OpenAI president Greg Brockman discusses the new GPT-5.5 model, its capabilities, and what it means for the company's competitive position.
要点 · TL;DR
GPT-5.5 从语言模型转向解决实际任务的有用 AI 助手。 GPT-5.5 shifts from language models to useful AI assistants solving real-world tasks.
OpenAI 在预训练、中期训练和强化学习全栈投入以提升实际效用。 OpenAI invests across pre-training, mid-training, and RL to improve real-world usefulness.
算力将成为驱动问题解决的稀缺资源，从日常任务到治愈阿尔茨海默病。 Compute will become the scarce resource driving problem-solving, from daily tasks to curing Alzheimer's.
核心观点 · Key points
GPT-5.5 标志着从语言模型向解决实际任务的有用 AI 助手的转变。 GPT-5.5 marks a shift from language models to useful AI assistants that solve real-world tasks.
OpenAI 在整个技术栈（预训练、中期训练、强化学习）上投资，以提升实际应用价值。 OpenAI invests across the entire stack—pre-training, mid-training, RL—to improve real-world usefulness.
模型更直观，需要更少的明确指令，减少了对详细提示工程的需求。 The model is more intuitive, requiring less explicit instruction, reducing the need for detailed prompt engineering.
OpenAI 的商业模式很简单：租用算力并以正利润率转售，受益于杰文斯悖论。 OpenAI's business model is simple: rent compute and resell it with positive margin, benefiting from Jevons paradox.
带有安全措施的迭代部署是最大化收益和降低强大 AI 风险的关键。 Iterative deployment with safeguards is key to maximizing benefits and mitigating risks of powerful AI.
反共识 · Contrarian takes
提示工程并未消亡；随着用户用更少努力获得更多价值，它反而更加活跃。 Prompt engineering is not dead; it becomes more vibrant as users get more value with less effort.
开源蒸馏无法复制全部能力；真正的护城河是制造机器的机器。 Open-source distillation does not replicate full capability; the real moat is the machine that makes the machine.
尽管价格更高，智能的回报如此之大，以至于需求将超过供给，而非萎缩。 Despite higher prices, intelligence returns are so large that demand will outstrip supply, not shrink.
在安全措施下公开发布强大模型，比限制少数可信测试者更安全。 Releasing powerful models publicly with safeguards is safer than limiting access to a few trusted testers.
智能体应被赋予高度自主权，但治理和监督必须随责任按比例扩展。 Agents should be given high autonomy, but governance and oversight must scale proportionally with responsibility.
算力将成为驱动问题解决的稀缺资源，从日常任务到治愈阿尔茨海默病。 Compute will become the scarce resource driving problem-solving, from daily tasks to curing Alzheimer's.
本期章节 · Chapters（共 12）
引言与 GPT-5.5 概览Introduction and GPT-5.5 Overview
两年研究历程与未来改进Two-Year Research Process and Future Improvements