An Alternative Trajectory for Generative AI
A new research paper proposes an alternative trajectory for generative AI, moving away from large, monolithic models towards domain-specific superintelligence (DSS). The authors argue that current LLMs are unsustainable due to high inference costs and limited reasoning capabilities outside of math and coding. Their proposed DSS approach involves creating explicit symbolic abstractions and using smaller models trained on synthetic curricula, forming 'societies of DSS models' that can be deployed on-device. AI
IMPACT Proposes a shift from large, energy-intensive LLMs to smaller, specialized models, potentially reducing environmental impact and enabling on-device AI.