Robert Ta argues that AI companies need to build closed-loop systems combining human expertise with proprietary AI models to avoid commoditization. He points to Harvey, a legal AI startup whose valuation collapsed when Microsoft released a free competing tool using the same underlying model, showing how easily AI expertise can be absorbed and given away.
A frontier without an ecosystem is not stable. I wanna give a quick take on like why I think Satya is saying this. I made an image, I just copy pasted his article and I was like, hey ChatGPT, make me like a diagram that explains this really well. Let me take you through it. Frontier without an ecosystem is not stable. The future of the firm, compounding human and token capital through a learning loop. So the transition here, past platform shifts, digital systems enhanced human capital. Now though, there's a real cognitive loop between people and digital systems. You feel this when you use ChatGPT or Claude Code. You make something, you create, you clear out some unknown unknowns or whatever, you get through some decisions. That loop is this, and this changes how we conceptualize work because now it's human attention and then the tokens that power the compute for the AI. That's what Satya is saying. So what's at stake? It's not a tool, but how organizations learn, build IP, differentiate, and thrive when AI models can continuously absorb and commoditize expertise. We've kind of seen this with— Harvey was sort of a darling at like a $10 billion plus valuation, if I'm not mistaken. Where they were just building an AI lawyer. And then Microsoft went off, Claude, I should say, first went off and released the legal plugin and tanked the entire legal tech market. Then tanked Harvey's valuation. Harvey came back and was like, oh, we're gonna make a harness. Microsoft comes out with, actually in Microsoft Word, you just get legal plugin for free now. So, and so that tanked Harvey's valuation even further. So yes, that is at stake. Anybody who could take your data, even outside of the models, but the models are, are the most susceptible because now we're seeing there's the model layer and people are trying to create an economy on platform and app layers on top of those models. So really we're selling picks and shovels and these picks and shovels have, uh, spy cameras in them. They're at that level. And hey, our spy cameras also take your IP and it makes our IP better. Maybe one day we can go upset your market. That, that's really what Satya's saying. Your knowledge becomes someone else's training data. That really precious, like the hard-earned secrets we hear about in startup life. Jonathan and I talk about this a lot, but you spend years in a domain. To earn the hard-earned secrets. Like, that's what Einstein did. That's what Galileo did. That's what any, like, invention that ever mattered, they did to find the hard-earned secrets. And now it's human attention and AI tokens that lead to those hard-earned secrets. And Satya is basically saying, like, you gotta be careful who's seeing that because if that's your IP, you might be just be giving it out depending on who you use. So new capital, human capital, token capital. Human capital is knowledge, judgment, relationships, ingenuity, pattern recognition, people. Token capital is AI capability your firm builds and owns. So the data model, the evals, the infra, the agents, all of that. This is what he's kind of saying: a compounding learning loop is where we're going, where the real opportunity is not quite picking the best model but building a loop where human capital and token capital compound. He's basically taking this position that this is the way all business will be done in the future. We take the time slider forward based on this current path. Human capital and token capital compounding loops are what we need. Beneath that, Jonathan and I believe autonomous closed loops fed from that to execute on your business towards business outcomes. That's obvious. Humans offload tasks to AI systems with those goals, agents, models, tools. People and AI work together, the human capital and the token capital. You learn and you improve. You get your private evals. That's something we help clients with. We've done a lot of this private reinforcement learning or RL on real traces, real data, real domain data that you can produce. And then ideally you get into a loop that makes your production better of whatever product or experience you deliver.
Evals are the moat, right?
Yeah. Evals are the moat.
We've been saying this for forever.
Evals are the moat. Evals are around your economically valuable workflows. Those are the moat. That's been so obvious. And it's obvious because when you look at the frontier labs, like, they're like, oh, there's so much money in coding. So we're going to try to dominate a moat around coding agents and then capture enterprise. That's what they're doing right now. And they're going to move into other areas. So the question for AI builders is, what are those areas? Where are the islands where I could survive and build a moat around?
This is a very welcome piece of thought leadership from Satya. Maybe some of the best corporate leadership that I've seen. This is a world that us as entrepreneurs want to be real. We want to manifest this world where there is a, there's a barrier in between the models coming in and using proprietary data to eat functionality within the organization. And there's an emphasis on these niche domains, domain expertise, private evals, and creating like unique competitive advantage from human capital. This is a world that we want, as Satya put it, politically and economically viable. So it, it, it alleviates a lot of the scare of like, oh, is AI gonna eat all the jobs? Is AI gonna like devalue companies, the SaaSpocalypse. Satya drew a line in the sand here. He created a possible future, and this is all a good future, you know, that kind of holds off all those fears. Good stuff.
