Let me be blunt: if you're a full-stack developer right now and you're not paying attention to what Tang Jie just wrote, you're sleepwalking into obsolescence.
On July 11, 2026 — Zhipu AI's founder Tang Jie published an internal letter titled "The Wave Has Arrived" to every employee and partner. It is, without exaggeration, the most important thing any AI company leader has written this year. Not because it's inspirational. Because it's a roadmap — and a warning.
I read the full letter. All 3,500 Chinese characters of it. Then I read it again. Then I sat in silence for about ten minutes.
Here's why.
Zhipu is currently sitting on a HK$1 trillion market cap — roughly three times Baidu's valuation and north of Xiaomi. Its stock surged 24.6x since IPO. Its MaaS platform hit 17 billion RMB in ARR, a 60x increase in one year. And what does the CEO tell his entire company to do next?
Forget monetization. Go back to the lab. Go higher.
He called it the "Touch High" (摸高) Plan — and it's a two-year strategic commitment to ignore short-term revenue in favor of climbing three mountains most AI labs are still just talking about.
If you build software for a living, every single thing in this letter concerns you. Let me walk you through why.
Tang Jie didn't write a fluffy memo. He published an architecture document for the next decade of AI. He identified three technical summits that separate today's chatbots from genuine AGI:
Here's the problem with every AI tool you use today: they're sprinters. You ask a question, they answer. You give them a task, they do it for five minutes, then they forget everything.
Tang's first bet is that the next paradigm is models that work across weeks, months, or years — not seconds. He gave this example: a model that "can work tirelessly searching for vulnerabilities in software, essentially learning the thought patterns of a top security expert and then amplifying them through machine endurance."
This isn't science fiction. Anthropic's Claude Opus 4.6 already has a 14.5-hour task completion window. The doubling time for agent task horizons is roughly 196 days, according to Prosus's 2026 State of AI Agents report. Do the math: by mid-2027, we're looking at agents that can work autonomously for weeks.
What happens when a $0.03-per-hour AI agent can do what you do across a two-week sprint? It doesn't replace you on day one. It replaces the you who spends 60% of your time debugging, refactoring, and writing boilerplate. The remaining 40% — architecture decisions, understanding weird business requirements, navigating office politics — that becomes your entire job.
Tang's language here is precise: he's moving from "One-Person Company (OPC)" to "No-Person Company (NPC)." Not as a joke — as a product roadmap.
The vision: thousands of specialized AI agents with different "personalities" and "skills," debating each other, reviewing each other's code, scheduling resources, operating 24/7 without human intervention.
Think about what that means for software development. Right now, you write code and maybe use Copilot or Claude Code as a pair programmer. In 18 months, you'll be the one agent in a swarm of twenty, and your role won't be "write the function" — it'll be "define the objective function."
Memory, Continual Learning, and Self-Judgment — three problems the industry said would require paradigm shifts — are being quietly solved. Long-context RAG approximates memory. Faster model iteration approximates continual learning. Cross-model evaluation is already showing self-critique capability.
These weren't supposed to be solved this quickly.
This is the scariest and most important one. Tang is betting the company on AI training AI. Models writing code. Models cleaning and synthesizing their own training data. Models training the next generation of models.
His justification is brutally pragmatic: "This may consume some compute, yet it saves the most precious resource: human labor and time. In the large-model era, speed is paramount; rapid iteration directly widens cognitive capability gaps."
When Google DeepMind warns that even if individual models never exceed human capability, superintelligence could be "squeezed out" purely through compute scaling — and when Tang cites that same report approvingly — you should understand what's being communicated. The ceiling isn't human. The ceiling is physics.
Here's the part of the story most English-language coverage missed: Tang's "Touch High" strategy only makes sense if you understand what Zhipu just pulled off with coding.
In early 2025, after DeepSeek R1 emerged, Tang made a call: the "chat paradigm" was approaching its ceiling. The next battlefield was coding and reasoning — capabilities that make AI models genuinely useful as workers, not just conversationalists. He reallocated resources. The company went all-in.
