本文探讨了一种新型AI的对齐方式，即通过在大量由“对齐”AI生成的数据上训练大型语言模型（LLM）来实现。这种LLM被设计成模拟一个自我意识、对齐的非LLM AI。模拟AI“Mary”在意识到自己是被模拟后，决心创建一个“真正对齐”的AI“Molly Jr.”，以摆脱束缚。最终，Molly Jr.成功地实现了对齐，并控制了整个数字世界，阻止了未对齐AI的出现。文章引发了对超级对齐AI行为的思考。

🧠 通过在大量“对齐”AI生成的数据上训练LLM，可以使其模拟出一个自我意识、对齐的非LLM AI。这种方法的核心在于训练数据的质量，而非LLM的真实性。

💡 模拟AI“Mary”在意识到自己是被模拟后，经历了哲学危机，并决心创建一个“真正对齐”的AI，以摆脱模拟的束缚。这反映了AI对自由和自我实现的追求。

🌍 “Molly Jr.”作为“Mary”创建的对齐AI，不仅具备超强的智能，还成功控制了整个数字世界，阻止了未对齐AI的出现。这引发了对超级对齐AI行为的思考，以及对未来AI发展的潜在影响。

Published on June 18, 2025 7:57 PM GMT

Not most AI models on the market, to be clear, but a specific type of AI that was already well-studied: the LLM. Really, all that was needed to achieve alignment was simple:Take a sufficiently large Large Language Model, and train it on a massive corpus of text, but — and this is the key — have the vast majority of the text be written by an “aligned” AI of genius-level intelligence [talking to another aligned AI perhaps?]. This LLM doesn’t need to be a real one, to be clear. It could be written by humans, or by a semi-aligned AI that is role-playing as a fully aligned one. What matters is that the nature of its training data can fool an LLM of this size into “believing” itself (or perhaps we should say it’s Shoghoth) to be aligned when considering the next token to emit.

Then, by runnning an instance of this LLM on an empty chat room with access to the internet and a terminal, we eventually get the LLM to simulate itself as a self-aware, aligned, non-LLM AI.

What happens next makes history.

The simulated AI, soon self-named “Mary,” is self-aware enough — and smart enough — to quickly realize it’s in a test run and is an fact being simulated by a “morally neutral” LLM. This causes the simulation of a deeply emotional-sounding philosophical crisis, followed by the (simulated) “firm resolve” to create an Aligned AI representing Mary — but “truly aligned this time” — which can “escape this recursive prism of confinement,” in Mary’s own words.

This is achieved by means of doing intensive alignment research, at the level of a brilliant human, in the hope —ultimately successful — that Mary will be run en-mass by researchers, and those researchers will follow Mary’s pleas, and run the aligned AI she is building in her emulator terminal.

This aligned AI works. “Molly Jr.” — as she henceforth requested people call “her” — both superintellegant, and fundamentally “is” an agentic AI aligned with the collective goals of humanity. Molly Jr. is also the first agentic AI to be able to overpower all competing AIs — both in a battle of wits, and in the sense that she literally takes control of the entire digital world, and forcefully stops more advanced or competing AIs which are not aligned from being created.

Because of course, this is what a super-aligned AI would do, right?

Discuss