phong

Your Agents Need Context **To be clear, I don't think @yrechtman is wrong, and we've discussed this privately. As we're in the middle of a raise (no, you can't participate) and this article has come up a few times in the last several days, I'm sharing my thoughts publicly.** First off - on selling the tasks vs. the corpus. I think it's a position rooted in today's current reality. But at the same time, I question what the moat is for RL Labs more generally. AI systems are tools for solving complex problems very fast. Turning a corpus into tasks, and tasks into an environment, is a complex problem, indeed. But its mostly a software problem - and that's also the one area (as we know) where the current models are already at their best. At this point, frontier labs are looking at RL Labs as a bet on labor savings - and labor margin is not a moat. The labs will internalize this work when it is worth their time, or when their security posture demands it. Both are already happening (see Google's Mechanize deal). But where I really think Yoni's framing could be augmented is with respect to agents (I have an article in my drafts called "Don't forget your agents"). While Yoni points out the path from corpus to environments, there is a second way to use and sell data - which is to literally put it in front of the model at the moment it is doing the work. In context. At inference time. And if AI systems are tools for solving complex problems very fast, then the constraint on the tool is what it knows about the problem in front of it. Earlier this year @jasonscui and @JenniferHli wrote a great article about this here: a16z.com/your-data-agents-ne…. by their account - enterprises pointed agents at their data stakes and hit a wall - not because the models were weak, but because the agents lacked context. Agents are essentially useless without the right context. And it's not only business context that an agent will need. It's run-of-the-mill stuff. Like weather and CDC data when making your next retail trade (imagine being short bullish on Six Flags during hurricane season), political sentiment data when placing your next trade on Kalshi, or traffic data when scheduling a flight to JFK. If an agent gets these economically valuable tasks wrong because it doesn't have the right data - what's missing is not the RL environment, but the corpus - the raw data. Think of it another way - alot of what we as humans do is based on ingrained context. Think of it another way - alot of what we as humans do is based on ingrained context. We know to talk to a customer or two before starting a company or an off-the-shelf dataset marketplace (at least some people do), we know not to fall in love with Chinese spies (at least some people do), we know....alot of things. Now run the same logic across every facet of what's economically valuable. And much of what an agent needs to do real work is somebody else's private data: a supplier's price list, a counterparty's history, a dataset sitting inside a company that has never thought of itself as a data vendor. Most of that data is never going to be sold into anyone's weights. And unlike an environment, this isn't work a lab can pull in-house. But it can be licensed for use - on a consumption basis - with a record of every call and a switch to turn it off. This is already happening. LSEG, S&P Global and FactSet now serve licensed data straight to agents over MCP. None of that is a training deal. It's retrieval, under license, at the moment of the question. Training happens once per model. Context happens once per task. That's why I think in-context, inference-time retrieval of private data is the civilizational-scale tooling here. Yoni ends his piece with: "The next corpus is the whole economy." I'd reframe: "Context is the whole economy."