AI is 'not smart' so what's next in artificial intelligence?
Despite advances, current AI systems are often described as "not smart" due to their lack of true reasoning and understanding. The article explores the limitations of today's large language models and discusses emerging approaches, such as neuro-symbolic AI, that aim to create more intelligent and capable artificial systems in the future.
Background
- Leading AI researchers (Yann LeCun of Meta, Yoshua Bengio) now publicly argue that today's Large Language Models (LLMs) — systems like ChatGPT — are not on a path to true intelligence. They call them sophisticated pattern-matchers, not reasoners.
- The article explains the gap between "narrow AI" (one-task systems) and AGI (human-like general intelligence). The key debate: are LLMs a stepping stone or a dead end?
- "World models" and "system 2 thinking" are proposed next steps: AI that builds internal simulations of physics and causality, and engages in deliberate planning rather than just predicting the next word.
- Why it matters: The industry has bet billions on the idea that simply making LLMs bigger and feeding them more data will eventually produce AGI. If the field's own pioneers say that's wrong, it could redirect the entire future of AI.