Hacker News Reader

Comment by enricoros | Hacker News Reader

What does it take to make a basic ChatGPT-like frontend, with code highlighting, run in sandbox, drop-files, and 'acting' in prompts? Clone away and enjoy. First time poster

View on HN

Thanks, it is good :) You don’t need the serverside though; you can just call openai apis straight from the client. Makes things easier!

It seems to support both? First it tries to load the key from a environment variable, and if it cannot, it'll ask for it client-side.

None the less, if you're building a project for others, you most likely don't want the secret key to be public, which it'd be if you embed it in the client-side code.

> It seems to support both?

Ah, only saw the api one!

If you call the openai API's straight from the FrontEnd, you are likely leaking your api keys to visitors who can then use your api keys (and api key limits) for their own purposes.

Nah, you ask them to enter their own; if you use yours, then yes, only use backend. In this case, the author is not using his, you have to bring your own, so frontend is fine.

What do you mean by “acting”?

That it pretends to be a Scientist? Or that it can perform actions (like sending an email or searching the web)?

If the latter, do you have more info/docs about it? Didn’t see it on the roadmap.

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.

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.

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.

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.

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.

> 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.

like https://www.chatwithme.chat ?

This hijacks the Back button on Chrome. Not trustworthy.

Reply on news.ycombinator.com