General intelligence does not necessitate complete architectural flexibility
For an AutoGPT to have general intelligence (be able to in principle learn any idea/solve any problem) it does not necessarily need a completely flexible architecture or the ability to modify its own code.
This is analogous to how humans do not need to modify their genetic code in order to increase their “intelligence”, but rather to do so, only need to learn new ideas, which involves modifications of neuronal connections.
The means by which an AutoGPT can increase its intelligence without needing to modify its code/architecture consist of learning and discovering/creating new ideas, which it can then employ to solve increasingly complex problems.
In order for this to occur though, the initial architecture must be designed to be capable of representing and reasoning about any possible idea. This is mostly already achieved by the LLM component of the AutoGPT, however other components are also necessary such as a long term memory mechanism able to store and relate such ideas to one another, akin to how ideas are represented in a zettelkasten.
The ability of the AutoGPT to store new ideas in its memory is critical to it being able to improve itself, given that what really constitutes and distinguishes one instance of an AutoGPT from another are the ideas in their memory.
If you found this interesting, have feedback or are working on something related, let’s chat: twitter (@0xdist) or schedule a 20 min call
Distbit
Interested in econ, cryptoecon, agents, finance, epistemology, liberty.