Interview with MiMo Head Luo Fuli: The AI era has entered its second act, "self-evolution" is the most critical event for AGI

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯1824 words

Luo Fuli, head of the Xiaomi large model team, released a 3.5-hour in-depth interview on bilibili, publicly discussing for the first time how AI has fully transitioned from the Pre-train-led Chat era to the Post-train-led Agent era, and naming "Self-evolution" as the most critical event for AGI progress in the coming year. (Previous coverage: Another name change! Moltbot (formerly Clawdbot) has been renamed "OpenClaw" and brings four major updates) (Background supplement: DeepSeek's valuation soared to $20 billion in its first financing round, behind which four core researchers were poached) Born in 1995 in Yibin, Sichuan, this young woman has walked through Alibaba DAMO Academy, High-Flyer Quant, and DeepSeek, and now helms the core large model R&D at Xiaomi—Luo Fuli has long been a legendary figure in the industry, yet she rarely speaks publicly. In April 2026, she finally accepted a 3.5-hour in-depth interview on bilibili (BV1iVoVBgERD). The information density was so high that many AI practitioners exclaimed, "Every minute is packed with substance." In this interview, Luo Fuli stated a core judgment right from the start: the competitive track for large models has shifted. She pointed out that in the Chat era led by Pre-train over the past few years, companies competed on base capabilities and general conversation quality; but entering 2026, the battlefield has fully shifted to the Post-train-led Agent era. Only those who can enable models to act autonomously in complex tasks without manual intervention will have a chance to capture the next wave of the market. Behind this judgment lies a technical node that excites her even more—"Self-evolution." Luo Fuli emphasized in the interview that "Self-evolution" will be the most critical event for AGI progress in the next year. She observed that top models are now able to autonomously optimize in specific tasks and execute stably for 2 to 3 days without the need for constant human intervention. This means that AI systems have begun to possess a certain degree of self-correction capability, and the boundaries are quietly shifting. Luo Fuli also shared a major shift in Xiaomi's internal computing power allocation strategy during the interview. She revealed that the industry's customary computing power ratio used to be roughly Pre-train : Post-train : Inference = 3:5:1, meaning the post-training stage consumed the most resources; however, Xiaomi's current strategy has been adjusted to 3:1:1, significantly compressing the post-training proportion while simultaneously increasing resource investment in the Pre-train and inference stages. The logic behind this is: as the Agent RL Scaling strategy gradually matures, post-training no longer requires massive brute-force computing power, and the efficiency of the model's self-reinforcement has improved significantly. The rebound in inference-side resources reflects the high demand for real-time response capabilities when Agents are deployed. Regarding the "pre-training generation gap" issue that domestic AI teams have long been criticized for, Luo Fuli gave a relatively optimistic judgment: the gap has shortened from 3 years in the past to just a few months. In her context, this figure is not self-aggrandizement, but a strategic turning point—since the Pre-train gap has almost been eliminated, the focus of resource investment must naturally shift toward Agent RL Scaling, which is exactly where Xiaomi is currently placing its main bets. She also discussed her judgment on the Anthropic path and the shock effect triggered by technical variables such as Claude Opus 4.6 and OpenClaw in 2026 across the entire AI ecosystem. She believes that the superposition of these technical variables is accelerating the entire industry's transformation from a "tool" mindset to an "agent" mindset. All talk and no action is not Luo Fuli's style. On March 19, 2026, the MiMo-V2 series she led was officially released, featuring three models at once: the flagship base MiMo-V2-Pro, which adopts a trillion-parameter scale with a hybrid attention architecture of only 42B active parameters, supports million-level context, and achieves an 81% task completion rate; MiMo-V2-Omni, which focuses on full-modal Agent scenarios; and MiMo-V2-TTS, which targets speech synthesis. Most notably, the open-source MiMo-V2-Flash took second place on the global open-source model leaderboard, with an inference speed 3 times that of DeepSeek-V3.2. For a team leader who only joined Xiaomi in November 2025, this scorecard is anything but low-key

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching1 items

2026-04-27

OpenAI releases five constitutions for AGI: AI cannot be monopolized by a few, and sacrifice can be exchanged for greater resilience

Similarity 120%關鍵字 agi同分類 zh

💡 Currently matching via keywords + symbols (MVP) · Will be upgraded to embedding semantic search later

Raw Information

ID:ec0ff7ea48

Source:動區 BlockTempo

Published:2026-04-27 03:10:01

Category:zh_news · Export Category zh

Symbols:Unspecified

Community Votes:+0 / −0 · ⭐ 0 Important · 💬 0 Comments