Auto Personas - Meta-analysis of chats and baby steps to AGI: Hyperlink in chat to launch new agents for specialized tasks.

[jondwillis]

Use-case:

In the course of consulting ChatGPT in many specialized fields, it will urge you to consult with a professional, e.g. a lawyer, a doctor, an electrician, peers, etc. If we are able to analyze portions of the bot's output for instructions to consult with a specialist, we can offer to launch a new chat with a new system prompt generated from that consultation instruction.

Example:

User: I want to start a business.
ChatGPT: Choose a business structure: .... Consult with a legal expert or business advisor to determine the best option for your startup.

Where clicking Consult with a legal expert, etc., would open a new tab with an editable custom system prompt.

More discussion needed:

ChatGPT's official interface, with plugins enabled, detects plugin-able requests, and this would be architected similarly. Except this would be open for custom plugins eventually!

Criticisms?

[enricoros]

I love the direction you're going in. Let's think of how to make this detectable, easy to experience, and how to factor with behavior in any possible system prompting.
Is it the case that if this feature kicks in, the user selected the wrong "purpose" to start with?

[enricoros]

Technology: How about a post-generation Enriching framework? An optional postprocessing step that takes the output of the assistant and can rewrite it in accordance of tools/rules (such as the one you mention, "sending you to a different expert")

Product: iteration 1 shall have links being injected (possibly with an indication of the tooling a link refers to), iteration 2 shall have augmentation of the UI (menu, widgets, text editing) to cast the information in a more usable way.

Naturally this would evolve to pre-processors as well. (E.g. expansion of URLs into their content, within the context window) - opinions?

[enricoros]

@jondwillis - we're at a point in technology where this is feasible.

Have you had more thinking about this feature from a UX perspective and use case perspective? How frequently do you bump into it?

[shiribailem]

This could connect to the suggestion I made in #35 of Personas into Identities, where the suggested identities would have associated tags that could be referenced.

This could also potentially bypass those prompts with an internal system prompt that provides the names of those identities and references, in the vein of "I know the following people who can help me where appropriate: Allison the doctor, Stephen the business analyst, Sam the lawyer...", which might cause the models to naturally and fluidly replace "Consult with a business advisor" with "You should talk to Stephen about this".