Comment by lee101 | Hacker News Reader

Comment by lee101 | Hacker News Reader

Parent Full thread

I run https://text-generator.io competing with OpenAI. I used to use kubernetes on Google cloud, but now run it from my house behind a CloudFlare tunnel believe it or not Using my own GPUs allows me to undercut them with around 10x cheaper pricing.

The product is built with some open source and some retrained models, it analyses input links or linked images with tesseract/unified IO / OFA / beautiful soup so it can answer questions about images/receipts etc.

The site itself is on app engine standard python with stripe/firebase UI for auth, which is running very cost effectively.

In future in planning to layer on more content understanding eg audio with whisper etc, then milvus/haystack for lookup phases for more factual question answering, then some work on automated retraining and regeneration etc for hard cases, then searching other systems/internal code and logic execution etc, eventually with a few more steps I'm hoping it will feel like I made AGI in a rack in my house lol...

Cool! I haven't tried, but thought I'd just let you know: it should be "Text Generator provides competitively (!) priced AI text generation..." you don't want people starting to question the grammar of a text generation tool right on the landing page

OT but can these AIs generate text based on existing corpus? eg helpdesk system that suggests responses to (human) agents based on previous (human) questions and replies or a bank of frequent saved snippets/replies.

edit: or find previous conversations with high similarity

Right now the text generation API can be used to do the conversational stuff along with a massive amount of other stuff like classification or adding questions about what is in text or how proper would respond to things for analysis purposes.

There's an embeddings API for text/code/images that can bed used for the similar snippets stuff if you bring your own vector search engine like milvus etc, but eventually I'll support uploading sets of documents and do the finding relevant snippets behind the scenes so you can ask these kinds of questions e.g. so people can do custom site specific FAQ answering bots etc.

There's an example in the playground of shorter autocomplete which is another way of helping agents too that it supports

Thanks for the tip about milvus, which might help on its own. I’ll try the playground when I get a chance as well.

shameless plug but that's exactly what we're building at Aide (https://aide.app), I would love to chat and see how we can potentially help if this is something you're interesting in learning more about or deploying!

Nice signup flow. Just tonight I was wanting GFN-GAN + a more light/noise/color cleanup ML model combined to restore old photos. Gluing a couple together would be super useful.

Yea, images/video is interesting, I think combining textual instruct and dalle/other models would be interesting so you could describe literally "restore this old photo and upscale+colorize it".

Was thinking of ways to do that kind of thing, I think unified IO is a good start at some flexibility but still lots of peices missing.

In the world of video training your own video filters based on editing a few frames is quite common E.g. something like nuke copycat or using ebsynth etc.

There's likely some new wave of DSLs or visual workflow builders coming that are either more ml friendly and or specific to building a conversational system/ image editing pipeline etc

This is so cool

Thank you! Means a lot :)

You're welcome. I have to guard my heart from envy when I read one person being able to pull something like that off.

Reply on news.ycombinator.com