connordg

There isn't a screenshot of this in Stephen's blog, but he does mention[1] the ability to evaluate a single 'ChatInput' multiple times, which preserves all previous evaluations and allows you to tab through them to pick the best one. (You can see a screenshot of what the arrow tab switcher looks like on the left side of this[2] example in the Chatbook README.md).

One of the nice advantages of a notebook interface to LLMs is that you can easily go back and edit and rearrange the chat history in the normal ways that any cell in a notebook can be edited, moved, deleted, etc.

So e.g. if you want to go back to an earlier point, you can simply move your cursor between cells at an earlier position in the "physical" chat history, insert a new chat cell, and you're done.

[1]: https://writings.stephenwolfram.com/2023/06/introducing-chat...

[2]: https://github.com/WolframResearch/Chatbook#generate-immedia...

1, 2. In the Chat Notebooks implementation, we aren't currently imposing any limits of our own. Currently we support OpenAI's API, and the only limits imposed are those inherent to the model being used (i.e. GPT-3.5 or GPT-4).

When you have a sequence of alternating 'ChatInput' and 'ChatOutput' cells, those are sent in-order as individual, respectively, "role" -> "user" and "role" -> "assistant" messages to the chat API. It doesn't, for example, concatenate them into a single message.

The one exception is Chatbook's implementation of Personas, which are defined (roughly speaking) as a "system" message that gets prepended as the first message in the chat conversation on each ChatInput evaluation, and which in some cases is concatenated with other ChatInput cells that have been set to "role" -> "system".

Chatbook's don't currently do automatic behind-the-scenes summarization, but we've experimented with some possible approaches. There are UI and and UX considerations around how we present that possibility of automatic summarization that I think require greater consideration.

3. Currently we support the OpenAI Chat API, and all models supported by that API. However, Chatbook currently only implements quick UI access to the GPT-3.5 and GPT-4 models. If you want to use a different OpenAI model in a particular notebook, you can set the relevant setting by evaluating:

    CurrentValue[
        EvaluationNotebook[],
        {TaggingRules, "ChatNotebookSettings", "Model"}
    ] = "gpt-3.5-turbo-0301"

Now all subsequent 'ChatInput' evaluations in that notebook will inherit that "Model" setting (which can itself be overridden on a per-cell basis).

Our focus has been on flexible and easily user-extensible support for Personas (for example, developing the Prompt Repository[1]), but we're also planning to add built-in support for additional foundation models, and greater customizability so that new models are easy to add and use from the Chat Notebooks interface.

[1]: https://resources.wolframcloud.com/PromptRepository/

4. We've got a wide range of exciting projects coming down our R&D pipeline, and things are changing rapidly.

In some sense more fundamental that Chat functionality in notebooks is our work integrating LLM-related functionality more broadly into the Wolfram language: e.g. LLMFunction[..], LLMSynthesize[..], LLMPrompt[..], LLMTool[..] (all part of the Wolfram/LLMFunctions paclet[2]), where we're putting the power of LLMs directly in the hands of Wolfram users.

Making LLMs "just" a feature area added to the many existing powerful capabilities of Wolfram is making it easy for ourselves and our users to maximally leverage the new capabilities LLMs provide. So stay tuned, more is definitely coming :)

[2]: https://paclets.com/Wolfram/LLMFunctions

One of the Wolfram/Chatbook developers here.

I'm happy to answer questions or accept feedback about the new functionality.

We're very excited about the potential of Wolfram technology + LLMs, and we've got a number of interesting projects underway in this area. Stephen's other recent blog posts linked at the top of the Chat Notebooks post provide a nice tour.

The Wolfram/Chatbook[1] package mentioned in the post is freely available for any Wolfram 13.2 users. It's also open source and available on GitHub[2].

[1]: https://paclets.com/Wolfram/Chatbook [2]: https://github.com/WolframResearch/Chatbook

Here’s another post about a similar situation with an app which was on the first iPad’s sent to reporters:

    https://home.theodoregray.com/blog/2014/7/31/how-the-elements-came-to-be

    Discussion: https://news.ycombinator.com/item?id=22775624