"The Wave Has Arrived": Zhipu Co-Founder Tang Jie's Letter to Staff

When Zhipu's shares dropped more than 19% following a lockup expiry, co-founder Tang Jie didn't respond with a reassuring investor call. He wrote a letter to the entire company instead, and the message was blunt: stop chasing application revenue, go all-in on foundation-model research, and do it for two straight years under a plan he's calling 'Touch High.' No quarterly-earnings comfort here, just a bet that the next leap in raw model capability matters more than near-term monetization.

Tang leans hard on history to justify the move. He points back to 2021 and 2022, when Zhipu poured resources into hundred-billion-parameter scale and produced GLM-130B roughly eighteen months before ChatGPT existed — a decision that looked reckless at the time and vindicated later. That's the pattern he wants staff to trust again: bet against consensus, then hold the position long enough for it to pay off. Zhipu listed on the Hong Kong Stock Exchange on January 8, 2026, and instead of treating that as a finish line, Tang framed it as a reset button pointing straight back at research.

The letter lays out four technical peaks standing between current models and something resembling AGI. First is long-horizon task execution — models that can plan and grind through work spanning weeks or months rather than answering one prompt at a time. Second is autonomous agent systems, swarms of specialized agents running and coordinating continuously, what Tang calls the shift from a one-person company to a fully automated one. Third is self-training, where models generate their own synthetic data and effectively teach themselves once human-generated data runs thin. Fourth, and the one Tang spends the most ink on, is safety — not bolted on after the fact, but baked into a model's value function from day one, alongside a commitment of tens of billions toward interpretability research meant to make model decisions auditable rather than opaque.

On open source, Zhipu isn't hedging. GLM-5.2 shipped under an MIT license with no entity restrictions, a million-token context window, and a spot in the top three on the Artificial Analysis leaderboard — free for anyone to download, deploy, or resell. Tang's argument is that real safety can't come from locking technology behind walls; it has to come from broad, visible oversight, which is why he frames open weights and the push toward AGI as two hands working the same problem rather than opposing strategies.

It's a strange moment to make this pitch. Zhipu's stock is sliding, investors want revenue, and the letter essentially tells them to wait. But Tang seems to be betting that the market pressure is a short-term nuisance next to what he sees as an irreversible shift in the ceiling of machine intelligence — and that being early and open, again, is worth the discomfort.