No-Code NLP - Automated text generation with Eleuther.ai's GPT-Neo (GPT-3 alternative) and Bubble.io

[@wonderfulvamsi]

Thanks for making this video this is exactly what I need

[@LaunchableAI]

Glad I could help!

[@Null-Red-Blue]

Best video covering this. Still waiting for a service/platform that handles the ugly database and API stuff. Who *knew, having beta access who be so useless for laymen like me lol

[@LaunchableAI]

Thanks! By ugly database stuff & beta access, are you referring to trying to build something around GPT-3?

[@Null-Red-Blue]

@@LaunchableAI Yep indeed. I didn’t realize hooking up text boxes and buttons to API calls and functions is still a bit of a learning curve *EVEN in “no-code”. I’m chained to the Playground only because of this. A platform can/probably will come about and allow people to use their API keys but with a more polished and formulated approach specifically for OpenAI and EleutherAI models. I wish I could just write get the whole UI situation with Buttons and textboxes/radio button/dropdowns and then have predefined places to put the Prompt formula to apply it to particular dropdown tools in a spreadsheet or something. It will come about probably with Codex integration. I’d pay quick. Instead of trying Zeroqodes GPT-3 Bubble template. I don’t think Bubble is for me.

[@rowelski]

Do you have tutorial with doing simple blog intro? using NLPclouid

[@LaunchableAI]

I haven't done a tutorial with their platform yet, but it looks excellent. If you have a try, let us know how you go.

[@YoungUrbanProject]

This was a very informative video. Thanks for making this. I have a question - can we train the model further for a particular use case - like writing an email or a Facebook ad copy - within this bubble app you created?

[@LaunchableAI]

Glad you liked it! Regarding training the model for particular use cases: this process of further training is typically called "fine-tuning", and is often applied in the context of "transfer learning" (taking a pre-trained model and fine-tuning it to do something else). You can do this if you're comfortable with training models, but you can't do it through the Bubble app demonstrated in the video. The documentation for training a GPT-Neo-based model is here: github.com/eleutherai/gpt-neo. Eleuther.ai is now recommending use of their more recently developed architectures, however, such as GPT-J and GPT-NeoX. Fine-tuning instructions for GPT-J are here: github.com/kingoflolz/mesh-transformer-jax/blob/master/howto_finetune.md If you're interested in a model for email / Facebook ad copy, you'll need some example text to work with. "The Pile" (pile.eleuther.ai/) is the 825 gigabyte dataset used to train the GPT-Neo model used in the video, which is accessed through the Transformers platform/API. A small portion of that dataset is the Enron email corpus, so you'll have some email-specific text in there. As for Facebook ads, I don't believe it contains any of that, so you'd need to source your own content. Hope this helps!

[@YoungUrbanProject]

@@LaunchableAI thank you!!!

[@woongda]

Skip to 14:07 - probably all you need ... thanks for making it, but it is very difficult to hear your voice.

[@LaunchableAI]

Yeah, I've been learning audio engineering as I go along lol. Thanks for checking it out.

[@ashmichael5679]

I absolutely loved this video thank you for making it !! I have already shared your channel to a lot of my colleagues I have a quick questions is it possible to create an Ai chatbot where there is a two way communication with both text and voice using gpt-Neo/gpt-6j that can resemble Elon musk if I was to fine tune model and I have a 3D render how could everything be incorporated together and can it all be brought together to make something functional on bubble.io?

[@LaunchableAI]

Thanks for the kind words! Glad you liked it and shared it :) What your asking about is certainly doable, but would be a fairly complex project. You'd be combining several pieces that can be challenging in their own right: - speech-to-text (for user voice input) - natural language understanding (to parse what the user is trying to say) - generative language modelling (a la GPT-J/GPT-Neo, etc.) with fine-tuning, to produce Elon-output - probably some sort of GAN-driven visualisation (to create/animate the Elon-avatar) The Bubble side of things would probably be the easy part, lol. Getting all the AI components working would be the real challenge. My hat is off to you if you even attempt this, let alone accomplish it! If you do try it, please keep us in the loop!

[@ashmichael5679]

@@LaunchableAI thank you !! will definitely keep you in the loop I would love to jump on a short zoom call with you where we can discuss what you guys do and I can interview you on no-code and AI If you're satisfied with the interview we can provide you with the clips and edited short versions for TikTok or youtube shorts is there any way we could make this happen.

[@LaunchableAI]

​@@ashmichael5679 Would be happy to schedule a call. You can get in touch via email and we can set something up: contact @ {our domain name} - see description if domain name not obvious :)