mpaepper

Yep, that's the way it's currently implemented in langchain.

The 4 is a hyperparameter you can change, though, so you could set it to 10 as well.

The way it works is that it first looks up the N most relevant documents (N being 4 in the default case) in the FAISS store relevant to the question, so it uses distance of embedding vectors for this lookup.

Then it uses GPT3 to get summaries of the 4 entries related to the question and finally all the summaries together with the question will lead to the answer.

In doing so, you can trace the source where the answer came from and can also point to that URL in the end.

When you make N larger it just gets more expensive in terms of your API costs.

Awesome, thank you so much. That is solid advice :)

Can you showcase any projects you've worked at in the mean time? That way you could fill some gaps with project stories which might make it easier for you. Good luck with the search!

I explained it in a more coding oriented style here: https://www.paepper.com/blog/posts/eigenvectors_eigenvalues_...

Thanks for the mentions everyone. I've already launched text2shirt some weeks ago - no need for audio, online and serving shirts and hoodies:

www.thisshirtexists.com

I work in machine learning for digital Pathology and I think the big problem here is the divergence between publishing papers and real life helpful models. What we see is that in the literature you often see a model trained on data from a single lab and gets crazy good results. However, apply it to a different lab (not in the paper of course) and it sucks. So what we do in practice for our models is to train on many different labs at the same time and have a hold out test set with labs not covered during training. That way you get a robust model which works well in practice (but doesn't have the 99.9% metrics reported in papers). The second thing is looking at what task to let the machine do. We typically go for tasks which are boring / repetitive for the doctor, but visualize the result, so the doctor double checks it before making the diagnosis. Still saves a lot of time.

Try it out yourself: simply answer anything to this thread https://twitter.com/mpaepper/status/1582622125390643200 and your text will be converted to an image and added to my profile banner. Takes around 20 seconds after you post sth.

If you use this in combination with Stripe, then what do you gain in addition to Stripe?

When you reply to my tweet, a bot will generate an image from your comment with stable diffusion.

The images are then shown in my profile banner after 20s.

Very cool idea! Unfortunately not working for me. I tried with the Chrome extension, but when I stop a recording I get forwarded to "https://app.supademo.com/demo/undefined" and there is no demo anywhere.

Yep, correct. But it's expensive if you have traffic :)

I also wrote a blog post about stable diffusion - how it works technically.

In case you want to go deeper down the rabbit hole:

https://www.paepper.com/blog/posts/how-and-why-stable-diffus...

Interesting idea, might consider to have it as an option. Thanks!

I have a small server which runs the main application which is quite cheap, but the stable diffusion generations are run via an API which costs around 1€ / 100 image generations.

This Show HN cost me around 150 € via the API so far and I had 3 orders, so I guess this will be more of a cost center :D

I might. Send me a private message on Twitter: Twitter.com/mpaepper Thanks!

Yes, I was torn between not censoring at all, but some results friends pointed out seemed too bad to have them be seen on my site, so I build a small filter for now.

Yep, that's correct. Thanks for the hints!

Currently, I review all created designs before printing, so that would be a way of not infringing, I. E. Cancel orders which are potentially infringing

You can somehow influence the emotions by the words you enter, though. So if you use intense words, it will take that into account as it has learned so from training on a lot of image Word pairs.

That's a very cool idea :) I would definitely wear that!

How would they know that a customer asked?

Haha, wow, that's not a prompt where I would've expected NSFW!

Thanks for the hint, I'm not a native English speaker, so I probably need to enter a couple more forbidden words.

Sorry, it might not be super clear everywhere on the site. That's because initially I only generated a single image per query, but recently in addition I added the Lexica API to showcase more pictures per query as inspiration.

No the license is the same, both CC0 free for all.

That's interesting, thanks! I actually generate one image per query and use the lexica API to sample others that are similar / related

It's called dropshipping. I have a provider who gets the GAN upsampled image and the order info. They print and send it to the customer.

Do they automatically upsample the images to print quality with a GAN model? That's what I'm doing in the background.

I see it this way: I provide a means for users to generate what they like, but I'm not mass selling those designs. If a single person uses tshirt pens to paint mickey mouse on their shirt, Disney would probably not sue them. Also there is no way knowing whether I sold any mickey mouse shirts.

Thank you :)

Yes, I use the lexica API and in addition generate a unique image with stable diffusion!