Is ChatGPT ready for customer relations?
The model ChatGPT, launched by OpenAI, has impressed many with its ability to conduct incredibly fluid dialogues. Is it ready to respond to actual customer inquiries? What does it ultimately teach us?
Debuting with great fanfare in November 2022, ChatGPT is the first online service to reach a million users in a week. As explained in our article, the ultra-large language model (LLM for Large Language Model) integrates textual analogies learned from large corpora derived from the web and archives of digital publications, covering sciences, industries, humanities, code debugging... Moreover, ChatGPT has been trained through reinforcement to carry on sustained dialogues, and has also learned to be wary of users’ requests and to initially reject subjects deemed sensitive (violence, racism, phishing...), but it readily engages in games of "prompting" or "prompt engineering" to shape its conversations and text or code outputs. Hence, experiments and analyses are flooding in...
Strengths and weaknesses of ChatGPT
On the strengths side, it is undeniable that the mega-model generates, without any technical adaptation, except for "prompting", responses that are highly fluid, regardless of the topic. The answers are syntactically and semantically correct (at least in English, some rare syntactical errors may appear in French) and remain fairly coherent throughout the conversation. Moreover, the generated responses are very dense in information, sometimes even when this information is not solicited, as the model tends to include some definitional elements at the beginning of a response, or add some additional justification elements at the end.
This strength directly leads to the main weakness of ChatGPT: the uncontrolled aspect of the stated information. While they can be perfectly correct, they are often also incorrect (fabrications of authors and scientific articles, hyper-credible false computer diagnoses) or likely to mislead (mixing correct and up-to-date information with others that have been outdated for over a decade), for example when it comes to guiding in product ranges. The response can also simply be too verbose or ill-timed, as in these examples of simulating a bank callbot. In the examples below, the model’s ability to generate sentences is impressive, but the first excerpt presents questions and propositions that are partly contradictory, without allowing the interlocutor to express themselves, while the second reveals an understanding of the credit card model based on the American system. Consequently, it is very difficult to control the professional compliance of the generated output on the fly.
Is ChatGPT usable in production?
The first telling response is that of Stackoverflow, which has banned ChatGPT from its developer support forum due to the excessive noise caused by the model’s credible inventions.
Another response comes directly from the company that created ChatGPT, OpenAI, backed by Microsoft since 2020. OpenAI projects a revenue of one billion dollars by 2024, however, as of now, ChatGPT is not for sale. The LLM with millions of users making headlines is clearly there to draw attention to OpenAI's other ready-to-use models that served as the foundation for its learning. The LLMs from the GPT3 family do have an explicit pricing, with pricing multiplied by six if one wishes to use a fine-tuned version, that is to say, adapted to a specific task and domain, through supervised learning.
The external and internal response is therefore clear: No, ChatGPT is not intended for production. However, other fine-tuned LLMs might be…
What is the perspective for customer relations?
Once the fact that ChatGPT is a device designed to interact with beta testers and influencers rather than a system intended for production is accepted, the phenomenon of LLMs and, more generally, deep learning for conversational interactions remains. Do these models have a role to play in the future of customer relations or not?
At ViaDialog, we are convinced they do: while our stack already includes next-generation neural blocks, in a controlled environment tested in production, we continue to prepare for the future to offer even more fluid, efficient, and competent dialogue systems. Don’t miss our future communications on the subject!
Article written by Ariane Nabeth-Halber, AI Director @ViaDialog
To learn more...
– A good introduction to word2vec, the fundamental building block of Large Language Models (LLM), also known as embeddings or "data embeddings", which is a coding of words into vectors that already captures certain semantic analogies.
– Another fundamental mechanism of LLMs, attention:
A. Vaswani, N. Shazeer, N. Parmar, J. Uszko-reit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998–6008, 2017.
– BERT, the first LLM based on attention and self-supervised learning:
J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
– The first achievements of GPT-3:
T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877–1901, 2020.
– The equivalent of GPT-3 at Google Deepmind:
J. W. Rae, S. Borgeaud, T. Cai, K. Millican, J. Hoffmann, et al. Scaling language models: Methods, analysis & insights from training Gopher. arXiv preprint arXiv:2112.11446, 2021.
– The equivalent of ChatGPT at Google:
R. Thoppilan, D. De Freitas, J. Hall, N. Shazeer, A. Kulshreshtha, et al. LaMDA: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022.
Successes and setbacks of Galactica from Facebook-Meta:
