Impact of GPT3 on Dialogflow and other chat platforms
tneogi
3 replies
With GPT3's latest model increasing the context window to allow up to 4000 tokens, its completely possible to get very advanced chatbots with almost no intent training. In my quick tests I have even found that not only doest the latest model quickly learn to identify intents and entities, but it also is able to identify unseen but correct entities.
While I don't have test or at scale data to back this, I am confident that GPT3 will become the defacto chatbot platform soon, if not already.
For all conversational AI product folks here, I would love to hear your thoughts on this matter and contrarian viewpoints as well, in case I am missing something.
What would be the impact of this on existing massive chatbot platforms like Google's diaglogflow and others? My sense is that chatbot platforms will be the first (in a series) of casualties of Open AI models.
Replies
Richard Gao@richard_gao2
Evoke
Although GPT-3 will become the go-to chatbot platform due to just being plain good, there are likely gaps left in the market due to OpenAI's content restrictions and GPT-3's poor ability with other languages.
You can see chatbot apps like Replika and a few others pop up that rely on GPT-J or NEOX (which are open source) rather than GPT-3 because they allow for more "content policy violating" material in OpenAI's eyes.
But bear in mind these produce worse results.
So I can definitely see GPT-3 being a very good corporate chatbot for customer service, personal assistant, and some niche applications as a children's chatbot, there are definitely some areas which I would lack that other players can fill in.
I'm currently also working on making uncensored open source LLMs more accessible through an API for developers making AI apps. If you're interested, check out evoke-app.com.
We're almost done working on our stable diffusion API, so you can sign up for notifications for now, but we'll begin stuff will LLMs in the near future.
Share
@tneogi Thanks for the support! I'll def let you know. I'm on twitter @TheRealEtch so you should be able to find it really easily if you follow and I post it on my timeline.
@richard_gao2 interesting inputs. Thanks for sharing! Yes, would definitely like to check out the models for our work on Adaptiv.Me. Please let me know when the API is ready!