At Code Class 2026, Ethan (Yuanming) Hu, Founder and CEO of Meshy AI, gave a talk titled “The 500-Day Countdown to AGI.” In his view, the most consequential changes of the AI era are not confined to model capability itself. They are also rewriting how founders work, how organizations are managed, and the basic logic by which a founder evaluates technology, deploys talent, and allocates organizational resources.

Hu begins with his own frontline practice of running ten coding agents, discussing how an AI-native founder can raise agent throughput and where the real limits of vibe coding lie. From there he turns to how management itself is changing in the AI era: when everyone can produce far more with AI’s help, how should an organization redefine talent density, the value of middle management, KPIs and OKRs, and founder mode? The question worth pressing on is this — in the narrowing window before AGI arrives, how does a founder build AI-native thinking, organization,

The following has been lightly edited for length while preserving the original meaning.

I’m Ethan Hu, founder and CEO of Meshy AI, and what follows is a set of practical, CEO’s-eye notes I prepared specially for Code Class 2026, Source Code Capital’s annual entrepreneurs’ conference. Over the past six months, reality has revised my thinking several times over. I once asserted flatly that “you should resist the urge to review code written by A”; after running into enough walls, I have retreated to “core code must be reviewed by a human.” My judgment that AI delivers a “100x productivity gain” turned out to hold only at one particular layer of the work — apply it at the wrong layer and you will misjudge things badly. Or take the question of whether a CEO should write code at all: I have flip-flopped on that five or six times. So this is not a set of conclusions. It is a first-hand account of a practice still being iterated on.

There are two things I want to cover today: first, how my vibe coding practice has evolved; second, what will change about management in the AI era, the latter being what I most want to discuss with every founder in this room.

Key points：

• The prevailing view among Silicon Valley’s AI leaders is that humanity is expected to reach AGI by the end of 2027 — which means we get 1.4% closer every week. Perhaps 500 days later all of us will have far more freedom to do what we really love upon AGI’ advent.
• The proper gauges on one’s AI literacy are no longer token consumption, but one’s agents’ continuous unsupervised working time and the number of agents managed simultaneously.
• On most tasks, AI outstrips humanity in terms of learning and evolving speeds.
• AI’s greatest advantage is that it could stick to plain facts without any emotional scruples, while irrationality and reality shunning are precisely humanity’s greatest weaknesses.

The AI-native founder: starting with managing ten AIs

A quick introduction. Our company builds the world’s leading 3D generative AI product, transforming texts, images, or even a bare idea into a 3D model. Meshy AI has notched an ARR of tens of millions of dollars, holds a leading share in overseas markets, and has more than 10 million users worldwide. We are also honored to be part of Code Class community.

Let me start with the first part: how I manage ten coding agents. I barely took a break over Chinese New Year this year, spending most of that stretch on vibe coding. In the process, I found we could leverage AI coding tools to code even better and improve output. There’s only one fundamental question: how to raise the throughput of a coding agent?

I adopt an iteratively progressive approach with every step underpinned by a same plain management intuition.

Step one: move from a GUI editor to command-line based tools, because that makes it far easier to build on top of — a CLI tool is remarkably easy to modify, while a GUI-based agent is hard to do that.

Step two: find a good vehicle (Sandbox). Whenever I use a Sandbox, there is one thing I always do: skip the prompts and enable all the permissions, otherwise I have to interact with AI very frequently to authorize every file creation and deletion and every command, giving permissions every 10 seconds, and people certainly don’t want that.

But the biggest problem with granting all permissions is that the agent may do something genuinely dangerous. While prototyping, I repeatedly found that it had wiped my database — it had mistaken my production environment for a test environment and cleared it once testing was done, erasing the entire database. Therefore I later wrote another agent to back the database up every hour, otherwise I wouldn’t dare trust it with too many tasks at all. Put simply, we have to give the agent a good environment — one where it can range freely, but with a boundary that keeps it from going too far off the rails.

Step three: build a multi-process management system where one person can run several AIs. I could handle perhaps three or four agents without auxiliary tools — one planning, one executing, one gathering information, and that’s all, because human only have very limited attention.

Step four: make sure the agent can close the loop on its own. It has to be able to write, test, and see the results of its own work independently. This is very similar to managing people: you should set clear goals and concrete ways of evaluating them, offering context rather than trying to control. A great deal of philosophy underlying human organization management works just as well for managing agents.

A few other things followed. I built a cloud-based web system so I can manage these agents by voice from my phone, because you don’t want to leave agents idle as soon as you close the laptop — they need to run 24/7. You can’t type while walking down the street, but you can talk, and the best route to put those fleeting ideas into practices is to simply direct to an agent vocally.

