Is ChatGPT ready for customer relations?
The ChatGPT model, launched by OpenAI, has made a strong impression with its ability to engage in incredibly smooth dialogues. But is it truly ready to handle real customer inquiries? What can we ultimately learn from it?
Launched with great fanfare in November 2022, ChatGPT became the first online service to reach one million users within a week. As detailed in our previous article, this ultra-large language model (LLM) integrates text-based analogies learned from extensive web corpora and digital archives, spanning science, industries, humanities, and even code debugging. Additionally, ChatGPT has been fine-tuned through reinforcement learning to handle sustained dialogues while being trained to avoid sensitive topics (violence, racism, phishing). Despite these safeguards, it remains open to "prompting" or "prompt engineering" to shape its conversations and text/code outputs. Thus, experiments and analyses have poured in...
Strengths and weaknesses of ChatGPT
On the strengths side, there is no doubt that this mega-model generates incredibly fluid responses across a wide range of topics with minimal technical adjustments, beyond basic "prompting." Responses are syntactically and semantically correct (especially in English, though minor syntax errors may occasionally appear in French) and tend to maintain coherence throughout the conversation. Moreover, the generated responses are often dense with information, sometimes providing additional context or definitions that were not explicitly requested, adding layers of justification or extra details at the end.
However, this strength directly leads to ChatGPT's main weakness: the uncontrolled nature of the information it generates. While responses can be accurate, they can just as easily be false (e.g., fabricating authors and scientific articles or offering highly credible but incorrect IT diagnoses) or misleading (e.g., mixing up-to-date and outdated information by a decade), such as when recommending product lines. Additionally, responses can be overly verbose or irrelevant, as seen in simulations of banking callbots. In the examples below, the model's ability to generate sentences is impressive, but the first snippet includes contradictory questions and suggestions without allowing the interlocutor to respond, while the second snippet reveals a misunderstanding of bank cards based on an American model. It becomes quite challenging to ensure professional compliance when responses are generated on the fly.
Is ChatGPT usable in production?
A significant response comes from Stackoverflow, which banned ChatGPT from its developer support forums due to the noise caused by the model’s highly credible but false answers.
Another response comes directly from OpenAI, the company behind ChatGPT, which has been partnered with Microsoft since 2020. OpenAI projects a revenue of $1 billion by 2024, but ChatGPT itself is not for sale at this time. The LLM, which has garnered millions of users and widespread attention, is mainly serving as a showcase for other ready-to-use models from OpenAI, which formed the basis of its training. OpenAI’s GPT-3 family of LLMs, for instance, are available with explicit pricing—pricing that increases sixfold if you want to use a fine-tuned version tailored to a specific task and domain via supervised learning.
So, both external and internal answers are clear: No, ChatGPT is not intended for production use. However, other fine-tuned LLMs might be...
What’s the Outlook for Customer Service?
Once we accept that ChatGPT is designed to engage beta-testers and influencers rather than being a production-ready system, the question remains: do LLMs, and deep learning more broadly, have a role to play in the future of customer service?
At ViaDialog, we believe the answer is yes. While our stack already integrates the latest neural modules in a controlled and production-proven environment, we continue to prepare for the future, aiming to offer dialogue systems that are even more fluid, efficient, and competent. Stay tuned for our upcoming announcements on this exciting topic!
Article written by Ariane Nabeth-Halber, AI Director @ViaDialog
Learn More...
A good introduction to word2vec, the foundational component of Large Language Models (LLMs), also known as embeddings or "data plunging." This is a word-to-vector encoding that 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 initial breakthroughs 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 GPT-3 equivalent from 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.Google's ChatGPT equivalent:
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.The rise and fall of Facebook-Meta's Galactica:
https://arxiv.org/abs/2211.09085
https://www.cnetfrance.fr/news/couac-pour-galactica-l-ia-de-meta-formee-a-lire-des-articles-scientifiques-39950046.htm
