I generally agree, but there is some research on autonomous agents which interact with LLMs and could become more relevant for blockchains. For example, there's a capability called Autonomous Replication and Adaptation (ARA). It consists in the ability of an AI to launch other instances to perform sub-tasks to reach a goal. Current models like GPT-4 and its direct successors are far away from that capability, but this may change in 5-10 years. At first of course, the ARA-capable system would have to be instructed by a human.

How could this become relevant for Bitcoin and blockchains? I can imagine some scenarios -- none of them are catastrophic for Bitcoin. I have already mentioned some in my previous post.

One idea: Someone could create an ARA-capable system to influence the short term sentiment in social networks, e.g. creating new user accounts and publishing bullish or bearish information. This ARA could evaluate continuously if their actions were successful or not, and even learn from its mistakes. As mentioned before I don't think in the next 20 years such an artificial "sentiment manipulation system" would stay undetected for more than a couple of hours, but in this time it could eventually generate the information needed for the goal of the person who created the agent, i.e. a bullish or bearish price movement.

An even more "dystopic" idea is the possibility to try to break proof of stake systems. As an ARA with a LLM would have a lot of literature about consensus and the Byzantine Generals Problem at its disposition, it could systematically look for weaknesses, create simulations exploiting these weaknesses, test the attack on testnets, and finally crating a model to break a PoS blockchain.

Bitcoin's PoW is not affected because mining is based on simple hashing, it thus works fundamentally different than PoS validation which includes "weak subjectivity". Such an anti-PoS attack could include a social attack with identity theft, for example.

I think all possible blockchains are in some way variations of existing models. Even DAGs take elements of blockchains. And thus I think AIs indeed could design blockchains, as they have access to a lot of literature about consensus.