AI Generative: from hype to concrete business applications25 October 2023

Unless you’ve spent the year on a desert island, you’ve heard of Chat GPT, Dall-E, LLM¹ and AI Generative. Perhaps you use them every day. Maybe your CEO asks you every day about their impact.
Well, we’ve got good news for you: we’re beginning to see things more clearly, and the impacts are going to be huge, but not necessarily where you’d expect them!

First of all, a few reminders if you haven’t done a PhD in Artificial Intelligence. What we pompously call Artificial Intelligence is a part of Machine Learning, basically a sophisticated application of the linear regression. You take a scatter plot and draw a line that follows the cloud: the slope of the line becomes the “rule” learned by the machine. If you combine this billions of times over, along with a number of other clever ideas that have emerged over the last few decades, such as word vectorization² or “attention”³, and allow super-powerful computers to learn from everything you can find on the Internet and beyond, something magical happens: the machine seems to understand what you’re saying and responds intelligently. It can translate, generate text, poems, summarize, generate images from text or sketches.

Of course, it’s amazing!

But there are limits

Firstly, the machine doesn’t really “understand”, but simply predicts the next most likely word – or pixel – based on everything it has read or seen. This results in some pretty comical responses, sometimes funny, sometimes shocking, but quite often wrong (37% according to one study⁴). So we try to “train” it with humans who fine-tune it.
Then, if you use the public version, your questions help it to progress. If you ask it to summarize your strategy, it can reuse it to answer a question from your competitor.
Finally, the machine doesn’t know your company, your systems, your processes, or the real world. So it’s completely irrelevant when it comes to answering a specific question about your business.

All this means that actual adoption in the enterprise is, in the end, fairly low.

But that’s all changing

Indeed, recent developments mean that there are many ways to teach the machine to reason, by following processes, calling on your knowledge bases or your systems. All in a private, secure environment, where you retain intellectual ownership of your developments and learning.
And that’s revolutionary.
Language models like GPT are great for understanding language and formulating an answer. Not necessarily for finding the right answer. The answer lies with you, but the models make it accessible, internally or to your customers, within a controlled framework.

From there, a host of applications are possible, from the simple interrogation of a corpus of documents or customer reviews, to a sophisticated automaton capable of accompanying a customer, taking over a repetitive task, or even helping you with more creative tasks.

We’ve successfully implemented such applications (bot) for our customers, and we’re at your disposal to imagine your use cases and develop them with agile, high-performance teams, within reasonable timescales.

Notes and references
1. LLM Large Language Model
2. https://fr.wikipedia.org/wiki/Word2vec
3. https://arxiv.org/pdf/1706.03762v3.pdf
4. https://arxiv.org/pdf/2302.04023.pdf