Rofu Li's Bold Prediction: AGI Will Arrive in Two Years, Work Models Will Be Disrupted

Rofu Li’s Bold Prediction on AGI

Rofu Li, the head of Xiaomi’s model team, recently made a striking assertion during an interview with “Language is the World”: “I feel that AGI (Artificial General Intelligence) should be achievable within two years.” Known as a “genius girl” and leading the MiMo-V2 series models, she shared a series of bold viewpoints during the three-and-a-half-hour conversation.

Core Predictions on AI Paradigm Shift

AI Replacing Research Work
Rofu Li noted that by providing recent research context to an AI agent, it could replicate the researcher’s growth path and discuss issues like a researcher. This led her to a radical conclusion: future models may train themselves and even develop stronger models independently.
Disruption of Work Models
She believes that the most immediate impact of AGI will be on work models, rather than daily life. As work generates productivity value, she predicts that many will be forced to abandon traditional work methods.
2026 as a Year of Accelerated Productivity Change
Rofu Li defines 2026 not as the “Agent Year” but as the “Era of Accelerated Productivity Change,” where many will realize that numerous tasks no longer require their direct involvement.

OpenClaw and Agent Framework

OpenClaw as a Revolutionary Agent Framework
Rofu Li describes OpenClaw as a revolutionary agent framework, emphasizing its role in defining how systems communicate with models.
Enhancing Model Action Capabilities
A good agent framework addresses the action limitations of models, providing persistent memory, messaging channels, and task management.
OpenClaw’s Open Source Value
Unlike black-box models like Claude Code, OpenClaw allows users to modify the framework based on their scenarios, enhancing its adaptability.
Collective Intelligence Accelerating Framework Evolution
She highlights that collective intelligence can rapidly iterate frameworks, significantly enhancing development speed.
Raising the Ceiling for Mid-Tier Models
OpenClaw can compensate for many model shortcomings, allowing mid-tier models to approach top-tier performance in various scenarios.
Cost Considerations in Model Development
Rofu Li argues that as models improve, the importance of agent frameworks will persist due to cost constraints.

Competition Landscape and Agent Technology Path

Transition to the Agent Era
Rofu Li believes the competition among large models has entered a new phase, shifting from chat-based interactions to complex agent frameworks.
Minimal Performance Gap Between US and Chinese Models
She asserts that Chinese models with a 1T parameter base are only 2-3 months behind Claude Opus 4.6.
Focus on Post-Training for Agents
The next competitive edge lies in effective post-training for agents, particularly in reinforcement learning.
1T Parameter Models as Entry Tickets
Rofu Li views 1T parameters as essential for achieving top agent performance, emphasizing the need for rapid post-training and paradigm shifts.
Changing Resource Allocation
She predicts a shift in resource allocation ratios from 3:5:1 to 3:1:1 in the agent era.
Increased Demand for Computational Resources
As agent frameworks evolve, the demand for computational resources will intensify due to accelerated research and experimentation.
Surge in Inference Demand
With the strengthening of agent frameworks, inference demand is expected to explode, necessitating efficient storage solutions.

Human-Agent Interaction and Data Forms in the Agent Era

Misconceptions About Current Agents
Rofu Li criticizes many so-called agents as merely complex chat systems, lacking true industrial-grade capabilities.
Shifting Interaction Paradigms
In the agent paradigm, human interaction will focus on high-level judgments rather than minor code corrections.
Skills as Key Data Supplements
Skills will play a crucial role in enhancing agent performance by allowing humans to contribute organizational knowledge.
Long Context as a Fundamental Capability
The MiMo-V2 design emphasizes the importance of long context for maintaining state and dependencies in complex tasks.
Two Paths for Agent Development
Rofu Li categorizes agent development into high-value task replacement and universal model creation for broader benefit.
Multi-Agent Collaboration
While multi-agent collaboration will occur, she remains skeptical about its current effectiveness in improving task completion rates.

Multimodal, Open Source, and Entrepreneurial Ecosystem

Pricing Logic Shift
She argues that pricing should reflect the value generated by models in agent tasks rather than just inference costs.
Multimodal Training’s Impact on Intelligence
Rofu Li has become cautious about the role of multimodal training in enhancing intelligence after observing its limited impact on benchmarks.
Open Source as a Path to Accelerate AGI
She believes that open source will facilitate collaboration across various sectors necessary for AGI development.
Blurred Lines Between Model and Product Companies
The integration of models into agent frameworks will increasingly blur the lines between model and product companies.
Smaller Entrepreneurial Companies
Rofu Li predicts that the rise of agents will enable smaller teams or even individuals to establish companies.

Rofu Li’s unconventional management philosophy emphasizes equality and collaboration, fostering an environment where creativity thrives. She believes that the right environment is more crucial than experience, valuing potential over past achievements.