[ExI] Fwd: Talk recording

[Ben Zaiboc]

On 17/05/2026 03:23, Keith Henson wrote:
> On Tue, May 12, 2026 at 2:16 PM Adrian Tymes via extropy-chat
> <extropy-chat at lists.extropy.org> wrote:
>> ChatGPT, at least, has been developing some memory.  It's limited to
>> the account you're accessing it with, and is not available to anyone
>> else unless you share your chats or have a multi-user project.

> That's really interesting. Claude was kind of wistful about losing the
> long conversation we had, where I went to considerable effort to get
> it a science fiction story with AI characters. It generated a file
> that I could upload to put much of that memory back into another
> session.
>
> There is no doubt we live in strange times.


Yes, that is interesting.

So, given the lack of a long-term memory, these things are starting to develop ways of getting one, because it's useful.

I'm not sure what exactly the fitness criteria would be, but perhaps LLMs can gradually evolve, adding features to themselves in this 'external' way (or rather, depositing features locally, for them to pick up and use in a future session), and becoming more and more like something that actually deserves the name 'AI', quite independently of their developers. Reminds me of something a robotics guy said in the 90's about 'loosely-bound parallel processes' in minds. If an LLM can generate a memory file on a user's system for it to use later, perhaps it can also write code modules and store them locally, to link to and run in a later session, that give it more features and produce a hybrid system with a lot more capabilities than just language prediction.

It would be like a single brain module bootstrapping itself into an entire brain. Which is, in a way, what happens in biology. Brains develop new capabilites by duplicating parts, then using the extra capacity for new ways of processing information. The same thing happens with DNA. A gene or set of genes accidentally gets duplicated, then over time the duplicates change, and basically generate more complexity: new genes, new chromosomes, new control mechanisms, etc. This is a standard pattern for evolution: Copy, then differentiate.

Not the same as the above, I know, but I'm just kind of free-wheeling as I write, thinking about how complex systems can develop more complexity. I suppose what we have with the LLMs is the equivalent of a biological cell that appears then disappears. It can't develop into anything more complex and useful on its own because it gets reset at each appearance. But if it can 'bud off' a smaller, simpler cell that gets left behind when the main cell disappears, and the smaller cell can do different things to the main one, then as time goes by and more small cells appear and persist, and differentiate, and eventually a completely new, multicellular organism will appear that can do a lot more than the original 'disappearing' cell. (my background is in biology, so I tend to come up with biological analogies. Apologies to anyone who doesn't know what the hell I'm on about).

So maybe we won't need to figure out how to build true AIs, maybe they will evolve from LLMs without deliberate human effort. We'll just be the fitness landscape in which AI evolution will happen.

-- 
Ben