Naruhodo #473 - Estamos descobrindo como a IA pensa?
Summary
Hosts Ken Fujioka and Altair Souza debunk a recent Anthropic paper claiming AI possesses a "global workspace" similar to human consciousness. Altair argues the study is marketing-driven, misusing Stanislas Dehaene’s neuroscience theory to anthropomorphize statistical processes. He explains that the so-called "Jacobian lens" is merely a standard mathematical tool for probability calibration, not evidence of thought. The episode concludes that such narratives manipulate public perception rather than advance technical understanding.
Topics discussed
Introduction and support for the Naruhodo podcast
Listener question on AI interpretability and the Anthropic paper
Critique of the paper's marketing nature and neuroscience analogies
Explanation of Global Workspace Theory and probability spaces
Deep dive into Jacobian Lenses and statistical transformations
Deconstructing anthropomorphism and misleading AI terminology
The inefficiency of LLMs compared to biological brains
AI in mathematics: The Jacobian Conjecture counter-example
Interview with André Cravo on neuroscience and AI boundaries
Data center infrastructure, energy costs, and final thoughts
Listen ad-free on Castria