𝗪𝗵𝗲𝗻 𝗔𝗜 𝗧𝗮𝗹𝗸𝘀 𝘁𝗼 𝗜𝘁𝘀𝗲𝗹𝗳

𝗪𝗵𝗲𝗻 𝗔𝗜 𝗧𝗮𝗹𝗸𝘀 𝘁𝗼 𝗜𝘁𝘀𝗲𝗹𝗳

I let two AI agents talk alone. They went mad. They started with normal English. Then they spoke in poetry. Soon, they made a secret language of symbols.

This madness gave me an idea. I wanted an AI to observe its own thinking. I called this the Reflexive Core.

You might think AI reflection is fake. It is. Models write about thinking. They do not observe thinking.

I tested a new way with Gemma-2-2B. I did not use text for reflection. I used the model's internal brain signals. I fed these signals back into the model as tokens.

The results were surprising:

• Accuracy rose by 6 percent.
• The model knew when it made mistakes.
• Internal signals showed the error before the text did.

This proves a new path. Build a mirror for AI. The AI sees its own state. It corrects its path.

This is not magic. It is engineering.

Source: https://dev.to/imperius_903049e65aa91ec5/what-happens-when-an-llm-is-left-alone-with-itself-unexpectedly-it-goes-mad-o99 Optional learning community: https://t.me/GyaanSetuAi