ChatGPT made a lot of noise late last year, when it was unveiled to the general public. Since then, it has been at the center of attention of the general public and the media, as well as players in the field of AI, with Google at the top of that list.
It has to be said that it has made a huge splash. As Tim O’Reilly has accurately suggested, adoption matters more than innovation. And that’s the main advantage of this service: anyone can take it on and immediately assess its potential. The ripple effect has been such that, after less than two months, ChatGPT already had 100 million active users.
A flurry of announcements soon emerged: Microsoft announced that it would be investing $10 billion in OpenAI and integrating ChatGPT with its Bing search engine (in beta testing at the time of writing). Almost immediately following the release of ChatGPT, Google unveiled Bard, a rival service, whose launch is imminent, and whose results are less impressive, at least for the time being.
The impact on content production, search engine optimization and product discoverability is also significant, although it has yet to be quantified. That being said, the community of SEO experts has taken a keen interest in the tool, partly to see how it can help them – and how it might possibly replace them one day (although not in its current state, in our opinion). Hundreds of articles have been written on this subject, so we won’t rehash it any further.
There is, however, one topic that has received relatively little coverage: how can this service be utilized in digital interfaces? And above all, what are its concrete use cases?
How does ChatGPT work?
ChatGPT is a large language model that has been trained on data from multiple sources, many of which were supervised by human users, allowing users to engage in natural, human conversation with the application. In a way, it is the culmination of the conversational chatbots that emerged with the arrival of voice assistants (Google Home, Siri, etc.), with the difference being that ChatGPT focuses solely on text, not voice. It should be noted that the current dataset’s cutoff year is 2021 and the service doesn’t cite its sources (unlike the chatbot used by Bing). So, if you ask about events that are too recent, ChatGPT may cobble together information that usually tends to be untrue.
Nonetheless, although ChatGPT is a powerful tool, it isn’t perfect. It may be inaccurate and incorrectly interpret sentences, especially wording that includes subtle nuances or sarcasm, or it may even completely invent information, usually to go along with the person asking the question (with the exception of cases where it has been specifically restricted by its designers, for example, to avoid explaining why the Earth is flat). ChatGPT can even go so far as to corroborate its statements by citing scientific articles that don’t exist. In a way, it could be the world’s best generator of fake news! This stems from the fact that it is, first and foremost, a language model that was designed to pair words together based on different degrees of probability, and is not, strictly speaking, an artificial intelligence model. In addition, it remains unspecialized, meaning it hasn’t been specifically trained in a particular field or sector.
What are ChatGPT’s integration capabilities?
ChatGPT is based on a general-purpose model, GPT-3 (Generative Pre-trained Transformer 3), and does not have its own API.
What’s the difference between ChatGPT and GPT-3? ChatGPT is a conversational model based on GPT-3’s set of APIs, which were developed by the company OpenAI, under a proprietary license.
ChatGPT has been optimized to be a conversational, while GPT-3 is better suited to data modelling tasks. Although both models are based on the same architecture and use similar learning methods, the major difference between ChatGPT and GPT-3 is linked to the targeted uses of each model.
GPT-3’s APIs let you configure various models (or engines), to deliver the most appropriate answers to questions, while optimizing their ROI. For example, the Ada engine is faster and costs $0.0004 per 1,000 queries, while Vinci is the most accurate engine and costs $0.02 for the same number of queries.
However, to date, these models can only work with predefined, non-dynamic datasets. For example, if your bot is trained to answer the question, “what are your business hours?”, with the answer, “9 am to 7 pm,” and you then change your hours, you will need to retrain the model.
Does that mean you can't integrate ChatGPT's with your website?
No, not in its current state. At least, not yet… OpenAI is working with our partner Microsoft on a dedicated API that aims to integrate the ChatGPT model with the catalog of services available through Microsoft Azure (which is already integrated with the GPT-3 APIs).
It remains to be seen with ChatGPT’s API will look like and, above all, its ability to integrate with third-party datasets. That’s the whole issue: it must be possible to inject data and expand the model’s knowledge, while making sure that the new knowledge will take precedence over the model’s own inventions.
We can already imagine some helpful use cases that would allow you to deliver advanced, tailor-made conversational AI experiences to your customers (and your internal organization):
- The most obvious one: designing virtual agents that can interact naturally with users, so as to offer them personalized experiences; this kind of AI can be used in many different situations, like customer service, online shopping, advertising, education, and internal HR applications
- Personalization management: ChatGPT can help you create lead generation campaigns by inviting users to ask questions related to their interests
- Search: integrating ChatGPT with your search engine to make it more conversational!
Finally, ChatGPT is a more conversational version of GPT-3, because it can remember the context and provides an enhanced natural language model. In a way, it’s an excellent illustration of what can be done with GPT-3!
While waiting for ChatGPT’s API to be released, can GPT-3’s APIs be useful to me?
Yes! GPT-3 provides services for copywriting, classification, syntactic analysis and text structuring tasks.
We have great confidence in GPT-3’s ability to structure data. It is capable of converting unstructured data into a structured format, which can help extract key information (entities) from a text.
Let’s look at the example of a product description with the following input:
The Hero World Cup 90 SC is a racing ski boot developed by Rossignol for junior racing champions!
Inspired by the Race range, the Hero World Cup 90 SC uses Dual Core technology to very precisely regulate the boot’s flex, rear support and rebound. It also improves lateral power transmission for optimal efficiency.
Its intermediate flex rating (90) provides the best balance between comfort and performance. The Full Custom T2 liner includes high-density padding to optimize precision and performance.
The Hero World Cup 90 SC is the perfect racing boot for young competitors who want to improve!
And the following output:
{
“brand": "Rossignol",
“model": "Hero World Cup 90 SC",
“level": “junior competition",
“technology": [
"Dual Core",
“enhanced lateral power transmission"
],
"flex": 90,
“liner": {
“model": "Full Custom T2",
“padding": “high-density"
},
“characteristics": [
“very precisely regulated flex",
“regulated rear support and rebound",
“best balance between comfort and performance",
“optimized precision and performance"
]
}In other words, you can generate coherent text from a more or less structured data source. It’s particularly useful for a PIM (product information manager), in which you can automatically generate wonderful product descriptions from an API! And vice versa.
While this use case is especially applicable to e-commerce, it can also be applied to all kinds of data storage systems (databases, data warehouses, etc.) and data flows (APIs, microservices, etc.) to federate data and transform it into a coherent text. This is particularly useful when you want to unify your data silos and create seamless digital paths!
That’s great! But couldn’t we already do all that?
Yes, but not with this ease of access. ChatGPT is revolutionary for its ability to make a conversational AI readily usable by the general public, suggesting new potential uses, although it’s true that there are alternative models that may be more relevant because they are trained on controlled datasets.
For example, ContentSide is a French vendor that has developed an AI model that you can train on your datasets and that is notably used by online media (https://www.contentside.com/en/presse-en-ligne/). Please don’t hesitate to contact our data experts for more information!
