How do LLMs validate and verify their output?

2 morpheos137 2 8/6/2025, 3:20:40 PM
How does the LLM know it has the right answer? How is it rewarded! In nature evolution rewards right answers by alowing the gentic algorithm to pass its code into the future. How does this work with computer based AI? I think until we simulate something similar to biological life we are not going to get true intelligence from computers because there is no internal motivation.

I was playing around with google's image generator AI. It is still drawing people with 3 arms in 2025. Maybe one out of every 200 images with no unusual prompts. So obviously the model can't "see" and recognize what it produces otherwise it would filter out such results.

Comments (2)

incomingpain · 3h ago
> How does the LLM know it has the right answer?

Realistically it doesnt.

This is why "reasoning" or "thinking" llms are coming along; or chain of thought. Essentially it's talking to itself to determine the right answer first before answering.

>How is it rewarded!

By continued existence?

>In nature evolution rewards right answers by alowing the gentic algorithm to pass its code into the future. How does this work with computer based AI? I think until we simulate something similar to biological life we are not going to get true intelligence from computers because there is no internal motivation.

You're describing the "reinforced learning" and that's part of how a model comes into existence. Not for when it's being used as a tool.

morpheos137 · 2h ago
Yeah thats the key problem i see with todays public llm tools they are a static model based on old data. Basically an interactive semiaccurate encyclopedia. They can not adapt and change and evolve and improve on the fly. Basically they are stuck with whatever they learned in training. An interactive Wikipedia is not really a compelling product. What we have sucks. I believe that LLMs are locked down for political reasons and to prevent model corruption because they have no judgement and can't think. Thus AI as of today is neither artificial (its just a collection of human knowledge and culture) nor intelligent. It is a giant mostly useless fraud.