OpenAI Optimised LLMs For Dialogue With ChatGPT
How does ChatGPT compare with the new DaVinci 3 model? Is ChatGPT an implementation of KI-NLP? And how OpenAI is avoiding possible disaster.
Since 2020 I have been doing research and writing on OpenAI and GPT3…
The release of ChatGPT needs to be considered in the light of the following points:
- GPT3 has always had a very good conversational capability where dialog state and dialog turns where managed successfully. Conversation context was also managed.
- GPT3 has not really been susceptible to hallucination and other ailments like BlenderBot.
- GPT3’s answers are well formed, succinct and coherent. This has improved considerably with the recent introduction of the new DaVinci 3 model.
- If I had to create/bootstrap a chatbot using the OpenAI Language API/GPT3, I will bootstrap the chatbot using the method I describe here.
⭐️ Please follow me on LinkedIn for updates on Conversational AI ⭐️
Critical Considerations:
- Considering the examples below, it does seem like the OpenAI langauge API has access to the internet and are able to retrieve answers.
- ChatGPT does not have access to the internet for information retrieval and seems to be an implementation of Knowledge Intensive Natural Language Processing (KI-NLP).
- What is KI-NLP? It is a very broad domain, knowledge intensive interface for question-answering or fact-checking tasks, known collectively as knowledge-intensive natural language processing (KI-NLP).
- The AI models underpinned by a KI-NLP framework searches through a digital archive for relevant information. The more comprehensive the digital archive, the broader and correct the answers.
- KI-NLP systems are self-contained and does not search the internet or make use of knowledge API’s.
- OpenAI is making sure that ChatGPT is not learning bad habits by questionable internet content or controversial user input, like in the case of Blender Bot.
⭐️ Please follow me on LinkedIn for updates on Conversational AI ⭐️
Practical Examples:
Below I ask the DaVinci 3 model two questions, the second being a follow-up question relying on context from the first question:
How do I create a chatbot?
What technology should I use?
The responses are much more verbose, well ordered and sequenced in Davinci 3 than in the Davinci 2 or Davinci 1 model. Read more on that comparison here.
However, posing the same two questions to ChatGPT, it is clear that the response are much more conversational and comprehensive.
In this second example I ask three questions:
Who is Peter Thiel?
Where does he live?
How old is he?
The Davinci 3 generative model manages the conversational dialog turns, state and context very well.
Below, you see the response to the same three questions from ChatGPT…it is clear that the integrity of ChatGPT is being protected…but also it is evident that ChatGPT is a self-contained LLM, or KI-NLP implementation.
Some Final Thoughts
- ChatGPT is a blend of all four of the elements listed below, Dialog Management, Generation, Question & Answer and Language Translation.
- Here is a perfect translation from Afrikaans to English from ChatGPT, the language was automatically detected by ChatGPT and only the correct sentence was translated.
- An ChatGPT API will have to be exposed for production implementations, which is not currently available.
- Fine-Tuning is a consideration for enterprise implementations.
- And related to fine-tuning, methods of incorporating specific knowledge is most probably on the roadmap.
⭐️ Please follow me on LinkedIn for updates on Conversational AI ⭐️
I’m currently the Chief Evangelist @ HumanFirst. I explore and write about all things at the intersection of AI and language; ranging from LLMs, Chatbots, Voicebots, Development Frameworks, Data-Centric latent spaces and more.
