Igor Babuschkin 分享他在 DeepMind、OpenAI 和 xAI 的经历，以及他创立 River AI 专注于企业和消费者个人 AI 的愿景。
Igor Babuschkin discusses his journey through DeepMind, OpenAI, and xAI, and his new venture River AI focused on personal AI for companies and consumers.
要点 · TL;DR
编码代理在 2024 年底达到变革性拐点，变得不可或缺，预示着更广泛的 AI 变革。 Coding agents hit a transformative threshold in late 2024, becoming indispensable and signaling broader AI changes.
AI 的下一个前沿是从可验证领域转向不可验证领域，需要新的训练方法。 Next AI frontier is moving beyond verifiable domains into non-verifiable areas, requiring new training methods.
个性化 AI，根据个人偏好定制并本地运行，是下一个重大演进，带来控制和隐私。 Personal AI, tailored to individual preferences and running locally, is the next major evolution for control and privacy.
核心观点 · Key points
编码智能体在2024年底达到了一个变革性的临界点，使其变得不可或缺，并预示着更广泛的变革即将到来。 Coding agents reached a transformative threshold in late 2024, making them indispensable and signaling broader changes ahead.
AI的下一个前沿是从编码和数学等可验证领域转向不可验证领域，这需要新的训练方法和更长的时间跨度。 The next frontier for AI is moving beyond verifiable domains like coding and math into non-verifiable areas, requiring new training methods and longer time horizons.
个性化AI，根据个人偏好定制并本地运行，是下一个重大演进，提供控制权、隐私和更好的用户体验。 Personal AI, tailored to individual preferences and running locally, is the next major evolution, offering control, privacy, and better user experiences.
专有模型提供商面临来自开源模型的日益增长的压力和规模扩张的收益递减，迫使他们创新或失去优势。 Proprietary model providers face increasing pressure from open-source models and diminishing returns on scaling, forcing them to innovate or lose ground.
基于实际奖励信号（如用户幸福感）的端到端训练对于开发有效的个性化AI智能体至关重要。 End-to-end training on the actual reward signal, such as user happiness, is crucial for developing effective personal AI agents.
AI的未来取决于保持人机共生，需要更深层次的对齐和集成，以确保AI服务于人类繁荣。 The future of AI depends on maintaining human-machine symbiosis, requiring deeper alignment and integration to ensure AI serves human flourishing.
反共识 · Contrarian takes
AI进步的最大瓶颈不是数据或算力，而是处理长时间跨度和不可验证奖励的新思路。 The biggest bottleneck for AI progress is not data or compute, but new ideas for handling long time horizons and non-verifiable rewards.
闭源模型提供商处于艰难的商业地位，受到开源进步和潜在监管限制的挤压。 Closed-source model providers are in a difficult business position, squeezed by open-source progress and potential regulatory restrictions.
AI的未来可能分化为少数人使用的超级智能AI和大众使用的日常AI，它们的优化目标不同。 The future of AI may bifurcate into super-intelligent AIs for a few and everyday AIs for the masses, with different optimization goals.
基于平均用户数据训练模型已经过时；相反，模型应该针对每个个体进行个性化，打破一刀切的假设。 Training models on average user data is outdated; instead, models should be personalized to each individual, breaking the one-size-fits-all assumption.
在个人设备上本地运行前沿模型是可行的，并且可能比基于数据中心的推理提供更低的延迟和更好的隐私。 Running frontier models locally on personal devices is feasible and could offer better latency and privacy than data-center-based inference.
美国应该训练自己的最佳开源模型，以避免依赖中国实验室，后者可能通过许可证或后门施加控制。 The US should train its own best open-source model to avoid dependence on Chinese labs, which could exert control through licenses or backdoors.
本期章节 · Chapters（共 29）
开场Introduction
小说灵感Inspiration for the fiction
智能体下一步Beyond coding: next steps for agents
灵感与智能体AIInspiration and Agentic AI
公司愿景与三大赌注Company vision and three bets
硬件赌注动机Motivation for hardware bet
构建个人模型方法Approaches to building personal models
个性化方法Approaches to Personalization
未来模型差异Future Model Differences
企业模型所有权Enterprise Model Ownership
领域专家与前沿实验室Domain Experts vs. Frontier Labs
后训练RL中的泛化Generalization in Post-Training RL
后训练专业化与企业价值Post-training Specialization and Enterprise Value
开源模型与地缘政治Open Source Models and Geopolitical Concerns
惊喜时刻与早期AI兴趣Surprising Moments and Early AI Interest
从物理到AIFrom Physics to AI
星际争霸项目及其影响StarCraft Project and Its Impact
转向编程与AlphaCodeTransition to Coding and AlphaCode
推理工作与OpenAIReasoning Work and OpenAI
DeepMind与OpenAI对比Comparing DeepMind and OpenAI
开源哲学与模型分发Open Source Philosophy and Model Distribution
对工程师的尊重与工作文化Respect for Engineers and Work Culture