Jeff Dean 回顾了他关于 AI 达到初级工程师水平的预测，讨论了长期运行的智能体系统的崛起，并分享了对定制推理硬件及 TPU 起源的见解。
Jeff Dean reflects on his prediction that AI is now at a junior engineer level, discusses the rise of longer-running agentic systems, and shares insights on custom inference hardware and the TPU's origin.
要点 · TL;DR
AI 现已能胜任初级工程师编码，且智能体系统进步速度超预期。 AI now handles junior-level coding, and agentic systems are improving faster than expected.
长时运行的智能体与上下文工程将解锁远超当前小时级任务的难题。 Long-running agents and context engineering will unlock harder problems beyond current hour-scale tasks.
真正的瓶颈是推理而非训练；专用低精度芯片能大幅降低延迟。 The real bottleneck is inference, not training; specialized low-precision chips cut latency dramatically.
核心观点 · Key points
AI 现在确实具备了初级工程师级别的编程能力，而且智能体系统的进步速度超过预期。 AI is now genuinely capable at junior-engineer-level coding, and agent systems are improving faster than expected.
下一个大趋势是通过紧密的实验、评估和整合循环，实现机器学习和科学的自动化。 The next major trend is automating ML and science via tight loops of experimentation, evaluation, and integration.
让智能体在分解后的任务上运行数天或数周，将解决比当前小时级使用更困难的问题。 Running agents for days or weeks on decomposed tasks will unlock much harder problems than current hour-scale uses.
上下文工程——技能、清晰规格和检索——是让强大模型真正完成实际工作的核心。 Context engineering — skills, clear specs, and retrieval — is central to getting capable models to do real work.
面向推理的专用硬件，采用低精度和低数据移动，是扩展智能体系统的关键。 Specialized hardware for inference, with low precision and low data movement, is key to scaling agent systems.
创业公司应瞄准前沿模型目前几乎完全失败的问题，而不是部分能解决的问题。 Startups should target problems where frontier models currently fail almost completely, not where they partly work.
反共识 · Contrarian takes
大多数所谓模型限制实际上是能量和数据 IO 瓶颈；数据移动的能耗约是计算的 1000 倍。 Most so-called model limitations are actually energy and data-IO bottlenecks; data movement costs about 1000x more than compute.
我们可以用不可靠的组件构建有用的 AI 系统，质疑晶体管绝不能出错这一假设。 We can build useful AI systems from unreliable components, questioning the assumption that transistors must never fail.
蒸馏论文曾被拒绝，被认为不太可能产生重大影响，却成为了最常用的模型压缩技术之一。 Distillation was rejected as unlikely to have significant impact, yet became one of the most used model-compression techniques.
像 AlphaFold 这样的专用垂直模型，如今能在高价值领域击败通用前沿模型。 Specialized niche models, like AlphaFold, can beat general frontier models in high-value domains today.
稀缺技能不再是编码，而是品味：选择让数百个智能体去解决哪些问题。 The scarce skill won't be coding but taste: choosing which problems to direct hundreds of agents to solve.
真正的瓶颈是推理而不是训练；定制低精度芯片可将延迟降低 20 到 30 倍。 Inference, not training, is the real bottleneck; custom low-precision chips can cut latency by 20-30x.
本期章节 · Chapters（共 35）
简介Introduction
初级工程师预测The Junior Engineer Prediction
2027年预测A 2027 Prediction
下一个内存适配时刻The Next Fits-in-Memory Moment
长期运行的智能体Long-Running Agents
粗略估算与TPUNapkin Math and the TPU
TPU起源与粗略估算TPU origin and napkin math
AI延迟数字Latency numbers: AI edition
能耗与数据移动Energy vs data movement
批处理与推理Batching and inference
AI作为压缩问题AI as a Compression Problem
上下文工程Context Engineering
精通上下文工程Becoming Exceptional at Context Engineering