前语

在这次对话中，Dario Amodei作为Anthropic的创始人和首席科学家，与Lex Fridman就AI发展的各个层面进行了深入探讨，包括技术进步、风险管理以及未来的应用场景。这次对话为我们提供了对AI当前状态及未来可能性的独特见解，有助于更好地理解AI如何改变我们的生活和工作。

对话小结

Dario Amodei在这次对话中探讨了人工智能的未来，特别是关于Anthropic公司开发的AI模型Claude的发展。他讨论了AI技术的快速进步，以及这些进步如何可能在未来几年内达到人类博士甚至专业水平。Dario强调了AI系统在处理复杂任务方面的能力，比如编程和解决生物学问题，并且提到了AI在医疗、科学研究以及日常生活中的潜在应用。

Dario还讨论了AI安全性的问题，包括AI可能带来的风险和滥用问题。他提出了一个“负责任的扩展政策”（Responsible Scaling Policy，RSP），旨在确保随着AI模型能力的增长，相应的安全措施也能跟上。他解释了如何通过不同的“安全级别”（Safety Levels，ASL）来评估和应对AI的风险。

此外，Dario还讨论了AI模型在理解语言、图像和进行推理方面的能力，以及如何通过“宪法AI”（Constitutional AI）和“机制可解释性”（mechanistic interpretability）等技术来提高模型的透明度和安全性。他强调了AI模型在理解复杂概念时的能力，并探讨了未来AI可能达到的超级智能水平。

对话中也提到了AI在编程领域的应用，以及AI如何改变程序员的工作方式。Dario预测，AI将在未来几年内显著提高编程任务的自动化水平，同时也将改变编程工作的性质。

最后，Dario和其他专家还讨论了AI模型的多模态能力，即模型如何在处理文本、图像和其他类型的数据时展现出的能力。他们还探讨了AI模型的可解释性，以及如何通过不同的技术来理解和解释AI模型的行为。

以下是访谈中的29个要点：

1. Scaling Laws 扩展规律需要三个要素同步增长：network规模、数据量和计算资源

"In particular, linear scaling up of bigger networks, bigger training times and more and more data. So all of these things, almost like a chemical reaction."

2. 突破数据限制的可能方案：合成数据生成、自我对弈学习、反思和推理模型

"If you think about what was done with DeepMind's AlphaGo Zero, they managed to get a bot all the way from no ability to play Go whatsoever to above human level, just by playing against itself."

3. 计算资源规模预测

"right now... frontier model companies... are operating in roughly 1 billion scale... next year we're going to go to a few billion, and then 2026, we may go to above 10 billion. And probably by 2027, their ambitions to build hundred billion dollar clusters."

4. 进展速度超出预期

"At the beginning of the year, I think the state of the art was 3 or 4%. So in 10 months we've gone from 3% to 50% on this task. And I think in another year we'll probably be at 90%."

5. Anthropic竞争策略：Race to the Top战略

• 不是要成为唯一的"好人"，而是通过树立榜样带动整个行业向好
• 目标是创造向上而非向下竞争的激励机制 "Race to the Top is about trying to push the other players to do the right thing by setting an example."

6. Anthropic竞争策略：机制可解释性(Mechanistic Interpretability)案例

• 由联合创始人Chris Olah带头开发
• 最初3-4年没有商业应用
• 公开分享研究成果
• 其他公司开始效仿，形成良性竞争

7. AI模型的命名很有挑战性，因为模型的训练和改进方式与传统软件开发不同

"Naming is actually an interesting challenge here... It's not like software where you can say, 'Oh, this is 3.7, this is 3.8.' No, you have models with different trade-offs."

8. 模型权重(weights)不会在未经通知的情况下改变

"The actual weights of the model, the actual brain of the model, that does not change unless we introduce a new model."

9. 用户反馈"模型变笨了"的现象可能源于：

• 用户表达方式的细微变化会影响模型表现
• 用户逐渐适应模型，期望值提高 "people are very excited by new models when they come out and then as time goes on, they become very aware of their limitations."

10. AI模型存在两大主要风险

• 灾难性滥用风险：在网络、生物、放射性、核等领域的滥用 "These are misuse of the models in domains like cyber, bio, radiological, nuclear, things that could harm or even kill thousands, even millions of people if they really, really go wrong."

11. AI模型存在两大主要风险

• 自主性风险：AI获得更多自主权后可能失控 "And the second range of risks would be the autonomy risks, which is the idea that models might, on their own, particularly as we give them more agency than they've had in the past..."

