FSD: Thoughts on AGI

In this FSD episode, titled AGI Just Passed the CAPTCHA Test, Ram Ahluwalia takes apart the claim that AGI has arrived. He reads Jensen Huang's declaration as positioning rather than analysis, given Nvidia's incentive to pat its largest customers on the back and to move before AMD does. The test the industry actually cleared, he says, was the CAPTCHA test. AGI is a far higher bar than telling a bus from a not bus.

The margin story is the sharper one. Open source models are caustic for the economics of Anthropic and OpenAI, and Ram expects a crowded summit in three years rather than a duopoly, closer to the search engine wars than to a winner take all. His own daily use of Astra backs the skepticism. It is a token hog, slow, and careless about oversight, prompting none of the workflow choices that would speed a build up.

Looking for a standard that holds, he lands on American pragmatism, where an idea is true if it proves useful. That thread runs through William James, Karl Popper on what we can know, David Deutsch on physics and knowledge, and the Chicago economists who applied it to law and behavior. Real AGI, he argues, is an agent handed a market and a budget. Memory, compute and natural gas still gate it, which keeps him on the MLP index.

[00:00] AI is creating more work, not less

[00:43] Why Jensen's AGI call is political

[01:19] Open source models are acid for margins

[01:52] Astra is slow, and it is not AGI

[03:21] Fireside chat with Michael Preck, September 16

[04:04] The practical test for AGI

[05:23] William James and pragmatism

[07:10] Bad philosophy leads to bad outcomes

[08:59] Karl Popper and what we can know