DSS Societies: A Sustainable Path for Generative AI
ยท Margarita Belova, Yuval Kansal, Yihao Liang, Jiaxin Xiao, Niraj K. Jha
A new paper from researchers argues that generative AI is heading for a wall. The problem isn't just the energy cost of training massive models anymore.
A new paper from researchers argues that generative AI is heading for a wall. The problem isn't just the energy cost of training massive models anymore. As these systems become real products, the real drain is inference: every time a user asks a question, the model does expensive, repeated computation. This gets worse with so called reasoning models, which can multiply costs by orders of magnitude per query. The paper points to grid failures, water consumption, and the simple fact that piling on more data yields diminishing returns. You get models that remember facts but can't reason through problems they haven't memorized.
The core insight is that current LLMs only show real reasoning depth in math and coding. Those fields have rigorous, pre existing formal abstractions like logic and symbolic notation. Other domains lack that structure, so the models flounder. The authors propose a different path: domain specific superintelligence, or DSS. Instead of one giant model that tries to do everything, you build explicit symbolic knowledge graphs, ontologies, and formal logic for a given field. You use those to generate synthetic training curricula for small, focused models. This avoids the dreaded model collapse that plagues most synthetic data methods today.
The vision isn't a single AI god. It is a society of specialized DSS models. Orchestration agents route tasks to the right expert back end. This decouples capability from raw size. Intelligence can move from energy hungry data centers to secure, on device specialists. The authors argue this shift makes generative AI sustainable, turning it from an environmental problem into a tool for economic empowerment that respects physical limits. The takeaway is clear: bigger isn't better if it breaks the grid and still can't think. Focused, grounded intelligence might be the only way forward.