The Oracle Problem and the AI Grounding Problem

There is a problem in blockchain called Oracle problem, and I think AI has exactly the same problem. Let me explain. Both blockchain and large language models share something important. They are both closed systems. A closed system means that it cannot see the real world by itself. A blockchain does not know what time it is now. It doesn't know the price of Bitcoin right now. It doesn't know what happened in the news today. Everything inside the blockchain is self-contained. And LLM is something similar. After training, it doesn't know what day it is today.

It doesn't know what just happened. The knowledge is frozen. Both systems are very powerful, but both systems do not have direct access to reality. So the question is, how do you bring the reality into a closed system? In blockchain, they use something called Oracles. Oracles are systems that bring off-chain data on chain. They bring prices, timestamps, events, results. But the moment you add an Oracle to the blockchain, you add a new risk. What if the data is wrong? What if the data is manipulated? This is the Oracle problem.

The blockchain community has been dealing with it for years. And now let's look at large language models. LLMs can generate text very well, but they struggle with factual accuracy. So we add bridges, search, documents, tools, APIs to ground the model to real data. This is something called grounding or RAG, retrieval augmented generation. And here is my thought on that. This is the same problem. In blockchain, to trust the Oracle, in AI, to trust the source, or the tool output. Different technology, but very similar problems. The weakest point in both systems is the bridge to reality.

So the real question is not how good the model is, but how good is the bridge and can you actually trust it? This was the thought on AI for today. I hope you liked it and I'll see you in the next one. Bye.