swiftlyTyped

Obviously we'll let customers decide how to deploy this tech, but our vision is less about "chat with AI customer support", and more "tap into intelligence to help you achieve your task".

I personally fell in love with copilots when all my engineering tools started adding them (e.g. Cursor IDE, Warp terminal, etc.).

And really, I wanted to tap into artificial intelligence everywhere, from excel to planning a vacation in AirBnb. "is this a good place to stay with a young kid?"

Just in recent months, the market has validated this intuition, e.g. Microsoft Copilot products (think 2nd-brain-on-tap for e.g. PowerPoint, not support chat) have boosted productivity by 30%, with lots of other impressive metrics

---

If you do go back to the customer support use-case, I'd see the support agent's software using CopilotTextarea, to let them reply to 10x more messages, but where they still drive the conversation.

Though I could see it coupled with AI chatbot on the consumer side, where it makes sense

Great minds think alike!

Our `useMakeCopilotActionable` is similar to your `useBeakFunction` and `useMakeCopilotReadable` is analogous to your `useBeakInfo`.

I don't know how much farther you're planning to take things or if it's a demo/hobby project.

But real talk, if you're excited about this space and would like to explore joining our team, I'd love to chat and take it from there.

atai at copilotkit dot ai

CopilotKit itself is open-source, and supports assistants API + langchain :) https://github.com/CopilotKit/CopilotKit

It's a fair point :)

AI Copilot started out synonymous with code completion, it still means that for many.

Our bet was this meaning would broaden with time.

Yes, exactly.

And thanks for the kind words!

Good question!

No, we actually don't even offer our own endpoints for inference yet, precisely because of the data concerns you site.

Right now it's self-host only, but it's very very easy to self-host, and we'll keep it that way for every feature touching user data.

Correct.

The default backend implementation uses OpenAI, but you can use any LLM pretty easily.

Correct, Copilot here stands for AI "second brain" that "co-pilots" the app alongside the user.

(the abstract idea behind GitHub Copilot).

I didn't know ChatGPT + Edge did that, pretty cool!

And yes, with CopilotKit you can implement the same to any text area in any app, that would work on any browser (+ other key features).

I can tell you that CopilotTextarea works very, very well if you give it the right context - but since I hadn't known about Edge's feature, I can't compare.

Even long term, I think this is functionality that will be better offered by the app itself rather than the browser - bc the app can control the relevant context beyond what's visible on the screen (e.g. RAG on users' backend state), customize the prompt, and potentially even design dedicated LLM chains (as needed).

Yeah, I hope that doesn't happen.

That said the concept of a copilot is pretty generic, so hopefully we can peacefully coexist.

They've doubled down on the Copilot naming after the project was well underway, and I don't really want to rename it in anticipation of trouble that may never materialize

Thank you!

We're starting with react web, planning to expand from there to other surfaces, including React Native.

I suspect the gap from the current version, to React Native support is rather small. It's open-source, so feel free to open an issue so that it can get upvoted/prioritized, (or even better, you can work on it too ;) )

Thanks!

And Pretty much! You have to wrap the app (or the part you want copilot in) with a <CopilotProvider /> - this is UI-agnostic, business logic only. Then there are a few customizable UI providers you can use, or you can just build your own if needed.

If you check it out do let me know what you think!

Thanks for the pointer, fixed!

Open-Source (MIT).

If you'd like to contribute to the repo, or to the open copilot protocol, please join the GitHub/Discord!

Makes me wonder if there are real (text) use-cases for the mega context windows we've been seeing.

Usually these types of articles are total nonsense, but this is legit.

A very cool result!

It'd be interesting to see how many other systems can be approximated by the system they've solved for (without incurring an exponential penalty in the translation).

I love clojure, but over time I've learned the painful lesson that at the end of the day, for complex modern applications, the ecosystem is all.

That said Clojure ecosystem isn't too bad

Been working on an open-source platform for building copilot experiences into applications.

It'd be interesting to test and find out.

As the article shows, there is some evidence that OpenAI may be using a new embeddings model under the hood of assistants retrieval. If they are, and if it's substantially better than the competition, then open-source RAG may lag for a while.

--

But if they're just using ada v2 (or if the embeddings improvement is in cost, rather than performance), there should be tremendous potential for open-source models in this space.

First of all, ada v2 is an aging model that has solid open-source competition.

But more importantly, it seems the key is an LLM agent loop that can best make use of the RAG primitives. Intuitively, I'd expect open-source models to be smart enough for very good results in this domain.

TL;DR: (RAG + GPT-4) delivers superior performance, at 4% of the cost.

Looks cool. What do you mean by cursor-style insertions/edits?

Copilots are considered unsexy bc they appear not to aim at AGI (i.e. go for full autonomy). But that gets the game theory completely backwards.

LLMs (and agents) can often do 90% of the heavy lifting, but those 10% failures (or just weak results) can unpredictably pop up anywhere. Even 1% failures can sometimes render an AI solution uneconomical (see: self-driving cars).

The reasonable path to AGI is to have a tight AI/human feedback loop, and take more and more off of the human's plate over time. That way you can deliver actual value to users right now (not just a cool demo), and provide more and more value as the AI ecosystem matures.