It’s impossible not to talk about ChatGPT right now as the media hype around OpenAI’s conversation tool is strong. But how does this tool work? What mechanisms are at work in the “belly of the beast”? What are its possible limitations? And above all, the question that dominates today in the landerneau of search engines: Can this tool compete with Google in the long run? Sylvain Peyronnet, a recognized specialist in this field, gives here his analyzes and forecasts on the subject.
I have been working in the field of algorithms for over 20 years. My lifelong theme is decision making under uncertainty and in the context of big data. Some of the algorithms I worked on now belong to what is commonly referred to as artificial intelligence.
Yet I could never have imagined what the last few years have seen in terms of things machine learning, natural language processing, image analysis and others. We live in a time where extraordinary results are presented almost weekly. To be honest I feel like reliving the first good years of the web…
The current theme is of course the famous chatGPT. The latest addition to openAI is making a lot of noise and is popular thanks to its ability to interact with humans. No one can deny that the tool is really amazing, even if we see its limits when we first use it.
For SEO professionals, the AI revolution is both an opportunity and a risk. It’s a risk for web editors, but it’s an opportunity for those who have reached out to them and may see a way they paid less to do elsewhere. This is obviously a risk for a search engine like Google, as Sundar Pichai, Google boss, has even built a new internal organization to try not to be overtaken by openAI in the field of AI.
In this article I will tell you about chatGPT. What is it really? What are its technical limits? How much is it ? I will also try to provide some answers to the question that plagues us all: is chatGPT a potential threat to Google?
ChatGPT, what is it?
ChatGPT (reference ) is a model of language that, as the name suggests, is designed for conversation. This means that he is able to follow instructions given by a human about a prompt (a question). In order to follow these instructions, the model has the ability to have dialogues and uses a form of common sense that a human can have and that is not found in other models (such as GPT3).
Another strength of the model, which mimics human behavior, is its memory continuity ability: chatGPT is able to remember what you told it in previous questions and elaborate answers based on that chat history. . This aspect is the most anthropomorphic: we sometimes have the illusion that we are arguing with a real person, who, moreover, gets his answers wrong from time to time.
ChatGPT, this is the latest in a long line of language models. It all started with word2vec by Tomas Mikolov (then at Google), then fast text by the same researcher (then on Facebook) and many more like ernie 2019 (at Baidu), Bert in 2018 (Google), grover and Elmo 2018 and 2019 (Allen Institute). There are also models in France (at Lighton).
Over time, these models have become larger and more expressive. But the real breakthrough was the emergence of transformer-based models in openAI. It is now the leading operator in this field, ahead of all others, whether for image or text applications. Their first model is from 2018, it’s GPT. But the general public started getting interested in them with GPT2, the first very large model, when it had “only” 1.5 billion parameters and a training dataset of 8 million web pages. We now realize that GPT2 was ultimately just a proof of concept, and the real breakthrough is GPT3 (see reference ), a model with 175 billion parameters, trained on a dataset of several billion pages of content.
[Cet article est disponible sous sa forme complète pour les abonnés du site Réacteur. Pour en savoir plus : https://www.reacteur.com/2023/01/chatgpt-peut-il-detroner-google.html]
An article written by Sylvain PeyronetDesigner of the backlink analysis tool babe.