Machine-learning system based on light could yield more powerful, efficient large language models

An MIT-led team has developed a new system that holds the potential to create machine-learning programs significantly more powerful than those powering ChatGPT while consuming far less energy than current supercomputers. The system relies on the movement of light rather than electrons and uses hundreds of micron-scale lasers to perform computations. This approach leads to over a 100-fold increase in energy efficiency and a 25-fold improvement in compute density compared to current digital computers. The researchers envision that this technology could pave the way for large-scale optoelectronic processors, enabling high-performance machine learning tasks on small devices like cellphones, which currently rely on large data centers for such computations. The technology could also accelerate the development of machine-learning models that were previously considered economically unviable due to computational constraints.

Link to the full article: Machine-learning system based on light could yield more powerful, efficient large language models | MIT News | Massachusetts Institute of Technology