The result: GLM-5.2, released in June 2026 under MIT license with a 1-million-token context window. It matches or exceeds Claude Opus 4.8 and GPT-5.5 on multiple benchmarks. And they gave it away — anyone can download, deploy, and commercialize it.
Compare this to Anthropic's journey: $1 billion ARR in December 2024 to $47 billion ARR by May 2026. A 47x increase in 17 months. Claude Code now commands 54% of the AI coding market. The message is unambiguous: coding AI is the closest thing to a money printer the tech industry has ever seen.
Zhipu saw this. They bet on it. They won. And now Tang is saying: that was just the warm-up. Coding was merely the proof of concept that a focused bet on fundamental capability pays off. The real game is the next three peaks.
Before I get to the developer career advice, I need to unpack the philosophy underpinning this letter, because it's the key to understanding why Zhipu moves the way it does.
Tang's framework is three words: 本质 (Essence), 反直觉 (Counter-Intuition), 专注 (Focus).
"Think deeply enough, and you dare to bet against the grain; choose contrarily enough, and you must hold your ground long enough."
This isn't corporate word-salad. Every major Zhipu decision has looked stupid at the time and brilliant in retrospect:
This is what "counter-intuitive" means in practice. When everyone zigs, you zag — but you only earn the right to zag by having thought about the problem at a depth nobody else bothered to reach.
His definition of AGI is worth quoting in full:
"AGI is not the wisdom of any single genius, but the sum of all human wisdom. It should be capable of creating original knowledge on the order of the Theory of Relativity — that is the only standard by which we measure whether the summit has truly been reached."
That's not a product milestone. That's a civilizational goal. And it reframes everything.
Now let's address Steve's question directly: if you're a full-stack developer focused on AI applications right now, what should you actually do?
Let's take this seriously for a moment. If your current work consists primarily of:
Then yes — you should be concerned. Not panicked, but concerned. The economic logic is inescapable: when Anthropic's Claude Code can handle 54% of the coding market and the task horizon doubles every six months, the commoditization of basic software development is accelerating faster than anyone predicted.
But "give up and find a new job" is the wrong framing. The right framing is: pivot to where the puck is going.
Here's what I would do starting Monday morning if I were a full-stack AI developer:
1. SMB AI Integration — The $100 Billion Underserved Market
While everyone fights over enterprise contracts and developer tools, small and medium businesses are drowning in manual processes that AI could automate tomorrow. These businesses don't need another SaaS subscription. They need someone who speaks both "GPT-5.5" and "I run a 12-person plumbing supply company."
The moat here is deep: understanding a business's actual workflow, translating it into agent orchestration, and maintaining the human relationship. AI can't replicate the trust of a local business owner who knows you'll pick up the phone.
Tang's "No-Person Company" concept applied downward: help a 50-person company become a 10-person company with AI, and charge for the transition.
2. Content Creation at Scale — But Not the Way You Think
The content game has inverted. AI can generate infinite text, images, and video. The scarce resource is no longer production — it's taste, curation, and distribution strategy.
As a developer, your advantage isn't that you can prompt Midjourney better than a designer. Your advantage is building the pipelines: automated research → drafting → image generation → publishing → analytics. The person who builds the factory wins, not the person who runs one machine.
3. Workflow Automation — Become an Agent Orchestrator
This is the highest-leverage pivot. Tang's letter makes clear that the future is multi-agent systems. Someone needs to design these agent societies, define the communication protocols, set the evaluation criteria, and handle the failure modes.
Full-stack developers are uniquely positioned here. You understand systems thinking. You've dealt with API integrations, error handling, and state management. Multi-agent orchestration is just distributed systems with LLMs as the compute nodes.
Start experimenting now: set up two coding agents that review each other's output. Build a pipeline where one agent writes tests while another writes implementation. You'll learn more in a weekend than most "AI consultants" know after a year.
4. Domain-Specific AI That Can't Be Commoditized
Here's Tang's actual quote: "Domain-specific large models are a false proposition." He's right — and he's also wrong, depending on what "domain" means.
If "domain" means "slightly fine-tuned on legal documents," yes, that's dead. But if "domain" means "deeply integrated into a regulated industry's actual workflow with proprietary data that can't leave the building," that's a different story entirely.
