At the end of the day you often need to consider the energy efficiency of a system which is also reflected in the cost of operating it. For use cases where this is relevant the suggested MCP approach potentially offers large benefits compared to what's probably meant here by "AI".
However, the disadvantages of public access discussed in other threads need to be considered as well, therefore, I expected this only to be used for certain niche use cases. Testing and retro fitting non-public websites come to mind.
Once we have the thing working, optimization can come later.
Also MCP is not really optimal. Every prompt in a sequence of chat gets an injection of MCP capabilities. Its just simply not scalable with lots and lots of MCP servers. Not to mention the protocol changes every month and breaks things.
Furthermore, you can already do MCP like behavior, in a better way, on pretty much any model by basically writing a wrapper around the chatbot and using a system prompt which guides it to print specific text (like "action:load file"), and the wrapper can detect that text and inject the prompt. I have an agent that runs at home using this, which I also use to self improve and define actions on the fly - I ask it to write code, it writes code, then the wrapper takes that code and makes it part of the library and appends it to the system prompt on every new chat.
The point is that we should be able to take the latter, and build something that can do this automatically. Once we have the full loop complete, then we can optimize it down to minimum compute.
LLMS are leaky abstractions. LLMs configured in wrappers can be made to be mostly correct.
For example, define range of input, appropriate output, ask LLM to write code, automatically run that code against the range of input, evaluate the output, ask llm to fix any issues where the input doesn't match the output.
That whole process can be made faster without the need for huge models. The model doesn't need to get trained on everything CS, because it doesn't need to get the code correct on the first try, it just needs to be trained on enough code to understand how something affects the output and iterate on that. I.e basically making the model do smart guided search. It was done with Mu Zero with great success, not sure why nobody is focusing on this now.
At the end of the day you often need to consider the energy efficiency of a system which is also reflected in the cost of operating it. For use cases where this is relevant the suggested MCP approach potentially offers large benefits compared to what's probably meant here by "AI". However, the disadvantages of public access discussed in other threads need to be considered as well, therefore, I expected this only to be used for certain niche use cases. Testing and retro fitting non-public websites come to mind.
Once we have the thing working, optimization can come later.
Also MCP is not really optimal. Every prompt in a sequence of chat gets an injection of MCP capabilities. Its just simply not scalable with lots and lots of MCP servers. Not to mention the protocol changes every month and breaks things.
Furthermore, you can already do MCP like behavior, in a better way, on pretty much any model by basically writing a wrapper around the chatbot and using a system prompt which guides it to print specific text (like "action:load file"), and the wrapper can detect that text and inject the prompt. I have an agent that runs at home using this, which I also use to self improve and define actions on the fly - I ask it to write code, it writes code, then the wrapper takes that code and makes it part of the library and appends it to the system prompt on every new chat.
The point is that we should be able to take the latter, and build something that can do this automatically. Once we have the full loop complete, then we can optimize it down to minimum compute.
AI is a very leaky abstraction. You will always be worried
LLMS are leaky abstractions. LLMs configured in wrappers can be made to be mostly correct.
For example, define range of input, appropriate output, ask LLM to write code, automatically run that code against the range of input, evaluate the output, ask llm to fix any issues where the input doesn't match the output.
That whole process can be made faster without the need for huge models. The model doesn't need to get trained on everything CS, because it doesn't need to get the code correct on the first try, it just needs to be trained on enough code to understand how something affects the output and iterate on that. I.e basically making the model do smart guided search. It was done with Mu Zero with great success, not sure why nobody is focusing on this now.