
LLMs, Neurosymbolic AI, and AGI Skepticism

Evolution of software architecture with the co-creator of UML (Grady Booch)
Topic 17 of 18
Published Dec 4, 2024 · Clip 1:07:20–1:24:58 (17:38)
Booch gives his AI credentials: he documented the as-built architecture of Watson Jeopardy (a pipeline of statistical systems, not a neural network), led a year-long study advising IBM on cognitive systems, and warned management against hype — an 'I told you so' when Watson's limits showed. He describes helping build Watson-based robotic concierges for Hilton (Aldebaran Pepper and Nao robots) and NASA's Robonaut 2 for Mars missions, where he architected the neurosymbolic 'Self' system combining Minsky's Society of Mind, Brooks's subsumption architecture, and Hofstadter's strange loops; the team worked with Soul Machines and Woodside until IBM cut the Watson group. On LLMs specifically: they're 'pretty freaking cool' but 'unreliable narrators' — stochastic parrots that don't reason or understand; with Gary Marcus he argues scaling is hitting diminishing returns and AGI hype diminishes human intelligence; he was blocked by Yann LeCun on Twitter for early warnings. He grounds this in cortical columns, theory of mind, embodiment, and the sparse multimodality of LLMs, and describes his tweet calling for a UML for AI, his internal visualization work, and plans to document AlphaFold 3's architecture in UML.