三位创始人探讨 AI 扩展定律、代币成本以及开源与专有模型之争，预测 2-3 年内推理能力增长 9 万倍。
Three founders discuss AI scaling laws, token costs, and whether open-source or proprietary models will dominate, with predictions of 90,000x inference growth in 2-3 years.
要点 · TL;DR
推理计算将在 2-3 年内增长 9 万倍，推动 AI 能力飙升。 Inference compute will surge 90,000x in 2-3 years, driving AI capabilities.
开源模型现在仅落后前沿模型 3-6 个月。 Open-source models are now only 3-6 months behind frontier models.
AI 取代工作的速度才是真正的问题，而非取代本身。 AI job displacement speed, not displacement itself, is the real issue.
核心观点 · Key points
推理算力将在 24-36 个月内增长 9 万倍，推动 AI 能力激增。 Inference compute will increase 90,000x in 24-36 months, making AI capabilities surge.
开源模型正在缩小与前沿模型的差距，目前落后 3-6 个月。 Open-source models are closing the gap with frontier models, now 3-6 months behind.
AI 写作目前质量不高，但通过适当的评估和技能文件可以媲美人类。 AI writing currently lacks quality, but with proper evals and skill files it can match humans.
AI 导致就业替代的速度才是真正的问题，而非替代本身。 The speed of AI-driven job displacement is the real problem, not the displacement itself.
中国的开源 AI 策略利用了硬件优势和软件商品化。 China's open-source AI strategy leverages hardware dominance and commoditized software.
反共识 · Contrarian takes
每年花费 10 万美元在 token 上，就能像 2028 年的普通公民一样生活。 Spending $100k/year on tokens lets you live like a normal citizen in 2028.
英伟达可能被严重低估，而非高估，低估了几个数量级。 Nvidia may be vastly underpriced, not overpriced, by several orders of magnitude.
美国并未与中国真正竞争；台湾将缓慢和平统一。 The US is not in real competition with China; Taiwan will slowly reunite peacefully.
AI 焦虑普遍存在，但前沿实验室的研究人员抑郁而非欢欣。 AI anxiety is widespread, but frontier lab researchers are depressed, not jubilant.
通用 AI 模型将击败专用模型，使垂直 SaaS 在 2027 年前过时。 General AI models will beat specialized ones, making vertical SaaS obsolete by 2027.
人类欲望不可替代；使用 AI 的人效率更高，而非被取代。 Human desire is irreplaceable; AI users become more productive, not displaced.
本期章节 · Chapters（共 30）
开场与嘉宾介绍Introduction and Guests
AI 成主导话题AI as the Dominant Topic
YC 初创与 AI：代币成本与算力扩展YC Startups and AI: Token Costs and Compute Scaling
AI 在数学与创造力上的进展AI Progress on Math and Creativity
人机对比：超越比较Human vs machine: beyond comparison
实用 AI：成本是瓶颈Practical AI: cost is the bottleneck
AI 接管后人类何去何从What will humans do when AI takes over?
AI 访问与控制AI Access and Control
AI 写作与人类互动AI Writing and Human Interaction
AI 写作与技能文件AI Writing and Skill Files
AI 创造力与编解码器AI creativity and codecs
开源与闭源模型Open source vs closed source models
开源模型追赶中Open source models catching up
中文预训练与蒸馏Chinese pre-training and distillation
算法突破与人才Algorithmic breakthroughs and talent
安全与泄露Security and leaks
软件商品化与硬件Software commoditization and hardware
硬件与软件商品化Hardware and Software Commoditization
优势时间收缩Time Contraction of Advantages
AI 写作与智能体通信AI Writing and Agent Communication
政治与 AIPolitics and AI
谷歌产品与文化问题Google's Product and Culture Issues
创业生态与模型专业化Startup ecosystem and model specialization