OpenAI’s hiring of Peter Steinberger, creator of OpenClaw, is more than just a story of corporate ambition.
It’s a story about software development.
On the surface, giving Steinberger what could be a 9-figure incentive package sounds like a return to the era of the singular wild genius. “Peter pushes more code than almost anyone,” Om Malik wrote.
But why, Om, why? How does Steinberger get so much code into the world?
It’s because Steinberger isn’t Bill Gates. He’s Linus Torvalds.
This is also why I think Steinberger went with OpenAI rather than Meta or Anthropic, which were also recruiting him.
Meta and Anthropic, like most organizations, are proprietary companies operating with structured teams. While alliances change with the times, and with the fashions (don’t go chasing waterfalls, be agile) there’s at least stability to corporate development. Like there is in Game of Thrones. Or, as we will see, in Lord of the Rings.
OpenAI, by contrast, is a hot mess. As Om writes, “OpenAI has become a weird collection of mini organizations,” fiefdoms if you will. All the internal resources are spoken for and, unlike the situation at Meta and Anthropic, the boss has no more clue about fixing this than anyone else.
Enter open source.
The Nature of Agents
OpenClaw is about AI Agents. OpenAI, like its rivals, is built around a Large Language Model (LLM). Agents are about orchestration, not computation.
An LLM may have just a few thousand lines of code, but it takes enormous resources to compile. It’s great for Edgelords who can just order Nvidia chips, build data centers, and rent it all out. (Sort of like the oil business in a way.)
An AI Agent takes a lot of code to integrate many data sources, although the code may be in huge hunks, hidden away, locked in a computer file somewhere. Finding that usable code, and adapting it for use in a larger structure, is the ball game.
LLMs are the cathedral. Agents are the bazaar.
But where did agents come from?
Palantir began with agents, using an entirely proprietary business model. Their outside teams would go into clients, find the pain points, create or find software that eased the pain, then take control of the client. The name comes from the indestructible crystal balls in Lord of the Rings. (Callback!)
But as the need for AI Agents becomes universal, they can be made more effectively, and deliver value for much less money, with an open approach.
Rather than building a team from inside OpenAI, the plan is for Steinberger to do what he did with OpenClaw. That is to get thousands of outsiders involved in building agents, using open source. As Om wrote to me when I asked about this, Peter “started an open-source project and galvanized hundreds, including a few key ones on the core team.”
Open Source vs. Palantir
In the end, Steinberger’s move may not be aimed at Meta or Anthropic at all.
Its target may be Palantir.
The dirty secret of modern software development is how much developer time is devoted to coordination, to setting priorities, to managing fiefdoms. Projects must be organized, groups developed, milestones set. It’s a team sport, and in a wide-ranging team that means meetings. Whether they’re done in an office or online, most programmer days are spent talking about doing stuff rather than doing stuff.
This happens naturally, as the corporate code base gets bigger, and as the corporation evolves. Note that with AI Agents we’re talking about building a giant code base, a library of code.
Steinberger went with OpenAI because OpenAI is desperate. As I have written OpenAI is in the process of “Rapid Unscheduled Disassembly.” It’s blowing up because it’s run like 18th century Germany, its factions wildly fighting for resources. There’s no leadership.
Steinberger is entering a vacuum. He won’t be in organizational meetings with other OpenAI executives. Instead, open source will build an AI Agent library.
Palantir stock is down 21% this year. Don’t buy the dip.
