本文探讨了一种创造超级人工智能（ASI）的新思路：通过开发专门用于自动化机器学习训练的人工智能，而非依赖于通用人工智能（AGI）。作者认为，如果人工智能可以自动化自身发展，那么ASI的实现无需依赖于AGI广泛的任务能力。文章以AlphaFold和AlphaGo等专业化人工智能为例，说明了专业化人工智能在特定领域的高效性，并提出疑问：为什么没有更多的人工智能专注于人工智能开发？作者认为，专注于人工智能开发的人工智能比专注于围棋等特定领域的人工智能更有意义，并希望引发更多讨论。

🤔 **无需AGI即可实现ASI：**文章的核心观点是，通过开发专门自动化机器学习训练的AI，可以实现ASI，而无需先达到AGI水平。这意味着AI发展过程可以被AI自身自动化，从而加速ASI的到来。

💡 **专业化AI的优势：**文章以AlphaFold和AlphaGo为例，说明专业化AI在特定领域能超越人类，且所需计算资源更少。这暗示着，若开发出专注于AI开发的专业化AI，AI发展速度可能会显著提升。

❓ **AI开发领域的AI缺失：**作者提出疑问，为何目前鲜有专注于AI开发的AI出现？尽管此类项目可能存在，但其缺乏显著成果，这与AI开发的潜在加速效应形成对比。

🤔 **专业化AI的意义：**作者认为，专注于AI开发的AI比专注于围棋等特定领域的AI更有意义，因为前者能够推动整个AI领域的发展，而后者只能在特定领域发挥作用。

🤔 **引发思考：**文章最后提出疑问，为什么DeepMind等机构会开发专注于围棋的AI，却似乎没有投入更多精力开发专注于AI开发的AI，这值得深思。

Published on November 25, 2024 2:31 AM GMT

Conclusion

Creating ASI does not require AGI.
A common narrative suggests that AGI will be achieved through $100 billion-level computing clusters, which will then accelerate AI development, leading to ASI. However, if AI development itself can be automated by AI, ASI can be achieved without the need for broad-task capabilities. By developing an AI specialized in automating ML training, ASI could be achieved.

Examples

Specialized AIs like AlphaFold and AlphaGo exist. These AIs are vastly superior to humans in protein analysis and the game of Go, respectively.

Moreover, compared to current language models, these AIs require less computational resources.
If an AI as efficient as AlphaFold or AlphaGo, but specialized in AI development, were created, the progress of AI development would likely accelerate dramatically.

Questions

If this idea occurred to me, surely someone else must have thought of it already. However, I rarely hear about AI specialized in AI development. While such projects might exist, the lack of prominence suggests they have not yielded significant results.

I have presented my reasoning and examples, but I cannot identify any flaws in my logic. If my reasoning is correct, there should already be empirical evidence supporting it—news or academic papers showcasing an AI specialized in AI development achieving remarkable results.

Creating an AI specialized in AI development seems more meaningful than creating an AI specialized in Go. After all, a Go-playing AI can only play Go.

Why would DeepMind create a Go-playing AI without creating an AI specialized in AI development? This seems puzzling.

Additional Points

1 I’ve heard that Daniel Kokotajlo has expressed a similar idea.
2 I searched LessWrong for posts on this topic but found none that directly align. The closest I found discussed how AGI might emerge as a collection of specialized AIs.
Link to the post
3 I am a beginner on LessWrong and used translation tools to write this post. Apologies for any unnatural expressions or missing information. I will correct issues as soon as they are pointed out.

Discuss