We're sunsetting PodQuest on 2025-07-28. Thank you for your support!
Export Podcast Subscriptions
cover of episode Ep 55: Head of Amazon AGI Lab David Luan on DeepSeek’s Significance, What’s Next for Agents & Lessons from OpenAI

Ep 55: Head of Amazon AGI Lab David Luan on DeepSeek’s Significance, What’s Next for Agents & Lessons from OpenAI

2025/2/19
logo of podcast Unsupervised Learning

Unsupervised Learning
AI Deep Dive AI Chapters Transcript
People
D
David Luan
Topics
David Luan: DeepSeek 的成功并非偶然，而是模型效率提升和智能提升同步进行的结果。降低成本并不意味着减少智能应用，反而会促进更多智能应用的出现。未来大型模型的训练会继续追求更高的智能，而效率的提升则会通过内部优化来实现，最终以更低成本提供给客户。 仅仅依靠下一个token预测不足以实现AGI，需要结合强化学习和搜索等其他机器学习范式来实现知识的发现和利用。当前模型的泛化能力比人们想象的要强，虽然在某些特定任务上的表现可能略有差异，但这只是模型发展过程中的小问题。 构建可靠的AI模型需要建立一个可靠的模型生产工厂，而不是仅仅关注算法本身。AlphaGo 等案例已经证明了模型具备原创性思维的能力，LLM 的局限性被夸大了。当前的AI代理模型虽然潜力巨大，但在可靠性和实用性方面仍有很大的提升空间。构建有用的AI代理的关键在于可靠性，而不是仅仅追求炫酷的演示效果。当前AI代理模型最大的挑战在于端到端可靠性不足，需要大量人工干预。 将基础多模态模型转化为大型行动模型需要解决两个问题：一是工程问题，即如何以模型可理解的方式向模型展示其能力；二是研究问题，即如何教会模型进行规划、推理、重新规划和遵循用户指令。当前AI模型与浏览器和程序的交互方式缺乏创意，未来需要更具创造性的交互方式来提升效率。 AI代理领域的关键里程碑是能够在训练阶段赋予代理任何任务，并在几天后达到100%的完成率。AGI 的定义是：能够完成人类在电脑上完成的任何有用任务，并且学习速度与人类相当。AGI 的普及速度可能会受到社会因素的限制，例如人们对新技术的接受程度和适应能力。未来会出现专业化模型，这并非出于技术原因，而是出于政策原因，例如数据安全和隐私保护等。 仅仅依靠简单的规模扩大并不能解决所有问题，还需要解决其他关键的技术挑战。高质量的数据标注在模型训练中仍然至关重要，但其作用会逐渐被强化学习所取代。过去一年，我对团队文化建设的重要性有了更深刻的认识。我改变了对AI技术长期差异化竞争的看法，认为不同领域的突破并非必然具有累积效应。数字代理的可靠性问题的解决可以为物理代理的研发提供借鉴和经验。世界建模可以解决在没有明确验证器或模拟器的情况下，如何训练AI模型的问题。

Deep Dive

Chapters
This chapter analyzes the market's reaction to DeepSeek, highlighting the initial panic and subsequent recovery. It discusses the model's efficiency and its implications for the future of AI development, including the increased consumption of intelligence despite cost reduction. The discussion touches upon the commoditization of previous levels of intelligence as newer, more complex models emerge.
  • Initial market reaction to DeepSeek involved panic and a stock market crash.
  • The market initially misunderstood the implications of increased efficiency in AI models.
  • Increased efficiency leads to increased consumption of intelligence, not decreased consumption.
  • AI use cases are categorized in concentric circles of complexity, with each circle requiring increasingly smarter models but commoditizing previous levels of intelligence.

Shownotes Transcript

David is an OG in AI who has been at the forefront of many of the major breakthroughs of the past decade. His resume: VP of Engineering at OpenAI, a key contributor to Google Brain, co-founder of Adept, and now leading Amazon’s SF AGI Lab. In this episode we focused on how far test-time compute gets us, the real implications of DeepSeek, what agents milestones he’s looking for and more.

[0:00] Intro[1:14] DeepSeek Reactions and Market Implications[2:44] AI Models and Efficiency[4:11] Challenges in Building AGI[7:58] Research Problems in AI Development[11:17] The Future of AI Agents[15:12] Engineering Challenges and Innovations[19:45] The Path to Reliable AI Agents[21:48] Defining AGI and Its Impact[22:47] Challenges and Gating Factors[24:05] Future Human-Computer Interaction[25:00] Specialized Models and Policy[25:58] Technical Challenges and Model Evaluation[28:36] Amazon's Role in AGI Development[30:33] Data Labeling and Team Building[36:37] Reflections on OpenAI[42:12] Quickfire

 

With your co-hosts: 

@jacobeffron 

  • Partner at Redpoint, Former PM Flatiron Health

 

@patrickachase 

  • Partner at Redpoint, Former ML Engineer LinkedIn

 

@ericabrescia 

  • Former COO Github, Founder Bitnami (acq’d by VMWare)

 

@jordan_segall 

  • Partner at Redpoint