Lucas Kaiser 解释为何 AI 放缓的说法是错误的，强调推理模型的新范式以相同成本带来更多收益。
Lucas Kaiser explains why the narrative of AI slowdown is wrong, highlighting the new paradigm of reasoning models that deliver more gains for the same cost.
要点 · TL;DR
AI 进步遵循平滑的指数增长，而非放缓。 AI progress follows a smooth exponential, not a slowdown.
带有思维链和强化学习的推理模型推动了近期进步。 Reasoning models with chain-of-thought and RL drive recent gains.
经济压力使焦点从扩大规模转向更小、更便宜的模型。 Economic pressure shifts focus from scaling up to smaller, cheaper models.
核心观点 · Key points
AI 进展遵循平稳的指数增长，而非放缓。 AI progress follows a smooth exponential increase, not a slowdown.
预训练的缩放定律仍然有效，但推理模型在相同算力下带来更多收益。 Pre-training scaling laws still work, but reasoning models offer more gains per compute.
推理模型使用思维链和强化学习来提高正确性。 Reasoning models use chain of thought and reinforcement learning to improve correctness.
后训练的改进（如强化学习和合成数据）推动了近期模型的大部分进步。 Post-training improvements like RL and synthetic data drive most recent model gains.
多模态能力仍落后于文本，但这是改进的关键领域。 Multimodal capabilities still lag behind text, but are a key area for improvement.
部署更小、更便宜模型的经济压力使焦点从规模扩张转移。 Economic pressure to deploy smaller, cheaper models has shifted focus from scaling up.
反共识 · Contrarian takes
AI 放缓的说法是错误的；进展迅速但常被忽视。 The AI slowdown narrative is wrong; progress is rapid but often missed.
前沿模型仍无法完成 5 岁儿童能解决的简单任务。 Frontier models still fail at simple tasks a 5-year-old can solve.
推理模型是“锯齿状”的——在某些领域出色，在邻近领域表现糟糕。 Reasoning models are 'jagged'—excellent in some areas, poor in nearby ones.
蒸馏和经济因素，而非技术限制，减缓了预训练的规模扩张。 Distillation and economic factors, not technical limits, slowed pre-training scaling.
泛化仍是关键挑战；仅靠推理可能不够。 Generalization remains the key challenge; reasoning alone may not suffice.
尽管 AI 进步，翻译行业仍在增长；信任让人类参与其中。 Translation industry grew despite AI advances; trust keeps humans in the loop.
本期章节 · Chapters（共 25）
引言与 AI 进展叙事Introduction and the AI Progress Narrative
教育差距与实际应用Education Gap and Practical Use
推理模型解析Reasoning Models Explained
推理模型及其领域Reasoning models and their domains
预训练与强化学习Pre-training and reinforcement learning
RL 泛化的条件What would it take for RL to generalize?
思维链工作原理How chain of thought works
推理模型中的自纠正Emergent self-correction in reasoning models
多任务模型与阻力Multi-task models and pushback
从谷歌到 OpenAITransition from Google to OpenAI
OpenAI 研究团队组织Research team organization at OpenAI
GPU 访问竞争Competition for GPU access
预训练的未来Future of pre-training
预训练缩放与蒸馏Pre-training scaling and distillation
可解释性与黑箱Interpretability and black box nature
从 GPT-4 到 GPT-5 到 5.1Evolution from GPT-4 to GPT-5 to 5.1
GPT-5.1 后训练改进Post-training improvements in GPT-5.1
模型命名与发布策略Model naming and release strategy
推理模型与思考时间Reasoning models and thinking time
多模态推理失败Multimodal reasoning failures
长时自主编码与压缩Long-running agentic coding and compaction
模型会取代所有人类工作吗？Will models replace all human work?
前沿 AI 研究课题Frontier AI research topics
对多模态与 RL 的兴趣Interest in multimodal and reinforcement learning