AI Is Dumber Than Investors Think — ft. Gary Marcus

The current AI boom relies on large language models that function as next-token predictors rather than true reasoning engines, resulting in persistent hallucinations and an inability to follow instructions reliably. This technological limitation is compounded by a widespread "over-attribution" of intelligence, where investors and policymakers mistake superficial mimicry for genuine cognitive capability. The industry’s economic model remains fundamentally unsustainable, marked by massive capital burn and intense price competition, leaving companies like OpenAI struggling to justify trillion-dollar valuations while competitors like Anthropic gain ground. Addressing these challenges requires a shift toward mandatory regulatory oversight and a move away from the current "exploit-only" approach to generative AI. By prioritizing fundamental research over the blind scaling of existing models, society can better mitigate risks like misinformation and cybercrime while exploring more efficient, reliable architectures for future intelligence.