140. 对姚顺宇的4小时访谈：请允许我小疯一下！在Anthropic和Gemini训模型、技术预测、英雄主义已过去

Summary

This 张小珺Jùn｜商业访谈录 episode interviews Yao Shunyu / 姚顺宇 about his move from theoretical physics into frontier model training at Anthropic and Google DeepMind. The technical core is that model progress now depends less on public benchmark ranking and more on problem definition, data construction, feedback signals, system engineering, and organization execution. The episode adds Long-Horizon AI and ML Coding as next-stage frontiers: models that can work across longer task chains, selectively manage context, and participate in AI research by writing code, running experiments, analyzing results, and proposing new hypotheses.

Key Claims

• Yao Shunyu / 姚顺宇 argues that public benchmarks make OpenAI, Anthropic, and Gemini look closer than their real workflow behavior: Claude feels stronger in tool use and agents, Gemini in reasoning and daily use, and Codex has narrowed the gap in pure coding.
• Coding broke out early because it has unusually clear feedback, large high-quality code data from GitHub, and relatively shared standards for good output; this makes AI Coding Verification more tractable than open-ended product, writing, or physical-world tasks.
• He thinks model progress has not meaningfully slowed; apparent slowdown can come from saturated benchmarks, bugs, data or token-horizon choices, and poorly defined experiments rather than a final scaling wall.
• His 2026 technical slogan is “train with finite context, use as infinite context”: models should be trained within bounded context but used through long interaction, selective forgetting, retrieval, and context management.
• Long-Horizon AI and continual learning are treated as closely related because a model’s current context can function like temporary weights; the key question is what to remember, retrieve, compress, or discard during extended work.
• ML Coding is framed as the next version of coding agents: AI should help with machine-learning research itself, including experiment code, result analysis, failure diagnosis, and hypothesis generation.
• Modern model training know-how is inseparable from infrastructure and large systems; tips that work in one training stack may not transfer without understanding data, sampling, compute, stability, and production constraints.
• Anthropic is described as unusually top-down and execution-focused during the Claude coding push, but also as culturally stressed as it scaled; Google DeepMind is described as more attractive for broad research freedom and learning across directions.
• He argues that the individual-hero era in frontier language models has mostly passed: model progress now depends on reliable people, shared goals, and system-wide responsibility more than isolated genius.
• He sees Chinese and U.S. model gaps narrowing in some areas, with Doubao voice and Seedance video as strong product examples, but he does not treat them as proof that the whole frontier paradigm has changed.
• Robotics and broad multimodal generation are described as earlier than language/coding: useful feature engineering exists, but the scalable route has not yet reached a GPT-1-like moment.
• He rejects the idea that any one AI company can stop AI progress by owning the frontier; if one company stops, others continue, so AI safety cannot rely on single-company monopoly control.

Key Quotes

“train with finite context, use as infinite context” — Yao’s slogan for long-horizon model use.

“个人英雄主义时代已经过去” — his summary of frontier language-model work after scale-up.

“靠谱、做事细、负责” — the traits he values most in AI researchers.

Connections

• Yao Shunyu / 姚顺宇 — guest and source of the episode’s model-training, organization, and career claims.
• Anthropic, Claude Code, Google DeepMind, and Gemini — main model-company and tool contexts.
• OpenAI, Codex, Cursor, Manus, and Open Cloud — peer products used to compare agent and coding-tool directions.
• Long-Horizon AI, ML Coding, Context Engineering, AI Coding Verification, and Model Workflow Fit — main technical concepts extended by the episode.
• Frontier Model Scaling, AI Organization Design, Long-Chain AI Competition, and Model Provider Tool Competition — strategy and organization concepts qualified by Yao’s inside account.
• ByteDance, Doubao, and Seedance — domestic AI product examples discussed through voice, video, and product execution.
• World Models, Vision Language Action Models, Embodied AI, and Video Models — physical and multimodal AI directions he treats as promising but still pre-scale.
• Research Taste and Problem Definition In Research — research-method themes reinforced by his physics-to-AI transition and emphasis on defining good tasks.

Contradictions

• No direct contradiction with prior wiki content.
• The source qualifies earlier Frontier Model Scaling doubts by arguing that “scaling wall” stories can be premature when experiments have bugs, unclear task definitions, token-horizon limits, or data mistakes.
• The source adds tension with broad World Models optimism by saying the term is often undefined and that robotics and multimodal generation have not yet found a scalable path comparable to language-model pretraining.