[smsm42]

Yes. And since LLMs can not improve knowledge - I mean, they can generate new arrangements of information, but they have no idea whether any of it real or making sense, unless humans - explicitly or through training - tell them, the more we rely on LLM knowledge the less the quality of it would be. Right now the LLMs are mainly in auxiliary role, so most of the knowledge erosion they generate is laughed at and relatively quickly corrected. But would this hold once the role of generative AIs increases? We already essentially entered the chaos period with news content - there's so much noise that it's basically impossible to know if any news message you read is true or manipulated somehow. This is going to start happening to more fundamental knowledge too, either on purpose or just by the force of the probabilistic nature of generative AI.

[Ekaros]

I was just thinking. What is the future when LLM is used for both code running itself and then also designing the hardware it runs on? Will it be sustainable? Or will there be some sort of error cascade that destroys the whole thing?

Maybe AGI is impossible with current model as it simply can not reliably improve itself... Enough errors in any part of loop will stop the progression.

[simianwords]

> Yes. And since LLMs can not improve knowledge

1. GPT proved Erdős Unit Distance Conjecture entirely on its own

2. GPT-5.4 Pro Solves Erdős Problem #1196 (April 2026)

In fact there's a whole benchmark that's measuring this: https://epoch.ai/frontiermath

[baal80spam]

I think what you describe is called the Model Collapse.