Post #1121559

"Scaling is no longer the essence of innovation," writes Gary Marcus on the fusing of LLM and symbolic AI that is happening in recent Claude coding models. https://garymarcus.substack.com/p/the-biggest-advance-in-ai-since-the It seems an acknowledgement that the central flaws inherent on LLM (being probabilistic and therefore unreliable) are, well, inherent and can only be dealt with through a different paradigm. If scaling is not the essence of improvement, then this makes the infrastructure investment boom even more unsustainable than it already is. Thus, the crash of the AI bubble might not slow down AI development, but rather accelerate of the adoption of of new, mixed models. Add to this the reduction in inference cost that not only DeepSeek seems to achieve, https://castbox.fm/episode/The-Reality-of-AI-Economics-With-Paul-Kedrosky-id6019928-id929815749?country=us (starting at 00:27:21 Speculation, Supply Agreements, and AI Efficiency)