𝗙𝗿𝗼𝗺 𝗔𝗚𝗜 𝘁𝗼 𝗔𝗦𝗜: 𝗔 𝗦𝘂𝗺𝗺𝗮𝗿𝘆 𝗼𝗳 𝗚𝗼𝗼𝗴𝗹𝗲 𝗗𝗲𝗲𝗽𝗠𝗶𝗻𝗱’𝘀 𝗟𝗮𝘁𝗲𝘀𝘁 𝗣𝗮𝗽𝗲𝗿

Most people talk about Artificial General Intelligence (AGI) with hype or fear. Artificial Superintelligence (ASI) is even harder to discuss.

A new Google DeepMind paper asks a better question. Instead of asking when AGI will arrive, it asks: What happens after AGI?

If we build AI that matches human capability, how does it keep improving until it outperforms entire groups of human experts?

Here is a simple breakdown for engineers and developers.

𝗧𝗵𝗲 𝗗𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲

• AGI: An AI system that performs at the level of a median human across many tasks. It can learn, reason, and solve diverse problems. • ASI: An AI system that outperforms large groups of human experts across almost all important fields.

A model that beats a human at chess is not ASI. ASI is an artificial system that can outperform research labs and entire companies.

𝗪𝗵𝘆 𝗔𝗜 𝘀𝗰𝗮𝗹𝗲𝘀 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁𝗹𝘆 𝗳𝗿𝗼𝗺 𝗵𝘂𝗺𝗮𝗻𝘀

Humans have biological limits. We cannot copy ourselves or run a thousand versions of our minds in parallel. AI does not have these constraints.

AI offers several advantages:

  • Machine speed: AI processes data much faster than biological brains.
  • Reasoning scale: You can give a model more compute to think longer or run many instances in parallel.
  • Memory: AI can connect to massive databases and infinite context windows.
  • Duplication: You can run millions of copies of a capable AI researcher.

𝗧𝗵𝗲 𝗙𝗼𝘂𝗿 𝗣𝗮𝘁𝗵𝘄𝗮𝘆𝘀 𝘁𝗼 𝗔𝗦𝗜

  1. Scaling: Using more data, more compute, and better hardware.
  2. Algorithmic shifts: Finding new architectures, better memory systems, or more efficient learning methods.
  3. Recursive self-improvement: AI helping to research and build better AI. This creates a feedback loop.
  4. Multi-agent systems: Many specialized AI agents working together like a large organization.

𝗪𝗵𝗮𝘁 𝗰𝗼𝘂𝗹𝗱 𝘀𝗹𝗼𝘄 𝘂𝘀 𝗱𝗼𝘄𝗻?

Progress is not guaranteed. Several bottlenecks exist:

  • Data limits: We may run out of high-quality human data.
  • Physical limits: Scaling requires massive amounts of electricity, chips, and hardware.
  • Research difficulty: As AI matures, the problems to solve may get harder.
  • Human intervention: Safety, regulation, and politics can slow development.

𝗧𝗵𝗲 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆

从 AGI 向 ASI 的演进不仅仅是一个研究问题，更是一个工程挑战。它将涉及分布式系统、基础设施以及复杂的编排。

如果你是一名开发者，请关注以下领域：

  • 如何构建可靠的 AI 智能体。
  • 如何协调多个智能体。
  • 如何评估长期任务。
  • 如何确保安全与控制。

来源：https://dev.to/marrouchi/from-agi-to-asi-a-summary-of-google-deepminds-latest-paper-2ndd