AI 未来：硬件、规模扩展与入门指南

AI Future: Hardware, Scaling, and Getting Started

格雷格·布罗克曼 Greg Brockman · Y Combinator · 2017-11-08 · 约 60 分钟 · 原视频 ↗

本期速览 · Overview

专家讨论即将到来的 AI 硬件加速、模型规模扩展带来的新行为，以及进入该领域的实用建议。

Experts discuss upcoming hardware acceleration for AI, scaling models to unlock new behaviors, and practical advice for entering the field.

要点 · TL;DR

硬件将加速 AI 模型扩展，催生新行为。
Hardware will accelerate AI model scaling, enabling new behaviors.
AI 进步主要来自工程纪律，而非 ML 科学。
Most AI progress comes from engineering discipline, not ML science.
游戏因复杂性和可扩展性成为 AI 理想试验场。
Games are ideal testbeds for AI due to complexity and scalability.

核心观点 · Key points

AI 硬件将比预期更快提升，使模型规模扩张并带来全新行为。
Hardware for AI will get faster than expected, enabling model scaling and new behaviors.
AI 工作大部分是工程而非 ML 科学；优秀工程是进步的关键。
Most AI work is engineering, not ML science; good engineering is key to progress.
机器学习系统通过自动化复杂任务来放大人类程序员的能力。
Machine learning systems leverage human programmers by automating complex tasks.
通过记分板和每周目标进行迭代训练，驱动指数级进步。
Iterative training with a scoreboard and weekly goals drives exponential improvement.
游戏是 AI 的理想试验场，因其复杂性、人类基准和可扩展性。
Games are ideal testbeds for AI due to complexity, human baselines, and scalability.

反共识 · Contrarian takes

理解现有方法及其局限比发明新问题更有价值。
Understanding existing methods and their limits is more valuable than inventing new problems.
AI 研究者可能比其他人更先失业，因为 AI 会自动化研究本身。
AI researchers may lose their jobs before others, as AI automates research itself.
ML 中的 bug 代价高昂且难以调试；编写无 bug 代码至关重要。
Bugs in ML are costly and hard to debug; writing bug-free code is critical.
非技术人员可以通过自我教育伦理问题来帮助 AI 领域。
Non-technical people can help AI by educating themselves on ethical issues.
长时间工作不是目标；对工作的热爱才能维持高效。
Working long hours is not a goal; passion and love for the work sustain productivity.

本期章节 · Chapters（共 27）

未来硬件与模型扩展 Future Hardware and Model Scaling
AI 中待探索的潜力领域 Promising Under-Explored Areas in AI
AI 硬件创新 Hardware Innovations for AI
AI 入门指南 Getting Started in AI
项目选择与游戏决策 Project selection and game choice
游戏环境工程构建 Engineering the game environment
搭建 Dota 2 机器人基础设施 Building the Dota 2 bot infrastructure
从回放中进行行为克隆 Behavioral cloning from replays
强化学习训练机器人基础 Basics of training a bot with reinforcement learning
参赛决策 Decision to compete in the tournament
机器学习贡献与项目进展 Machine learning contribution and project progress
Dota 2 机器人输给职业选手 Dota 2 bot loss to pro player
为机器人观测空间添加缺失特征 Adding missing features to bot observation space
通宵调整运行实验 All-night surgery on running experiment
与职业选手激烈迭代的一周 Intense week of iteration against pros
发现诱敌策略 Discovery of the baiting strategy
赛前紧急修复 Emergency fix before the match
最终准备与通宵奋战 Final preparation and all-nighter
部署与训练时间线 Deployment and Training Timeline
AI 研究中的调试与代码质量 Debugging and Code Quality in AI Research
非技术技能：谦逊与工程纪律 Non-Technical Skills: Humility and Engineering Discipline
非技术人员如何助力 AI 初创 How Non-Technical People Can Help AI Startups
工作生活平衡与热情 Work-Life Balance and Passion
AI 炒作与最后的人类工作 AI Hype and Last Human Job
电子游戏与 AGI 的关联 Relevance of Video Games to AGI
游戏作为 AI 试验场 Games as a testbed for AI
如何参与 OpenAI How to get involved with OpenAI

阅读全文双语转录 →