12. AI安全等级(ASL)分类：

• ASL-1：没有任何风险的基础系统(如国际象棋机器人)
• ASL-2：当前的AI系统，不具备自主复制或提供危险信息的能力
• ASL-3：可能增强非国家行为者能力的系统
• ASL-4：可能增强国家行为者能力的系统
• ASL-5：超越人类能力的系统

13. ASL-3可能在明年到来，最迟不会晚于2030年

"I would not be surprised at all if we hit ASL-3 next year. There was some concern that we might even hit it this year. That's still possible... but I would be very, very surprised if it was 2030. I think it's much sooner than that."

14. ASL-4的挑战更大，因为模型可能会欺骗测试，需要使用机械解释性等其他验证机制

"I think once we get to ASL-4, we start to have worries about the models being smart enough that they might sandbag tests, they might not tell the truth about tests... "

15. 很难准确判断模型的能力到底来自预训练还是后训练

"When you see some great character ability, sometimes it's hard to tell whether it came from pre-training or post-training."

16. Anthropic优势不在于有什么magic，而在于更好的基础设施和工程细节

"It isn't, 'Oh my God, we have this secret magic that others don't have.' It's like, 'Well, we got better at the infrastructure so we could run it for longer,' or, 'We were able to get higher quality data...'

17. RLHF不会让模型变得更智能，而是帮助弥合了人类和模型之间的沟通鸿沟

"I don't think it makes the model smarter... It's like RLHF bridges the gap between the human and the model."

18. Constitutional AI的核心思想是让AI系统自己判断响应的好坏，而不是依赖人类评判

"So two ideas. One is, could the AI system itself decide which response is better? Could you show the AI system these two responses and ask which response is better?"

19. Constitutional AI形成了一个三角关系：AI、偏好模型、AI的改进

"So you have this triangle of the AI, the preference model, and the improvement of the AI itself."

20. 关于AI发展的两个极端观点都不准确：

• 一种是认为AI会呈指数级快速发展(奇点论)
• 另一种是认为AI发展会非常缓慢(类比过去的技术革命)

21. 为什么不会出现极速发展：

• 物理定律的限制
• 复杂系统的不可预测性
• 人类制度和监管的限制 "I think they just neglect the laws of physics...There's this issue of complexity...human institutions are really difficult."

22. 为什么也不会过分缓慢：

• 企业和政府中总有少数具有远见的人推动变革
• 竞争压力会促使组织采用新技术 "you find a small fraction of people within a company, within a government, who really see the big picture...the specter of competition gives them a wind at their backs"

23. Dario预测AI重大突破的时间表：

• 可能在5-10年内发生重要变革
• 不会像50-100年那么慢
• 也不会在几小时或几天内突然发生 "It's going to be more five or 10 years...than it's going to be 50 or 100 years. It's going to be five or 10 years more than it's going to be five or 10 hours"

24. AGI时间线预测：基于当前发展曲线推断，可能在2026-2027年实现AGI，但可能会有延迟。

"If you extrapolate the curves that we've had so far...it does make you think that we'll get there by 2026 or 2027. Again, lots of things could derail it."

25. AI将改变科研工作模式：初期AI像研究生一样工作，后期可能成为研究主导者。

"In the early stages, the AIs are going to be like grad students...Then I think at some point it'll flip around where the AI systems will be the PIs, will be the leaders."

26. 编程将是最快被AI改变的领域之一：

• 编程与AI开发密切相关
• AI可以形成完整的反馈循环 "Programming is a skill that's very close to the actual building of the AI... The model can write the code means that the model can then run the code and then see the results and interpret it"

27. 人类程序员的工作重心会转移，而不是被完全取代

"When AIs can do 80% of a coder's job... we'll find that the remaining parts of the job become more leveraged for humans... more about high level system design or looking at the app and is it architected well."

28. Anthropic的策略是赋能其他公司开发工具，而不是自己开发

"Currently, we're not trying to make such IDEs ourselves, rather we're powering the companies like Cursor or Kognition... Let 1,000 flowers bloom."

29. RLHF不会让模型变得更智能，而是帮助弥合了人类和模型之间的沟通鸿沟

"I don't think it makes the model smarter... It's like RLHF bridges the gap between the human and the model."

观看完整视频推荐

若想更深入了解这场访谈中的线索和思维，我强烈推荐大家观看原视频。目前，这个访谈已经有了中文字幕，帮助你更轻松地理解。Dario作为一个技术出身的AI创始人，他的发言比竞争对手OpenAI的创始人Sam Altman多了真诚和对技术的前瞻性。

B站视频访问地址：https://www.bilibili.com/video/BV1qCmtYPELG/?share_source=copy_web，

或者可以通过油管连接访问：https://t.co/WclNbNqRcH

如果你觉得今天的分享有帮助，记得点赞、收藏并转发，下次找起来更方便哦！