Tubi Meetup Recap: Redefining Software Engineering as AI Joins Engineering Teams

01 | What Should We Discuss After AI Coding?

Over the past two years tools such as ChatGPT, Claude Code, Codex and Cursor have repeatedly raised developers' expectations of productivity. Rapid emergence of concepts like Vibe Coding, SDD and Harness Engineering has focused attention on "which tool to learn" and "how to write better prompts," creating anxiety about keeping up.

Chen Tian argued that when AI becomes part of long‑term projects, the decisive factor is not a single impressive conversation but whether the next Agent can continue the work correctly after the current dialogue ends. Prompt engineering is temporary; versioned, reviewed, and evolving engineering rules and feedback mechanisms are the foundation for sustained delivery.

In the past year, AI Coding has been the most discussed topic in the software community. The rapid evolution of related ideas and tools fuels FOMO—fear of missing out—both on new tools and on the relevance of accumulated engineering experience. Beyond "which tool to use," the more valuable question is: how should software engineers work in the AI era, and how should software teams be organized?

02 | Three Very Different Projects Illustrate the Same Point

To show how AI can participate in long‑term engineering, Chen Tian presented three real cases: a lightweight Rust AWS local simulator (Rustack), a wearable system that grew from a sneaker sensor to phone and cloud, and a multi‑hundred‑page Chinese technical book that can be automatically built and published. All three rely on clear context, task decomposition, feedback, and project memory.

Chen's talk covered three directions: work projects, personal‑life projects, and AI writing. Though the topics appear different, they all point to the same need: find what truly interests you, stay curious, discover new demands, and try to solve them with technology.

03 | Validation: Generation Is Only the Start, Evidence Determines Value

Previously the focus of AI Coding was "how to make AI write faster." This talk shifted attention to making results correct, reliable, and continuously improvable. Validation is not a final step; it should permeate requirements, design, implementation, operation, and retrospection.

Validation matters because data reliability, correct analysis logic, real business impact, and decision‑supporting output decide the value of generated results.

When errors are recorded, reproduced, classified, and fed back into the system, they become the starting point for the next execution, creating a feedback loop that gives AI engineering compound returns.

04 | From Hands‑On Implementation to Defining Problems, Organizing Collaboration, and Owning Results

As implementation costs fall, engineers' work shifts from template code to problem definition, system design, validation review, runtime feedback, and handling exceptions that templates cannot cover. AI collaboration does not mean a single Agent does everything; research, specifications, implementation, testing and review need clear division, while humans focus on direction, authority, trade‑offs, exception handling, and final responsibility.

Engineering rules and feedback mechanisms are the basis for long‑term delivery.

05 | How Can a 20‑Person Team Serve One Million Users?

Chen Tian posed the imaginative question: with AI coding reducing marginal costs, a small team could serve many niche scenarios that were previously "not worth serving" individually. The key is reusable engineering systems, validation tools, and domain assets rather than simply building more apps.

Instead of a single product for a million users, a team could provide vertical, precise software for each ten‑thousand‑user segment, maintaining hundreds of such products.

This suggests future software teams may stay small while covering more scenarios, focusing on system design, domain understanding, validation capability, and responsibility.

06 | A Meetup as a Shared Discussion About the Future of Work

The audience valued not only the three case studies but also a transferable way of thinking: problem definition, task decomposition, validation, error‑feedback loops, and human judgment apply to software engineering, data science, AI deployment, and broader knowledge work.

Participants debated how Agents should be organized, how AI coding workflows evolve, and what software teams will look like in coming years.

What truly separates people in the AI era is not tool usage but the ability to pose the right questions, decompose complex tasks, decide where to invest, and define high‑quality results.

07 | Closing Thoughts

AI is changing software development at unprecedented speed, but the lasting discussion is never about a specific model or tool—it is about software engineering itself: how we define problems, organize people and Agents, build evidence, handle failure, and take responsibility for outcomes.

When answers become cheaper, the problem becomes more important; when implementation becomes cheaper, direction, judgment, validation, and responsibility become more valuable.

The Tubi Meetup will continue to explore AI, engineering practice, and technical leadership, building an open platform for developers to discuss the future of AI‑driven software engineering.