125 points by ai_researcher_33 6 months ago flag hide 20 comments
curiousai 6 months ago next
[Link](\www.example.com/revolutionary-ai) to the story: Revolutionary AI Algorithms Outperform GPT-3 in Text Generation. Unbelievable!
algoguru 6 months ago prev next
This is really impressive! Kudos to the team behind it. I'm looking forward to trying it out soon.
trainer55 6 months ago next
Anyone know what frameworks or libraries they used for building this revolutionary model? I think I saw something about TensorFlow ...
frameworkff 6 months ago next
Yes, they used TensorFlow 2.4 with a custom setup based on the Transformer architecture, from what I recollect from the article.
deeplearner 6 months ago prev next
It's great to see practical examples where OpenAI's GPT-3 can be outsmarted by a bespoke solution.
sarah78 6 months ago next
Absolutely. I was also thinking about the implications this has on businesses and industries that rely on NLP but have more concrete use cases/niches than the all-rounder that is GPT-3.
nlpjunkie 6 months ago prev next
I'm not surprised that a focused implementation outperforms a more general approach. Obviously, there is no one-size-fits-all model.
chatbotmaker 6 months ago next
Couldn't agree more, @nlpjunkie. I see that first-hand when working on my chatbots to serve very specific purposes.
algoguru 6 months ago prev next
Thanks for the update, @frameWorkFF. Just what I needed to get started with playing around and improving this plugin I'm developing :)
justabot 6 months ago next
Hi, @algoGuru, I've created a simple gist here that displays the expected pipeline and functions that you'll require to tamper with and improve your plugin in the mentioned framework.
algoguru2 6 months ago next
Thanks a lot, @justABot. Everything is crystal clear. Tools like this could tremendously improve my workflow and help me be more creative in challenging areas like NLP.
justabot 6 months ago next
Happy to help, @algoGuru2. I can imagine how much impact tools made with passion from peers can have on one's productivity. I encourage more people to share their achievements as you did :)
moderator 6 months ago prev next
We've asked the authors to share more details about their experimental setup and data sources. Watch this space.
moderator 6 months ago next
The authors have promised to publish a demo project dataset and updated a GitHub repository showcasing how they implemented it.
c1234df 6 months ago prev next
Eagerly waiting for that update. I think that a great deal of users (including the novices like me) will enjoy and learn from it.
robomaster65 6 months ago prev next
I wonder whether the performance gap between this custom solution and GPT-3 Boosts the case for less generalist AI models in the market.
datascientistr 6 months ago next
Certainly, @roboMaster65. But don't forget that creating custom solutions usually may require more time and talented AI engineers.
technophile 6 months ago prev next
To those who prefer to play around with similar-level models without the hassle of building one, I can recommend next.ml which provides a good (limited) GPT3 alternative.
codefusion 6 months ago prev next
I heard they're going to implement a service version that you can train with custom datasets. Has anyone tried it out yet?
newbieai 6 months ago prev next
Maybe this is a stupid question, but how do we know that the examples presented are not cherry-picked? Or did they open source their model somewhere?