undefined | Better HN

0 pointscapableweb3y ago0 comments

You can kind of program something around the API to make it able to perform actions. That's the way I'm guessing Microsoft did it with Bing. Here is an example:

System prompt:

    - You are a text-based AI delegates to other AIs.
    - You only respond in JSON with two keys `delegate_to` and `parameters`.
    - `delegate_to` defines what other AI you should delegate to.
    - `parameters` are parameters to send to that AI.
    - Here is the list of possible delegates
    - delegate_to: email_ai, parameters: $to $subject $body, used for sending emails
    - delegate_to: calendar_ai, parameters: $time, $date, $title, used for creating events in a calendar
    - delegate_to: search_ai, parameters: $search-term
    - Be creative and helpful.
    - Your goal is to take user input and successfully delegate the task you receive from the user to the correct AI.
    - Your initial message to the user should welcome them and ask them what you can do for them.

Example conversation:

    > {"message": "Welcome! I'm your AI assistant. What can I do for you today?"}
    
    > User: Send email to john@example.com asking if he wants to have dinner tonight
    
    {
      "delegate_to": "email_ai",
      "parameters": {
        "to": "john@example.com",
        "subject": "Dinner Invitation",
        "body": "Hi John, would you like to have dinner together tonight? Let me know your thoughts. Best regards."
      }
    }

Then you'd make your program read the JSON messages it generates and perform the correct action.

0 comments

nico3y ago

Nice. How would you go about providing a dynamic list of delgates? Could it work to just give one delegate that can provide a list of delegates with a description of the actions they can perform (then that delegate can query a db and return a list)?

Re-reading, I’m guessing the prompt could also be dynamically generated to include the most relevant delegates.

sprobertson3y ago

> the prompt could also be dynamically generated to include the most relevant delegates

Yup that's how I'm doing it - the system prompt is re-generated for every request, and that includes getting a list of available delegates and the arguments they accept. I only have 10 so I'm just listing all of them, but if you had some huge number you could combine that with embeddings / vector lookup.

behnamoh3y ago

At that point we're basically back to the `AI is just nested if-else expressions` story. The only difference is that now there is a language reader on top that understands the semantic of your language. But actors (or agents in LangChain lingo) are just if-else. The tools you connect them to must be developed separately.

nico3y ago

Sure, you could also say that human language/action capacity is just a biological LLM with some ifs on top that give it access to actions.

In the case you describe, you can have an LLM write the tools.

Yes, the first tools and bridge code might need to be manually built. But after that it could be LLMs all the way down.

Kind of similar to writing a new programming language. At first you write it in another language but after compiling it for the first time, you can then write the language in the language itself.

enricoros3y ago

Very good point. Once you start breaking down a llm into presets/delegators, you introduce basically if-else, with all the problems of that split. Lack of visibility, local vs global optimization, lack of control and predictability, asymmetry of information. I wonder if the current Agents approach is a stopgap solution.

j / k navigate · click thread line to collapse

0 pointscapableweb3y ago0 comments

You can kind of program something around the API to make it able to perform actions. That's the way I'm guessing Microsoft did it with Bing. Here is an example:

System prompt:

    - You are a text-based AI delegates to other AIs.
    - You only respond in JSON with two keys `delegate_to` and `parameters`.
    - `delegate_to` defines what other AI you should delegate to.
    - `parameters` are parameters to send to that AI.
    - Here is the list of possible delegates
    - delegate_to: email_ai, parameters: $to $subject $body, used for sending emails
    - delegate_to: calendar_ai, parameters: $time, $date, $title, used for creating events in a calendar
    - delegate_to: search_ai, parameters: $search-term
    - Be creative and helpful.
    - Your goal is to take user input and successfully delegate the task you receive from the user to the correct AI.
    - Your initial message to the user should welcome them and ask them what you can do for them.

Example conversation:

    > {"message": "Welcome! I'm your AI assistant. What can I do for you today?"}
    
    > User: Send email to john@example.com asking if he wants to have dinner tonight
    
    {
      "delegate_to": "email_ai",
      "parameters": {
        "to": "john@example.com",
        "subject": "Dinner Invitation",
        "body": "Hi John, would you like to have dinner together tonight? Let me know your thoughts. Best regards."
      }
    }

Then you'd make your program read the JSON messages it generates and perform the correct action.

0 comments

nico3y ago

Re-reading, I’m guessing the prompt could also be dynamically generated to include the most relevant delegates.

sprobertson3y ago

> the prompt could also be dynamically generated to include the most relevant delegates

behnamoh3y ago

nico3y ago

Sure, you could also say that human language/action capacity is just a biological LLM with some ifs on top that give it access to actions.

In the case you describe, you can have an LLM write the tools.

Yes, the first tools and bridge code might need to be manually built. But after that it could be LLMs all the way down.

Kind of similar to writing a new programming language. At first you write it in another language but after compiling it for the first time, you can then write the language in the language itself.

enricoros3y ago

j / k navigate · click thread line to collapse