Unreliable Statistics at Scale (aka AI) with Craig Martell, Lockheed Martin's CTO

Craig Martell has seen AI adoption from more angles than almost anyone: a Silicon Valley machine learning leader (LinkedIn, Dropbox), the first Chief Digital and Artificial Intelligence Officer (CDAO) of the Department of Defense, and now CTO of Lockheed Martin. He brings that rare Valley-to-Pentagon vantage to one question: what does it actually take to field AI you can trust when lives are on the line.

Craig makes the case that AI is "statistics at scale" — gather data from the past to predict the future, and therefore guaranteed to be sometimes wrong. So the real work is defining the error envelope use case by use case and building "human escape hatches" where mistakes are cheap to detect and correct. We dig into why he's deeply bullish on the technology yet skeptical of a "not fully baked" product that ships with an asterisk, the product-management gap between how government writes thousand-page requirements and how the best new entrants build, and how the US and Chinese defense industrial bases really differ.

We then get into buy vs. build and learning from Ukraine, Silicon Valley's 70%-good-enough mentality vs. defense's demand for 100%, the fight over IP ownership and the stewardship of taxpayer dollars, and where an Apollo-scale public bet might still be worth making. Finally, the frontier Lockheed is investing in: quantum computing, sensing, and PNT; human-machine teaming and collaborative combat aircraft; microelectronics and friend-shored fabs; and Craig's "principle of stuff" — why all computing is ultimately manufacturing, and why low-power inference at the edge is going to matter.

–

This and all episodes are enhanced with lots of useful links and transcripts which you can read at typhoonbearing.com

–

Timestamps:

(00:29) Ukraine, Iran, and the future of modern war

(01:29) AI as "statistics at scale," guaranteed to be sometimes wrong

(06:44) Human escape hatches and error envelopes

(09:59) CDAO to CTO: the product-management gap in defense

(11:38) How the US and Chinese industrial bases differ

(14:06) The "not fully baked" product and the asterisk

(24:45) When will AI make a lethal decision on its own?

(28:58) Buy vs. build, and learning from Ukraine

(31:01) Silicon Valley's 70% vs. defense's 100%

(38:03) IP ownership, taxpayer dollars, and parallelism

(43:44) Quantum, GPU concentration, and an Apollo-scale bet

(48:34) Quantum sensing, PNT, and human-machine teaming

(55:15) Microelectronics, fabs, and friend-shoring

–

I love mail. Send comments, critiques, and takes to ⁠⁠⁠typhoonbearing@gmail.com⁠⁠⁠

⁠Visit the Typhoon Bearing website: typhoonbearing.substack.com⁠

Follow me on Twitter:⁠⁠⁠⁠ @ChaseHDalton⁠