AI2 发布完全开放的 Almo 3 模型，挑战闭源 AI

AI2 Launches Fully Open Almo 3 Models, Challenging Closed-Source AI

内森·兰伯特 Nathan Lambert · Matt Turck 的 MAD 播客 · 2025-11-20 · 约 88 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

AI2 发布 Almo 3 系列，包含完全开放的模型、数据和配方，包括推理模型，美国开源 AI 与中国 AI 巨头竞争加剧。

AI2 releases Almo 3 family with fully open models, data, and recipes, including reasoning models, as US open-source efforts compete with Chinese AI powerhouses.

要点 · TL;DR

• AI2 完全开放的 Almo 3 以完全透明挑战闭源 AI。
  AI2's fully open Almo 3 challenges closed-source AI with full transparency.
• 中国开源模型填补了 Meta 转向 Llama 后美国的开源真空。
  Chinese open models fill the US open-source vacuum after Meta's Llama shift.
• 后训练是一门艺术；SFT、DPO 和 RL 工具对未来模型至关重要。
  Post-training is an art; SFT, DPO, and RL tooling are critical for future models.

核心观点 · Key points

• 预训练提供昂贵的初始化；强化学习摘取低垂果实，但两者缺一不可。
  Pre-training provides a costly initialization; RL captures low-hanging fruit, but both are essential.
• 开源 AI 需要完全透明：数据、中间检查点和配方，而不仅仅是权重。
  Open-source AI needs full transparency: data, intermediate checkpoints, and recipes, not just weights.
• 中国开源模型如 Qwen 和 DeepSeek 因战略开放和市场动态而占据主导地位。
  Chinese open models like Qwen and DeepSeek dominate due to strategic openness and market dynamics.
• 后训练是一门艺术：监督微调和直接偏好优化带来巨大收益，而强化学习工具对未来的模型至关重要。
  Post-training is an art: SFT and DPO yield large gains, while RL tooling is critical for future models.
• AGI 将通过混乱的共同进化到来，而非突然的奇点；进步是平稳但受限的。
  AGI will arrive via messy co-evolution, not a sudden singularity; progress is smooth but constrained.

反共识 · Contrarian takes

• Meta 的 Llama 转向后，美国开源真空由中国实验室填补，而非美国实验室。
  The US open-source vacuum after Meta's Llama shift is filled by Chinese labs, not American ones.
• Qwen 模型中的预训练数据污染可能夸大强化学习研究结果，质疑可重复性。
  Pre-training data contamination in Qwen models may inflate RL research results, questioning reproducibility.
• 直接偏好优化的收益来自选中和拒绝示例之间的对比，而非绝对答案质量。
  DPO gains are from contrast between chosen and rejected examples, not absolute answer quality.
• 长上下文扩展更多关乎架构而非数据；糟糕的设置无法通过好数据修复。
  Long-context extension is more about architecture than data; bad setup cannot be fixed by good data.
• 基于可验证奖励的强化学习比基于人类反馈的强化学习更难，因为存在内核不匹配和数值不稳定等系统问题。
  RLVR is harder than RLHF due to systems issues like kernel mismatches and numerical instability.

本期章节 · Chapters（共 24）

• 引言与模型发布 Introduction and Model Release
• 数据与预训练细节 Data and Pre-training Details
• 长上下文数据与开源 Data for long context and open source
• AI 开源：风格与承诺 Open source in AI: flavors and commitment
• 2025 回顾：开源 AI 大事记 2025 recap: key events in open source AI
• 中国开源模型生态 Chinese open-source model ecosystem
• 中国开源策略与美国回应 China's Open Source Strategy and US Response
• OLMo 起源与 AI2 开源文化 Origins of OLMo and AI2's open-source culture
• AI2 项目概览 AI2 Projects Overview
• 预训练与后训练概述 Pre-training vs Post-training Overview
• 预训练方法与数据整理 Pre-training methodology and data curation
• 中训练与尾部修补 Mid-training and tail patching
• 长上下文：重要性与技术挑战 Long context: importance and technical challenges
• 长上下文对模型性能的影响 Impact of long context on model performance
• 预训练与长上下文 Pre-training and Long Context
• 监督微调与蒸馏 Supervised Fine-Tuning and Distillation
• SFT 与 RL 及数据创建 SFT vs RL and Data Creation
• DPO 与偏好调优 DPO and Preference Tuning
• 实验验证与数据工作 Experiment Validation and Data Work
• RLVR 历史与定义 RLVR History and Definition
• RLVR 与 RLHF 对比 RLVR vs RLHF
• 后训练中 RL 的挑战 Challenges in RL for Post-Training
• 复杂性与物理约束 Complexity and Physical Constraints
• 社会贡献闭幕词 Closing remarks on societal contribution

阅读全文双语转录 →