat, and writes and edits about science fiction and how our world works. You can find his previous ones on Gizmodo.
