Intelligent Alpha CEO: Letting AI Run the Portfolio
In the latest episode of Invest with AI, Brett and Khe sat down with Doug Clinton, founder and CEO of Intelligent Alpha, an investment company where frontier AI models do the analysis and build the portfolios. Doug is also a partner at Deepwater, the venture firm he co-founded almost a decade ago.
It started as one experiment. ChatGPT launched, and Doug asked whether it could beat the S&P 500, which most human managers can't do. The early results were good enough, and maybe beginner's luck by his own read, that he built a company around it. He now sits on both sides of the table: making the call as a human at Deepwater, letting the machines make it at Intelligent Alpha.
Brett came in skeptical and said so. He's been a hater on the idea of letting machines pick stocks, or at least the one asking for evidence. He left more open-minded than he came in.
Doug grades the models a B+ analyst. Good at spotting what's likely to move a stock in a quarter. Not as creative as a strong human analyst, not yet. His read is that the difference between a mega-cap list and a portfolio worth looking at has almost nothing to do with which model you pick. Ask one for ten stocks and you get eight of the ten biggest names in the index. What matters is the universe you define, the data you feed it, and whether anyone at your firm has ever written down how you actually define EBITDA.
Do that work and the models are useful. Skip it and you have a chatbot with opinions.
We get into:
The progression from prompt engineering in 2023 to context engineering to agentic workflows, and the frontier Doug is working on now, which is organizational context rather than task context — how to make the entire corpus of every portfolio your firm has ever generated useful to the agent doing the work today
Where AI-run portfolios actually land, which is between a fundamental book and a quant book: not a thousand positions chasing a 52-54% batting average, not fifteen names, but a couple hundred up to 500, with model-selected conviction bets of several percentage points
The three buckets of data — public and third-party fundamentals, LLM-augmented estimates built from consensus plus Reddit and X, and proprietary feet-on-the-street research — why bucket three is the bottleneck and the durable source of edge, and whether agents will eventually run channel checks by talking to other agents
What a knowledge graph actually is under the hood, why EBITDA means fifteen different things to fifteen people inside one firm, and how you give a model a map — often just markdown files — so it knows your software definition differs from your industrials one before it hands you a number for $CRM
Why a dev shop building this spends a big chunk of the engagement interviewing your PM, the pushback that follows — wait, I thought you were going to give me the answer — and why taste is what gets model output from B+ to A
The case for trading perfection for speed in an industry with real penalties for being wrong, why a six-month project to build an elegant knowledge graph will be obsolete before it ships, and how much Intelligent Alpha has built and thrown away, some of it inside a month
Model routing and the IA 500, the AI-constructed benchmark where ten frontier model families each pick a portfolio, plus the contrarian call that GPT rather than Claude is the better model for stock work — and Khe's counterpoint that harness beats raw intelligence when a 2-3% delta won't move anyone off the workflow they already run
The one thing every firm below the Citadel tier needs, which is someone in leadership willing to nudge compliance, accept a governance risk, and fund an analyst to go figure things out — and why funds still running Claude with web search turned off in 2026 are the counterexample
If you are building an AI-native research process at a fund, or trying to work out how much of the portfolio decision you would actually hand to a model, listen to the full episode.
Chapters (Timestamps)
[00:00] Intro
[02:39] — Can ChatGPT Beat the S&P 500?
[04:11] — Why LLMs Keep Handing You $NVDA
[05:05] — Prompt Engineering to Agentic Workflows
[06:45] — Organizational Context: Every Portfolio You've Ever Built
[09:07] — Public Markets Alpha Is a Power Law Game
[10:09] — Where Human Intuition Still Beats the Model
[11:14] — The Hybrid: Part Quant Book, Part Fundamental Book
[12:30] — The Three Buckets of Data
[13:58] — Can Agents Do Channel Checks?
[14:49] — Is MCP Institutional Grade Yet?
[16:11] — Grading the Models: A B+ Analyst
[16:55] — Knowledge Graphs Are the Frontier Right Now
[18:41] — Schemas, Ontologies, and the Gray Matter
[19:21] — Taste: Getting AI From B+ to A
[20:26] — What Building an EBITDA Ontology Actually Looks Like
[24:04] — The Dev Shop Wants to Interview Your PM
[25:37] — Trade Perfection for Speed
[26:36] — We've Built a Lot of Things We No Longer Use
[28:25] — The Vendor Cycle and Controlling Your Own Destiny
[30:23] — Model Routing and the IA 500 Benchmark
[32:36] — The Contrarian Take: GPT Over Claude for Stock Work
[33:53] — Codex vs. Co-work and a Branding Problem
[37:32] — Why There's No Harvey or Rogo for Investing Yet
[40:45] — Do Funds Want the iPhone or the Android?
[43:13] — You Need Someone in Leadership Who's AI-Pilled
[45:19] — Fable Built a 10-Stock Portfolio. It's All AI.
[48:28] — Brett Updates His Skepticism
[49:09] — Advice for Building an Asset Manager With AI
[51:33] — Getting Comfortable With Uncertainty
[53:03] — Go Get Lost in the Models
Want to actually build these workflows yourself?
The AI Accelerator is Fundamental Edge's 6-month cohort for investors who want repeatable AI workflows. Learn More below:
https://www.fundamentedge.com/ai-accelerator
Follow Invest with AI on:
Spotify: https://open.spotify.com/show/033xcEEovVViS7hIYwNuGZ
Apple Podcasts: https://podcasts.apple.com/us/podcast/invest-with-ai/id1896918892