I see this is using GPT4o, any plans for something more sustainable? Would be interesting to see an https://openfreemap.org for LLMs.

Perhaps via an RNN like in https://huggingface.co/spaces/BlinkDL/RWKV-Gradio-2

Or even just leverage huggingface gradio spaces? (most are Gradio apps that expose APIs https://www.gradio.app/guides/view-api-page)

I wonder if a 1B model could be close to free to host. That's an eventuality, but I wonder how long it'll take for that to be real.

I’m planning to deploy a 1B model, feed it all the documents I’ve ever written, host it on a $149 mini-PC in my bedroom, and enable you to chat with it.

I’ve released similar projects before.

I’ll drop a post about my plans in the coming days and I’ll build and document it about two weeks later if there’s enough interest.

joeldare.com

The post about my plan is now online:

https://joeldare.com/my_plan_to_build_an_ai_chat_bot_in_my_b...

Sounds cool!

A 1B model at 2-bit quantization is about the size of the average web page anymore. With some WebGPU support you could run such a model in a browser.

I'm half joking. Web pages are ludicrously fat these days.

Something like this? https://github.com/huggingface/transformers.js-examples/tree...

That was the exact thing I was thinking of but couldn't remember the project name. Thanks!

These videos are incredibly life like. I can see many virtual companions being made with this.

Glad you like it!

I don’t see how this is different from using the pdb Python module? You can catch exceptions and use to write functions.

The biggest distinction I see is the ability to continue from exceptions.

I find the other differences to be minor in comparison - to be able to recover and continue execution following post-mortem debugging is amazing.

Here the neural network was given examples of how to use the calculator for each question which means it wasn't generating it's own abstractions.

If you wanted to use this to solve other (e.g. programming) problems you would need examples of every step required for almost every problem.

Using neural networks in this way is akin to locality sensitive hashing, instead it should understand what it's lowest level operators do and discover useful combinations of them that can solve new problems.

I've seen one that pairs people to meet up every week.

It's just hard to incentivise. I wonder what would? Would being payed work?

I was inspired to work on this by Bret Victor's "Inventing on Principle" lecture https://youtu.be/PUv66718DII

Any feedback appreciated!

I found this really useful for my articles. The short snippets would work well as tweets, a tweet bot for article summaries would be cool!

We are glad that you liked it.

That's a very good suggestion. By tweet bot, you mean when someone tweets an article link to the bot, it would respond back with the summary of that article, right?

Thank you very much for taking the time to use the app and provide your feedback.

Yeah, or I could comment on a tweet linking to an article with "@TLDR" and your bot would comment the summary.

The previewing execution is cool, reminds me of my Python debugger https://marketplace.visualstudio.com/items?itemName=fraser.l...

If this takes of it will create an awesome dataset for a description to code ai!

Creator here,

The idea is to have one site that can brings you everything you check for into one place but also helps you discover great sites you've never seen before.

Just finished this today so Im super keen for any feedback anyone might have.