Frank Long

Key takeaways from my convo with @lennysan! When you can build anything, I think the skill becomes knowing which of the thousand possible things is actually worth building, and getting everyone pointed at it

Watched an awesome talk by the Head of AI at Coatue ($93bn hedge fund) today. @fyxlong walked through his approach to applying AI in financial services + the complexity & compression challenge enterprises must wrestle with as they diffuse the technology. Here are some of the notes I took: The decision function. - His mental model: a hedge fund is a function that takes in two parameters (external context & internal context) and spits out one tuple (aka decision): investment selection, size, and timing - He says achieving this is a compression problem of taking in all of the world's information & spitting out a single decision. - Principle: everything an investor can see on their screens at work, an agent should be able to see via API. - Data sources he focuses on: external is news, prices and filings via api. Internal is current positions, liquidity and risk requirements, past case studies, and history with a sector or management team. - Huge advantage: Coatue has a decades-long culture of rigorously documenting and capturing all of the internal context fed into the function. - The reality of this function: "I probably can't just take all of that data, jam them into a context window, and get my three magic numbers and go home. It can't be that easy." The over-simplified version of this decision function. Step 1: Gather internal/external data Step 2: Analyze the data Step 3: Make the decision - Every step has an input and an output and the input for the subsequent step is derived from the output of the previous step. - Each one of these steps has dozens of substeps and within each one of these substeps, you have to pick the optimal tool for the job. Picking the right tool for the job. - Every line in the function should leverage the optimal runner: who or what does each step. - Runner options: deterministic code, class ML, an LLM, or a person. - Code is fast, reliable, and cheap. Great for use cases like "if PE ratio <15, do this". - Classic ML is great because you can model much more complex relationships vs. pure deterministic code. - LLMs are great for even more complex relationships like summarizing a transcript. - People still matter. For meeting a company's management team, "I probably don't want to send an LLM or an AI agent in to do that. I probably don't want to send an if statement into that meeting." - "There is no glory in automating things for the sake of automating things. The goal is not to take as many things LLM'd as possible. The goal is to maximize absolute quality for each of these steps." The realistic version of this function. - The real function isn't just a linear statement. There's all kinds of branching logic & tons of ways to fork the path the function takes (see image below). - And then each one of these steps in this far more complex branched function has an optimal runner, or really in practice, multiple runners per step. - And then there are so many other nobs you can pull outside of just the branched logic. Model size, thinking effort, hooks, skills, the list goes on. Simplifying such a multivariate problem. - The key question: "What can I actually compartmentalize in a way that we can understand and evaluate?" - "If you don't have a North Star to hill climb against, you're just kind of… fidgeting around." - Every data source has its own chain of steps, so the firm's real graph of its decision function is "massively complex." (see image below) - He estimates at least 10 knobs per node with at least 20 valid values each, which makes the search space "basically infinite." - The hard part is laying it out in a way people can understand technically and also "absorb culturally", inside the org. "That is the great challenge of applying AI in investing." Beyond investing - What he's really describing is "creating a state machine that can take information from the world and understand it well enough to make coherent decisions." - The same pattern applies to policy, nonprofits, drug discovery and corporate strategy.

Agents are the new users of CPUs. Coatue’s Frank Long and Nick Gagnet on the agentic shift inverting the GPU-to-CPU ratio.

Coatue hired a Claude Genius. aka "AI Mad Scientist" His name is Frank (@fyxlong) “He’s the one trying everything out there that is new, creative, on the cutting edge.” “Figuring out how do we implement it. How do we as an organization utilize our 20 years of data that we’ve been collecting to give us another leg up against the competition & allow us to succeed.” @coatuemgmt

Today we're announcing @Corridor's $25M Series A led by @Felicis. More code will be written this year than ever before. At Corridor, securing AI coding at the source, enabling companies to their development without security being a blocker. 🧵

friend of mine in finance has used mcp to automate a lot of his financial research and analysis - recently he built himself a slide gen mcp and said it plugged into claude code works better than any slide gen tool he’s tried github.com/FrankLong1/SlideA…

I keep changing my mind about the AI jobs crisis so jotted down 42 notes on where I’m at: 1) You don’t need mass unemployment to inspire mass fear—merely its shadow is enough. Just look at SAG-AFTRA and the port strikes last year. 2) Most AI backlash is economic anxiety coated in a veneer of social justice. Alfalfa farming consumes 19x the water that data centers do; there’s no sound environmental reason to boycott Claude but not GPS. When people say “AI is a moral scourge,” they really mean: I am scared that I won’t be able to pay my bills. 3) To be fair, the labs are definitely trying to automate everyone’s jobs. 4) Carl Benedikt Frey: “There is no iron law that postulates that technology must benefit the many at the expense of the few.” 5) In my last week as a product manager, I realized I didn’t have a single task to document and offboard. My role was relational, not task-based. Someone had to be the fall guy; someone had to herd the cats. 6) The map is not the territory. The org chart is not the org chart. Systems are much more unruly than they appear. 7) A common argument says that AI capabilities are fast but diffusion is slow. But it didn’t take students long to start ChatGPTing all their homework. “Diffusion lag” reflects a lack of product-market fit. 8) The real world is all edge cases, all the time. 9) Increasingly, fewer jobs will look like doing tasks ourselves, and more will involve teaching AIs to do them for us. How can we transfer context to the machine? Can they adopt the values and instincts we’ve evolved over millennia? When you pair with a model, will it remember what it sees? Can you teach taste? Creativity? Learning to learn? This is the great pedagogical project of our time. 10) Both human and machine intelligence seem infinite to me. Here’s the rest: jasmi.news/p/42-notes-on-ai-…