AI 不只比智商，WAIC 和 Kimi K3 透露了什么新竞争

Summary

This Keji Luandun episode uses WAIC and a hands-on [[KimiK3|Kimi K3]] coding test to argue that AI competition is moving from “which model is smartest” toward deployment, cost, stability, model routing, and commercial closure. The hosts treat exhibition booths, embodied-intelligence demos, small AI applications, and speech-to-text services as evidence that real advantage depends on [[AIDemoDeploymentGap|demo-to-deployment discipline]], [[AIApplicationLayerMoat|application moats]], and [[AIInferenceCostStructure|unit economics]]. The source also clarifies [[OpenWeightReleaseBoundary|open weights versus open source]] and frames strong frontier models as more valuable for building tools and handling unknowns than for every runtime call.

Key Claims

• WAIC showed a stronger push toward application, industrialization, monetization, and commercial closed loops, but exhibition visibility still did not prove deployability or buyer demand.
• Embodied AI progress appeared uneven: robot supply chains and domestic component substitution looked stronger than autonomous task planning, with many demonstrations still relying on teleoperation or narrow behavior.
• Small AI applications face weak moats when generic AI can help competitors copy the visible feature; domain know-how, customer understanding, accumulated data, and workflow integration matter more.
• [[KimiK3|Kimi K3]] was good enough to build a podcast-host dialogue agent from requirements, but the source emphasizes its slowness, high token use, and need for fit-based deployment rather than treating it as a universal best model.
• Model Routing Cost Control should route simple intent classification, routine work, and cheaper runtime tasks away from the most expensive model while reserving frontier models for planning, coding, review, tool-building, and unknown problems.
• The episode-dated claim that Kimi K3 would open weights on 2026-07-27 is explicitly separated from full open source: downloadable/self-hostable weights do not imply released training code, data, or process.
• Speech-to-text cost reduction came more from engineering optimization and batch processing than from expensive fine-tuning alone, showing why users may prefer stable, cheap, good-enough AI over maximum single-point accuracy.

Key Quotes

“开放权重不等于开源” - the episode’s release-governance distinction.

“客户在哪” - the recurring application-business test.

“顶级模型仍有价值” - the source’s boundary around frontier models.

Connections

• WAIC - exhibition setting used to read the industry’s shift toward landing and monetization.
• [[KimiK3|Kimi K3]] and Kimi - hands-on model/product case for AI coding, cost, latency, and open-weight discussion.
• AI Demo Deployment Gap - main exhibition-to-production pattern added by the source.
• Open Weight Release Boundary - open-weight versus open-source boundary made explicit by the K3 discussion.
• Top Model Build Runtime Split - source’s claim that frontier models may be most valuable while building tools, with cheaper models running many delivered workflows.
• Speech To Text Cost Optimization - concrete example where engineering and batching lowered one-hour transcription cost from about 0.6 yuan to under 0.1 yuan.
• AI Commercialization Pressure, AI Application Layer Moat, and AI Startup Unit Economics - business discipline behind the episode’s critique of application hype.
• Model Workflow Fit, Model Routing Cost Control, AI Engineering Thinking, and AI Coding Verification - AI coding and model-selection practices reinforced by the K3 test.
• Embodied AI, Robot Teleoperation and Remote Takeover, and Robot Demo Authenticity - robotics branch sharpened by the WAIC embodied-intelligence comments.
• Persistent Agent Memory - project knowledge files and daily summaries used to preserve work context across AI coding sessions.

Contradictions

• No direct contradiction found. The source qualifies optimistic AI-application and embodied-AI narratives by separating demonstrations from deployable autonomy, customer pull, cost structure, and stable runtime operation.