There have been a few examples of single-person projects hitting it big in the AI age, such as OpenClaw, but the future could be stranger still.
Tang Jie, founder of Zhipu AI, the Beijing lab behind the record-breaking open-weight GLM 5.2 model, thinks the one-person company is already a transitional phase. In a lengthy note supposedly posted on Rednote to mark the company’s new strategic push, Tang argues that the endpoint isn’t a single founder running a business with an army of AI tools, but a company with nobody running it at all. He calls it the NPC, or the no-person company.
Tang’s reasoning starts from where the industry currently sits. He credits the one-person company, or OPC, as a real phenomenon, pointing to how far a single individual armed with capable agents can now stretch. But he frames it as a stepping stone rather than a ceiling, arguing that “technology is advancing faster than expected” and that the natural extension of an agent doing a founder’s work is a network of agents doing everyone’s work, with no human required to hold it together. His argument rests on three capabilities he says are converging faster than most people assume: models that can hold a task for weeks or months instead of a single session, systems of agents that can coordinate with each other independently, and models that can train and improve themselves without waiting on human engineers. Put together, he says, these three unlock a company that plans, executes, reviews its own output and corrects course, all without a person in the loop.
He is careful to frame this as a technical prediction rather than a product roadmap. Tang breaks it down into what he calls three problems that once looked unsolvable — memory, continual learning and self-evaluation — and says progress on long-context handling and retrieval-augmented generation has already made a workable form of memory possible, that faster model release cycles are functioning as a crude version of continual learning, and that frontier models are starting to show early self-evaluation. None of these are fully solved in his telling, but he treats the direction as settled, describing agents that will eventually “independently debate, collaborate, review code, and allocate resources” the way a self-driving system handles traffic, without a driver checking in.
The rest of the note is Tang’s account of Zhipu’s own history and the strategy the company is committing to next. He traces the company’s lineage back to an academic search project he built on a single desktop computer in 2006, and to the decision in 2021 to commit to a large model with hundreds of billions of parameters over a year before ChatGPT existed, using both as evidence that Zhipu’s pattern is to make contrarian bets years ahead of the market. He then describes the company’s Hong Kong listing this January not as an endpoint but as a reset, with Zhipu choosing to redirect its resources back into foundational research rather than treat the IPO as a finish line.
From there, Tang lays out what he’s calling the “Touch High” initiative, a two-year push built around four areas: long-horizon task execution, autonomous multi-agent systems, self-training models that generate and learn from their own synthetic data, and what he describes as safety governance built into the model’s foundations rather than added as a patch afterward. He pairs this with a commitment to keep releasing Zhipu’s models openly, pointing to GLM 5.2’s MIT license and its million-token context window as evidence that the company intends to make frontier-level capability broadly available even as it chases harder technical ground. The note closes on Tang’s framing of this as a responsibility as much as an opportunity, arguing that once a technology is powerful enough to reshape how civilization works, treating safety as optional stops being viable.