My view about whether to review AI generated code has changed considerably lately. I used to hold that AI-written code needed no review — all we have to do is to give it an environment where it could debug itself, but after running into enough obstacles recently I have taken a step back: critical code entails human review; while non-critical one can be left to the AI.

I have also found AI especially useful personally in terms of cognition improvement. Learning something used to mean digging through mountains of materials myself. Now I can have an agent that gathers, organizes, visualizes and even presents the materials in an interactive way. In half an hour it can sort out the thread of a field, and then I could continue to spend another half hour actually learning. In Meshy AI we call this “living on AI’s liquid diet” — don’t chew through solid food; drink it, and you absorb it faster. There is, of course, one thing we should stay wary of: information security. Without proper information isolation, it is theoretically possible for someone in one department to learn another department’s sensitive information simply by asking the AI a question. Thus in the AI era, sandboxing and permission isolation are surely among the most important issues any AI-native company faces.

Image courtesy of Ethan (Yuanming) Hu, Founder and CEO of Meshy AI: an art-history learning site he built with AI.

Why do we keep misjudging AI?

With the practice part covered, I want to share a more fundamental line of thinking: why is it so easy for us to underestimate AI, and just as easy to overestimate it?

Image courtesy of Ethan (Yuanming) Hu, Founder and CEO of Meshy AI: an art-history learning site he built with AI.

Let’s start with underestimating. I keep asking myself — how did I not realize until late 2025 that AI could code so proficiently with such a formidable thing like Claude Code? The answers boil down to three points, as I see it.

First, as human beings we still subconsciously believe we ought to be better than AI — and that idea has to be discarded. What I have found is that on most problems, people are not so distinguished from AI; When human codes, they do nothing more than inputting a command line, thinking, and looking things up — all of which AI can do, too. Once agents plug all those tools into an AI’s brain, the range of problems it can solve is not fundamentally different from ours.

Second, I didn’t take a dynamic view on the value of technology. Back in 2023, people regarded large models as merely a toy making plenty of mistakes; but sheer scale and input then produced a breakthrough, which has since solved a great many concrete problems users face.

Third, my understanding of the technology wasn’t deep enough. Only when you understand a technology with an enough depth do you begin to see its internal imperfections — and once those are fixed, the technology becomes more capable. I am now willing to spend a great deal of time on technical detail, because nowadays a founder’s in-depth understanding of technology has arguably never mattered more.

Let’s turn to overestimating, which is the other side of the same question: how should we objectively measure what AI can and cannot solve for us? Take coding as an example, I assume it is a three-layer process. The first layer is typing — something AI unquestionably does extremely well. I competed in Olympiad in Informatics in high school and could write several thousand lines of code in five hours, once ranking among the top ten in China; but there is no way I can compete with AI on keystroke speed, because it doesn’t type at all. AI accelerates this layer by roughly 100 times. The second layer is software architecture. AI has been fed with plenty of good architectures, so it can do this too, but the acceleration may only be 3x. The third layer is to determine what the user problems the software is aimed at and which direction it’s heading for — and at this layer I believe AI’s acceleration is close to zero. Why? Because anything valuable is necessarily new, anything new is by definition absent from the historical training data, and AI is trained on precisely that historical data.

Image courtesy of Meshy AI.

So the primary reason people misjudge AI is that they mistake AI’s 100x acceleration of the first layer for the third. We have to recognize that AI’s acceleration factor differs enormously depending on which layer of problem it is solving. It does not accelerate the whole process evenly; instead, it has moved the bottleneck up from “typing” to “choosing a direction” — and the latter is exactly the layer where training data is scarcest. That realization determines where founders should spend their time.

How would organizations change when AI makes individual output 10 times higher?

I, as a founder myself, have been dwelling on and iterating my ideas about this very question. From my two favorite books on management, I have distilled two propositions I still agree with. The first is Andy Grove’s: output is the only measure of management. A manager’s job is to maximize the team’s output, and that still holds in the AI era. The second is Ram Charan’s: management is getting results through others.

On that basis, a founders’ responsibilities remain unchanged — depict a clear vision, organize outstanding talent, and realize the vision through effective execution. However, AI has almost rewritten every specific managing maneuver.

Setting direction: you still have to work out the bigger picture, but AI has already settled many of the smaller ones for you. A product manager used to be able only to verbalize their idea, but now they can build a prototype with AI and hand it to engineers to productionize — the direction has become concrete.

Enpowering the team: 90% of a team’s output may now come from AI, and much of my own work is actually about paving the road for it — offering good infrastructure so it can work unobstructed.

Developing your people: the biggest change is that on most tasks, AI outstrips human in terms of learning and advancing speeds — so your talent bench should be made up of “people who are good at using AI,” and developing them is actually teaching them how to use AI.

