Vasant Dhar has been working on AI since way before it was cool (and before most people thought it was possible). I've admired his work for a long time and since I started this show have wanted to have him on as a guest to talk about the strengths and limitations of LLMs. A few days ago, he published an article in The Hill about how DOGE is using AI wrong, and I knew I had to talk to him immediately.
We discuss, indeed, what DOGE is doing wrong, as well as the broader lessons we all can learn from this example about what makes a good vs. not-so-good use case for LLMs. We also talk about his recent and completely earth-shattering Severance-esque research where he and a colleague create a bot with the goal of simulating the valuation reasoning of investing legend Aswath Damodaran. Naturally, throughout the conversation, I pepper him with questions about how scared or optimistic we should be about all of this, as well as how AI can somehow be both incredible and limited at the same time.
Follow Vasant!
I'm @jonesrooy on Instagram and at jonesrooy.com.
And, because I'm trying to make friends with AI, below is a chapter summary! Woo hoo.
Chapters
00:00 The Evolution of AI: A Personal Journey
01:49 Understanding LLMs: Are They Thinking?
04:35 The Fear of AI: Balancing Optimism and Caution
06:57 Explaining LLMs to Newcomers
09:37 Learning and The Integration of the Senses
12:24 DOGE's Approach to AI in Government
15:22 The Need for Better AI Implementation in Government
17:47 Could This Actually Work?
20:16 A Problem with Using AI on Data from Humans
21:47 Contextualizing Project Failures
22:33 Applying AI to Investing
23:45 Building an AI Investor
27:36 The Science of Long-Term Investing
28:40 Simulating Expertise: The Damodaran Bot
32:33 The Art of Questioning in AI
36:07 Expectations and Realities of AI Development
40:10 Be the AI Change You Wish to See in the World
","author_name":"Andrea Jones-Rooy"}