The next 2–3 months represent a critical window for the transition from simple chatbots to AI Agents, with value shifting toward framework layers like OpenClaw and Llama. Investors should prioritize companies utilizing efficient architectures like MTP and MIA (e.g., DeepSeek and Xiaomi), as these "cost-performance" leaders are narrowing the gap with US models to just 2–3 months. Expect a 10x surge in demand for inference-specific hardware as persistent agents begin scanning screens and performing long-term tasks. The most reliable ROI over the next 24 months will likely come from AI for Software Engineering, as coding provides the most stable environment for AI self-evolution. Avoid AI applications in sectors with "messy" feedback loops, such as certain quantitative finance models, where the lack of clear rewards hinders model training.

This analysis extracts investment insights from the interview with Luo Fuli (Roderick), a key AI figure at Xiaomi, regarding the massive paradigm shift in AI agents, model training, and the competitive landscape for 2025-2026.

AI Agents & Open-Source Frameworks (OpenClaw)

The discussion highlights a fundamental shift from "Chat" models to "Agent" models. The guest emphasizes that the next 2–3 months will be a "spectacular" period for AI agent evolution.

• OpenClaw Impact: Mentioned as a pivotal product/framework that allows for better agent structure, long-term memory, and self-correction.
• Paradigm Shift: AI is moving from simple dialogue to complex task execution (Agents). This requires a change in model design—moving away from just "scaling" to focusing on "post-training" and "agentic workflows."
• Customization: Unlike closed systems (like Claude 4.6 Ops), open frameworks allow developers to change the entire multi-agent logic, which is seen as a "skill" rather than just a "tool."

Takeaways

• Investment Theme: Look for companies transitioning from "LLM as a chatbot" to "LLM as an OS/Agent." The value is shifting toward the Agent Framework layer.
• Actionable Insight: Open-source ecosystems (like those surrounding Lama or OpenClaw) are becoming "amplifiers" for mid-tier models, allowing them to compete with top-tier closed models by optimizing the agent architecture.

Computing Power & Hardware Allocation (GPUs)

The interview provides a specific "formula" for how leading AI teams are now allocating their hardware resources, which differs significantly from the early days of the AI boom.

• The 3:1:1 Ratio: A "good" allocation of computing cards (GPUs) is now 3 (Pre-training) : 1 (Post-training) : 1 (Research/Experiments).
• Research is the Bottleneck: The guest notes that research requires a massive amount of "cards" to test new paradigms. Without dedicated research cards, teams cannot innovate on the next architecture.
• Inference Demand: As agents become more powerful and "scan" screens or perform long-term tasks, inference demand is expected to rise by 10x.

Takeaways

• Bullish on Inference Chips: While training chips (NVIDIA) remain dominant, the shift toward persistent agents will drive massive demand for Inference-specific hardware and low-cost token production.
• Risk Factor: Companies that do not allocate at least 20% of their compute to "Research" (the 3:1:1 ratio) risk falling behind as the "Pre-training" era hits diminishing returns.

Model Architectures (MTP & MIA)

The technical discussion touches on specific architectural choices that determine the cost and speed of AI models.

• MTP (Multi-Token Prediction): Used in the Flash and Pro series to increase the ability of the model to predict and reason. It significantly lowers the cost per token if the "hit rate" is high.
• MIA (Multi-Head Latent Attention): A structure used by teams like Kimi and DeepSeek. It is highly efficient for KV cache management, making models faster and cheaper to run.
• Hybrid Attention: The guest believes a hybrid structure (combining sliding windows and full attention) is more "elegant" and provides more room for Agents to function effectively in the future.

Takeaways

• Efficiency as a Moat: In 2025, the "winner" may not be the largest model, but the one with the most efficient architecture (MTP/MIA) that allows for the lowest Token Cost.
• Xiaomi (MIMO/Flash): Positioned as the "early Xiaomi phone" of AI—focusing on extreme cost-performance (low price, high speed).

Competitive Landscape: US vs. China

The guest provides a timeline for the gap between Chinese AI models and top-tier US models (like Claude 3.5/4.0).

• The 2-3 Month Gap: The guest claims that leading Chinese teams (mentioning Kimi, Mimo/Xiaomi, and DeepSeek) are only about 2 to 3 months behind the latest US releases (like Claude 4.6 Ops) in terms of actual effect and performance.
• 2026 Outlook: By 2026, the "table stakes" for staying in the game will be the ability to integrate agent structures directly with the model's core.

Takeaways

• Sector Sentiment: Highly bullish on the agility of Chinese "Post-training" teams. They are described as being able to achieve "驚艷" (stunning) results with 1/10th the model size of US competitors.
• Key Players to Watch: DeepSeek (for architectural innovation), Kimi/Moonshot (for long-context), and Xiaomi/Mimo (for infrastructure and cost-efficiency).

AGI Timeline & Investment Risks

• AGI Arrival: The guest estimates AGI is roughly 2 years away.
• Definition of AGI: It is defined here as when AI significantly changes every person's life and work through "persistent" agents that don't need rest and evolve on a single curve.
• Risk - "The Reward Problem": In fields like quantitative finance (factor mining), the "reward" is often unclear. AI models struggle when the feedback loop is messy. Investors should be cautious of AI applications in sectors where the "ground truth" or "reward" is not easily defined.

Takeaways

• Timeline: Investors should look for significant ROI from AI integrations within a 24-month window.
• Focus Area: The most "elegant" path forward is Coding. Code is natural language, it is easily scalable, and it provides a perfect environment for AI to learn and "self-evolve." Companies focusing on AI for Software Engineering are likely the safest bets.