这段对话围绕着安德烈·卡帕蒂（André Carpati）展开，他讨论了人工智能代理的未来、他在人工智能教育领域的工作，以及他对智能和技术进步的更广泛看法。卡帕蒂认为，关于人工智能代理快速发展的预测被夸大了，他认为“代理的十年”比“代理的一年”更现实。他强调了目前人工智能代理在智能、多模态和持续学习方面的局限性，并指出它们还不适合像员工级别的工作。 卡帕蒂凭借他在人工智能领域15年的经验，将当前的乐观和悲观情绪放在了背景中，回顾了早期围绕深度学习和游戏中的强化学习的兴奋之情。他批评过度依赖游戏作为通用人工智能的训练场，主张将现实世界的互动和知识工作作为更相关的目标。 他用一个建造“幽灵”而不是“动物”的比喻，强调当前人工智能的发展根植于模仿人类生成的互联网数据，而不是复制进化或动物的学习过程。卡帕蒂将此与理查德·萨顿（Richard Sutton）构建人工智能来模仿动物的方法进行了对比，并对这是否完全可行表示怀疑。 对话探讨了上下文学习的细微之处，认为它可能是一种令牌窗口内的梯度下降形式，完成模型权重中的模式。讨论涉及预训练和上下文学习之间的差异，强调了模型可以存储多少信息。他还提出了一个类比，将模型比作人类，并认为模型回忆信息就像“模糊的回忆”。 卡帕蒂谈到了如何通过剥离模型的知识并保持“认知核心”来打造模型的最佳版本。他概述了人们如何通过使用推理和前额叶皮层来推动该领域的发展，而对基底神经节这样的功能并没有太多关注。 卡帕蒂就人工智能在软件工程中的应用提供了见解，指出自动补全现在很有帮助，但尚未为高级编码和从未编写过的代码做好准备。他强调，代理往往会犯太多的认知错误，并且他认为它不是实际的净收益。 讨论涉及强化学习的挑战以及在过程监督中分配部分功劳的难度。他提出了使用大型语言模型（LLM）作为裁判的想法。 展望人工智能，他表达了对人工智能变得更强大，但也更容易失控的担忧。他描述了这个领域如何变成一个完全自主的锅。 在讨论Eureka时，卡帕蒂设想建立像“星际舰队学院”一样的东西，这是一所专注于技术知识的精英教育机构。他强调最好的选择是一对一的辅导，并希望围绕它建立一个教育模型。此外，他希望建立“知识的阶梯”，以帮助推动该领域的发展并教授学生。 最后，卡帕蒂讨论了他之前在特斯拉构建自动驾驶汽车的角色。他表示，这还需要一段时间，并认为演示阶段通常是无用的。他也在为自动驾驶汽车构建一个教育工具。 他还为任何领域的教育者提供了一些见解，指出他们应该首先列出第一性原理，然后再解决其他方面的问题。

The conversation features André Carpati, discussing the future of AI agents, his work in AI education, and his broader views on intelligence and technological progress. Carpati argues that predictions of rapid advancement in AI agents are overblown, suggesting a "decade of agents" is more realistic than a single year. He highlights current AI agents' limitations in intelligence, multimodality, and continuous learning, stating they are not yet suitable for tasks like employee-level work. Carpati draws on his 15 years of experience in AI to contextualize present optimism and pessimism, referencing the early excitement around deep learning and reinforcement learning in gaming. He critiques the over-reliance on gaming as a training ground for general AI, advocating for real-world interaction and knowledge work as more relevant targets. He uses an analogy of building "ghosts" rather than "animals," emphasizing that current AI development is rooted in imitation of human-generated internet data rather than replicating evolution or animalistic learning processes. Carpati contrasts this with Richard Sutton's approach of building AI to mimic animals, expressing skepticism about whether this is fully achievable. The conversation explores the nuances of in-context learning, suggesting it might be a form of gradient descent within a token window, completing patterns within the model's weights. The discussion touches on the differences between pre-training and in-context learning, highlighting how much more information the models can store. He also brings up an analogy of models resembling humans and the models recalling information like a "hazy recollection" Carpati talks about how to make the best version of models by stripping them from knowledge and maintaining a "cognitive core" He outlines how people are pushing this field forward by using reasoning and the prefrontal cortex, and not doing much for like the basal ganglia. Carpati offers insights on software engineering with AI, noting that autocomplete is now helpful but isn't ready for advanced coding and code that has never been written before. He emphasizes that agents tend to make too many cognitive errors, and he doesn't think it's an actual net gain. The discussion touches on the challenges of reinforcement learning and the difficulty of assigning partial credit in process supervision. He brings up the idea of using LLMs as judges. Looking at AI, he describes his concerns with AI becoming more powerful, but also more out of control. He describes how the field can become a pot that is completely autonomous. Discussing Eureka, Carpati envisions building something like "Starfleet Academy," an elite educational institution focused on technical knowledge. He stresses that the best option is a one-on-one tutor and wants to build an education model around that. As well, he hopes to build "ramps to knowledge" to help advance the field and teach students. Lastly, Carpati discusses his previous role at Tesla in building a self driving car. He states that it will still take a while, and thinks the demo stage is often useless. He is building an educational tool for self-driving cars as well. He also offers insights for any educator from any field, stating that they should start by laying out the first order and then attack the other terms.