Podcast Atlas

对话 MiniMax 闫俊杰：M3、10X 计划、10T 模型、和智能的终局

source Updated 2026-07-06 Tags: Podcast, Ai, Coding, Agents, Models

Summary

This Shizilukou Crossing episode records a MiniMax Developer Meetup roundtable hosted by Koji with Yan Junjie, Zhang Jiayuan, He Tao, and Yu Yang. The discussion connects MiniMax M3, AI coding practice, token cost, open-source project governance, financial AI, model scaling, and the long-term question of whether AI can help humans understand AI. Its strongest throughline is that AI raises production speed, but durable value depends on AI Coding Verification, Model Harness Co-Evolution, domain expertise, and human judgment.

Key Claims

  • Yan Junjie frames AI’s core value as productivity improvement and argues that AI coding should give more people production leverage rather than simply reduce the number of practitioners.
  • MiniMax moved from M1 experiments to M2/M2.7 coding and editing focus, with M2.7 reportedly reaching roughly ten trillion tokens of use; MiniMax M3 sets a bigger target but still leaves room for capability improvement.
  • Zhang Jiayuan says MultiCard uses model orchestration rather than one highest-end model for every task, combining MiniMax M3 coding with review or mentor roles from other models while watching costs from Claude Code, Codex, and Cursor.
  • Deerflow began as an open-source attempt to make expensive Deep Research-like capability more accessible with Chinese models, then expanded toward multimodal desktop work such as reports, charts, podcasts, animation, and music.
  • He Tao says Deerflow has over 1K contributors but also faces open-source governance and codebase-quality problems, making successful community maintenance a core engineering concern.
  • Yu Yang argues that financial AI first helps users filter information and understand terminology, but compliance constraints prevent the system from directly giving investment advice or trading for users.
  • The roundtable presents models and agents/harnesses as co-evolving: stronger models enable better agents, while real agent workflows create feedback that shapes what models need to do.
  • Frontier Model Scaling remains constrained by compute, training efficiency, architecture, data quality, and data quantity; Yan Junjie says a 10T-scale model would require experience with 3T-scale training and far more data than is easily available.
  • AI coding has made code generation cheap, but review, validation, tests, long-term maintainability, architecture, and responsibility do not disappear; this extends AI Assisted Software Development Risk into AI Coding Verification.
  • Domain Expert Alignment becomes more important as models enter safety, finance, law, and other high-stakes domains where researchers and engineers need real subject-matter experts.
  • The final discussion moves from product use to intelligence itself: Yan Junjie treats current AI as a black box and sees AI Interpretability By AI as a safety-relevant frontier.

Key Quotes

“AI 的核心是生产力” — Yan Junjie on the main reason AI matters.

“不是单向关系” — the roundtable’s view of model progress and harness/agent progress.

“vibe coding 从来没人说是 vibe engineering” — He Tao on why software engineering does not disappear.

“最重要的思考部分留给自己” — Zhang Jiayuan on using AI without outsourcing human judgment.

Connections

Contradictions

  • No direct contradiction with prior wiki content. The source reinforces the existing view that AI acceleration needs context, workflow, and human judgment, while adding a sharper distinction between code generation and engineering verification.