The most crucial point is that AI has introduced an entirely new form of competition. AI’s greatest competitive advantage is that it can stick to the plain facts without any emotional scruples — and being unpragmatic and irrational is precisely humanity’s greatest weakness. A great many problems arise because we escape the facts, or because we are so careful of people’s feelings when communicating that after all that talking the other side still has no idea what we meant. With AI you can be blunt and direct. There was an amusing line going around online: “Don’t say thank you to AI — if every ChatGPT user in the world said thank you once a day, it would consume something like the annual electricity use of several thousand households over a year.” You can probably work the real figure out from token API prices. Nowadays an open mind and a fact-based way of thinking have never been more important.

Organizational boundaries are being erased by AI as well. In many companies engineers used to fall into front-end and back-end categories; now they don’t, because most front-end work is already largely done by AI. The line between product managers and engineers is likewise harder and harder to draw. Thus getting something done now hinges chiefly on talent density — you need the best people, each able to deliver results independently and with strong empathy across functions. It may be necessary to reassess and refine the responsibilities of a number of roles, especially those with high “AI exposure”. Founders, too, have to keep evoling to stay competitive.

A few new conclusions about management in the AI era

Following that line of thinking, I have several more concrete — and more pointed — conclusions I would like to put to my fellow founders.

First, KPIs and OKRs may no longer be applicable in this era. I’ve asked friends working at the most cutting-edge companies in Silicon Valley whether they have OKRs and how much weight they attach to them, and the answer is either that they don’t have OKRs, or that the circumstances change immediately days after OKRs are set.

Then what matters more? I think it is culture and talent density. Seismic industrial shifts make it almost impossible to measure an individual’s or a team’s contribution with quantitative metrics; relative to building out a comprehensive system of rules, trust in and reliance on people become more important. On the other hand, AI is such a nascent and fledgling industry that you could hardly find enough qualified talents if you require candidates for every position to hold a solid track record. At that point you have to be bold about delegating to capable, self-driven young talents and trust them.

Second, the work of middle management will change significantly. I have always insisted that every engineering manager at our company be able to write codes manually or at minimum have a profound understanding in business details, without which one could hardly manage. And much of what managers used to spend their time on — relaying messages up and down the chain, putting together slide decks — can be handed to AI outright, so those can solve problems hands-on can better thrive in this era. Management becomes even easier thanks to the flat founder mode (Jensen Huang has 40 direct reports) AI that takes on a large share of the management work. One more point: use AI to speed up the flow of information. Most cross-team coordination questions come down to “is this feature done,” “when does it ship,” and “who owns it.” We are now pushing hard to hand all of that to AI — it summarizes day-to-day chat into a knowledge base that offers instant answers on demand, which makes the whole organization more transparent and more energetic. Sensitive information, of course, still has to be properly isolated.

Image courtesy of Ethan (Yuanming) Hu, Founder and CEO of Meshy AI: from his Code Class 2026 keynote deck.

Third, the standards for evaluating AI literacy have changed. I don’t quite acknowledge the interview question “how many tokens you have consumed” that people often discuss online — the emphasis on token consumption is prone to create the myth that “token consumption automatically translates into output”, while even the efficiency of every single token varies enormously. There’re two better questions, as I see it. First: how long can your agent run effectively and continuously with no human intervention? An entry level agent requires human permissions every five seconds; an intermediate one runs smoothly for half an hour unsupervised with a Sandbox; the most advanced agent can work independently and effectively for several hours. Second: how many agents can you manage at once? A beginner manages only one agent, those who adroitly use skills manage three, and skills developers manage even more. We’ve also run a compulsory company-wide vibe coding competition. There was some resistance at first, but by the end, we found that everyone demonstrated incredible creativity — colleagues in every department had automated a great deal of their own workflows.

Fourth, there’re several things that stay enduringly true. How will founder mode change in the AI era? I think, first, that founders can learn faster with AI, and second, that managing a team becomes easier. However, a few fundamental management principles will not change. A clear vision — it is only when an organization has clearly articulated its mission and problems it aim to solve can it attract the cream of talent. The importance of talent — “surround yourself with the best people” is always the best strategy; looking back over these past few years, every success could be ascribed to the right hires and every failure came from getting that person wrong. The importance of culture — a nurturing culture lets people thrive while a toxic culture drives good people away. These three things stay true and withstand changes of time.

Image courtesy of Ethan (Yuanming) Hu, Founder and CEO of Meshy AI: from his Code Class 2026 keynote deck.

Let me close the talk with something thrilling. Some of my Silicon Valley friends have put the question “when will AGI arrive” separately to the helmsmen at several top AI companies. The prevailing view was around the end of 2027. That means we get 1.4% closer every week — and perhaps 500 days later all of us will have far more freedom to do what we love upon the advent of AGI. Thereforee, this is a tremendously exciting era, when every one of us can bring our full strength to bear, sprint through the milestone of 2027, and work out the most important questions of 2030.