From Code to Capital: Why the Real AI Edge Is Systems, Not Summaries
I’m an equity analyst who’s been learning to build with AI tools over the past several months. Two things hit my feed today that are really the same conversation happening in two different worlds.
The finance side
Y Combinator’s 2026 “request for startups” includes a section on AI-native hedge funds. The thesis: the next Renaissance, Bridgewater, and D.E. Shaw will be built on AI. Swarms of agents will do what hedge fund traders do now.
Brett Caughran (ex-Maverick, D.E. Shaw, Citadel, Schonfeld - now building an analyst training firm) pushed back. His argument: big funds aren’t threatened because LLMs are still an “unreliable world calculator” — they can complete quantitative tasks about 75% of the time, which isn’t exactly keeping Renaissance up at night. And you’re not going to unseat quant firms running on massive leverage, compute budgets, and decades of infrastructure with “two guys and a swarm of agents.”
The real opportunity, he argues, is somewhere else entirely (emphasis mine):
“The much bigger opportunity is to help fundamental investors become more ‘quanty’ by regaining some of the stacked alpha into the fundamental process that quants have been eating for the last 10+ years. Building the ‘Iron Man’ suit that becomes an AI Native operating system about both internal & external data (i.e. a swarm of agents can’t walk the floor of a tradeshow or attend the JPM HC conference). No one has really done this yet, because it is an incredibly difficult, but incredibly high ROI task if done right.”
In a separate thread, he showed an example of one small tool that could arise from a mindset and a system like this — a Claude prompt that generates a structured 7-page institutional-grade technical analysis report on any ticker in about 10 minutes. One signal in a larger system.
kwak (Point72 / Perplexity) quote-tweeted Brett and added a couple of important points. The core of it:
“Most AI ‘research’ tools for finance today is simply a regurgitation of the web, which consists of the opinions of others... it generates almost no alpha because it is simply an average of public sentiment. This is like paying a HF analyst to summarize 2-3 news articles as his ‘recommendation’ - would be fired instantly.”
His argument: the real shift would be AI that can formulate a plan of attack, retrieve primary data from filings or expert calls, build analysis tools, and synthesize. He points out that the “holy grail analyst” - someone who could take any dataset (credit card data, 300 expert call transcripts, weather forecasts) and build the tools to analyze it AND turn it into a recommendation - was always the dream at big funds. Analysts refused to learn Python (and even if they did, it still took time to write). With Claude Code, it has become dramatically easier to become that person.
The software side
The same day, Gergely Orosz (who runs The Pragmatic Engineer, one of the most-read software engineering newsletters) published a conversation with Grady Booch. Booch co-created UML, spent decades as an IBM fellow, and is one of the founding figures of software engineering as a discipline.
His thesis: software engineering has gone through three “golden ages,” each defined by a rise in the level of abstraction. Assembly to high-level languages. Individual programs to object-oriented systems. And now, apps and programs to integrated systems and platforms. AI coding tools are the latest step in that pattern.
His take on the current anxiety about AI replacing developers:
“Fear not, O developers. Your tools are changing, but your problems are not.”
And on Dario Amodei’s prediction that software engineering would be automatable in 12 months, Booch pushed back. His point: Amodei is talking about automating the lowest layer (writing code), not the actual job of software engineering, which is balancing technical, economic, and human forces to build systems that endure.
The connection
These are the same conversation.
In software, the easy stuff automates first - boilerplate, repeated patterns, standard web applications. In investing, the same thing: “summarize the web” is useful but it’s consensus regurgitation. No edge.
What doesn’t automate is similar in both worlds: systems thinking, judgment calls, relationships, and field work. The scarce skill isn’t writing code — and soon it won’t be just reading 10-Ks or building models either. It’s designing and managing systems, then allocating attention, time, and capital around them.
People and companies worth following on this
Brett Caughran
Check out his Cutting Edge podcast
kwak
Gergely Orosz
Grady Booch
Khe Hy
Portrait Analytics
Rogo
Finpilot
Clarity Markets (@IndraStocks)
Jim Moran (Flat Circle)
Who am I missing?
It’ll be interesting to watch who actually builds the “Iron Man suit” - and whether it comes from a startup, a fund building it internally, or something nobody’s thinking of yet.
PS: I’m having a lot of conversations with funds and analysts figuring out this shift — what the AI-native research workflow (or asset management firm more broadly) actually looks like, and who can build it. If that’s you, I’d love to hear what you’re seeing. DM me, reply, or book a call here.