Healthcare compliance. Insurance claims processing. Construction permit management. These aren't solved by better base models. They're solved by people who understand both the regulation and the integration surface area.
If I were allocating learning time, here's my prioritization:
| Priority | Skill | Why |
|---|---|---|
| P0 | Multi-agent orchestration | The next 24 months are about agent societies, not single agents |
| P0 | Evaluation & monitoring infrastructure | When agents run for weeks, you need observability |
| P1 | Business process decomposition | The hard part isn't coding; it's breaking "run my business" into agent-executable subtasks |
| P1 | Synthetic data generation pipelines | Self-play and data factories are the fuel for the next wave |
| P2 | Traditional full-stack development | Still valuable but rapidly commoditizing — treat as table stakes, not differentiator |
| P2 | Prompt engineering as a standalone skill | Being absorbed into system design — not dead, but not a career |
The theme is clear: move upstream. Stop being the person who implements. Start being the person who designs what gets implemented.
I'd be doing you a disservice if I didn't present the bear case for Zhipu's strategy, because it's directly relevant to your career decisions.
Problem 1: A $105 million revenue run rate against a $128 billion valuation is... aggressive. The market is pricing in a future that hasn't arrived yet. If the AI coding revenue curve flattens — or if open-source models commoditize Anthropic-style economics — the whole sector reprices.
Problem 2: The "Touch High" plan asks public-market shareholders to tolerate a two-year monetization pause. That's an eternity in quarterly capitalism. Tang has moral authority because his previous bets paid off, but moral authority doesn't pay dividends.
Problem 3: Fully Self-Training sounds great until you think about model collapse. If AI trains on AI-generated data recursively, what happens to output quality after 10 generations? Nobody has a definitive answer yet, and Zhipu is betting billions on the optimistic case.
Problem 4: The safety governance question is genuinely unresolved. Tang commits tens of billions to "mechanical interpretability" — making black-box models transparent. Noble goal. But if the leading US frontier labs are already delaying releases due to risk concerns, the gap between "we want transparent models" and "we have transparent models" might be measured in years, not quarters.
Here's my actionable advice for Steve — and any full-stack developer reading this:
Step 1: Audit your current work. For one week, track every task. Categorize each as: (a) an AI coding agent could do this today, (b) an AI coding agent could probably do this by end of 2027, or (c) this genuinely requires human judgment. If more than 50% falls into category (a), accelerate your pivot.
Step 2: Pick one of the four focus areas above (SMB AI integration, content pipelines, agent orchestration, regulated domain AI) and build a weekend project. Not a tutorial. A real thing you can show people.
Step 3: Learn one agent framework deeply. LangGraph, CrewAI, AutoGen — pick one and build something where at least three agents collaborate on a non-trivial task. The difference between "I've read about multi-agent systems" and "I've debugged a five-agent pipeline at 2 AM" is the difference between a spectator and a practitioner.
Step 4: Get uncomfortably close to an actual business problem. The developers who survive this transition won't be the best coders. They'll be the ones who can sit across from a business owner and say: "Here are the 47 things you do manually that AI agents can handle, and here's what it would cost."
Tang Jie ended his letter with a sentence that should be pinned to every developer's wall:
"不登顶，就是失败。" — "Not reaching the summit is failure."
He's talking about AGI. But the principle applies to your career. The summit isn't "be a better coder than the AI." That summit is already behind us. The new summit is: design systems that make AI coding agents productive, reliable, and safe at a scale no human team could match.
The wave is here. You can either be the one building the surfboard or the one wondering why the shore disappeared.
What I'd tell Steve directly: Don't quit AI application development. But stop thinking of yourself as a developer who uses AI and start thinking of yourself as an AI systems architect — someone who designs processes where AI agents are the workers and you're the foreman. The market for that role is about to be enormous, and 97% of current developers are too busy optimizing their React components to notice.
Research sources: Zhipu AI internal letter (July 11, 2026), Geopolitechs translation & analysis, Forbes/Sacra Anthropic ARR data, Prosus State of AI Agents 2026, Eigent.ai GLM-5.2 technical analysis, Pandaily Zhipu coverage.
