AI's Self-Evolution Threatens Work | Generated by AI

Home 2026.04

Question: What did Luo Fuli (罗福莉) say in her interview with Zhang Xiaojun (张小珺) on the WeChat Video Channel / podcast?

Answer:

Overview of the Interview

This was Luo Fuli’s first-ever interview, and also her first long-form technical interview. The 3.5-hour conversation systematically covered the AI earthquake triggered in 2026 by Claude Opus 4.6, OpenClaw, and other technological developments, along with their structural downstream impacts.

Who Is Luo Fuli?

Luo Fuli previously worked at Alibaba DAMO Academy and DeepSeek, and is currently the head of Xiaomi’s large model (MiMo) team, leading the development of the MiMo-V2 model series. She carries many labels online such as “AI genius girl,” a label she dislikes.

Key Technical Judgments and Views She Shared

1. AI Paradigm Shift: From Pre-training to Post-training

The large model war has fully escalated in 2026, entering its second act — a shift from the Chat era dominated by Pre-training to the Agent era dominated by Post-training.

2. On Anthropic’s Path Being Correct

Luo Fuli stated that Anthropic’s technical path is correct — this is now a consensus in the industry. With the path clearer, domestic large model teams in China have entered a phase of accelerated catch-up.

3. AI’s Threat to Her Own Work — “Self-Evolution”

She expressed a sense of crisis even for researchers who personally train models: “I previously thought that the work I do is creative enough, that it wouldn’t be reduced to a ‘Skill’ or a ‘Workflow’. But now I find out — it actually can! So could it train an even stronger model? Could it bootstrap itself into improvement? — This is something that will happen in the next year or two.”

4. On OpenClaw as a “Landmark Agent Framework”

Luo Fuli described OpenClaw as a “landmark Agent framework” and noted that before and after using it was a major dividing line. She spent one week of intense collaboration using only Claude Opus 4.6, saying only it gave her a truly stunning experience.

She also modified OpenClaw’s Agent architecture herself — redesigning its memory system and multi-agent logic — because OpenClaw’s openness allowed this. She pointed out that Claude Code, by contrast, has a black-box Agent architecture design that users cannot modify.

5. On Future Agent Self-Evolution

She said that the next frontier she has not yet seen emerge at scale is: the self-evolution of the framework itself, the self-evolution of the Agent, and the co-evolution between Agents and humans — three distinct layers that have not yet appeared.

6. On Why All Top Models Focus on Code Data

She made a pointed technical observation: “The only data types that can truly reach 128K to 1 million tokens in length are Code data and books. Books have too scattered signals; Code files have stronger inter-file relationships.” This single statement explains why all leading models in 2026 are focusing hard on Code — not because Code tasks themselves are the most important, but because Code is the only long AND dense training data that exists.

7. On Team Management — Using AI for Leadership Advice

One particularly striking moment was that Luo Fuli said she consults OpenClaw (an AI) to learn management methods and how to lead a team. She also mentioned she had required her team to have at least 100 conversations with AI or else quit on their own.

8. On Chinese Models Closing the Gap

She put forward the view that “AI needs to move from language toward interaction with the physical world,” and emphasized that the gap between Chinese models and the world’s top tier has narrowed to just a matter of months.

Public Reception of the Interview

Reviewers noted that while Luo Fuli demonstrated deep thinking on certain issues, her communication style was scattered and non-linear — jumping from topic A to B to C without clear structure. However, the interview was considered information-dense and highly technical.

References:

Back Donate