To understand the change induced by ChatGPT, we must go back to the Paleolithic of the Web. In the beginning was the yearbook. Yahoo! listed lists of links classified by headings. It was 1994, when the entire World Wide Web was less than 3,000 sites, but was tripling every year.
The big leap took place in 1998, when two Stanford students came up with a brilliant idea: an algorithm allowing sites to select the most popular among them based on the number of links that will converge on each one. It is on this principle of voting by the multitude that Google was founded. Advantage of the system: it works without human intervention and it scale wonderfully, in other words, it can handle the immensity of the Web without perceptible limit (the Web today has 200 million active sites).
Succession
Twenty years later comes GPT-3, the first text generator available to the general public. Its operation could not be more intuitive: you submit a subject to it, and it develops it in such good English that its use is prohibited in American education. In fact, he is content to regurgitate the 500 billion words he has been stuffed with: press articles, blogs, scientific papers, the entirety of Wikipedia and all the books available. This is called an LLM, for large language model.
2022 marks an essential evolution, with the possibility of communicating with GPT, hence its name, ChatGPT. Unlike a traditional search engine, it returns complete paragraphs for each question asked. Major disadvantage – and risk –: it gives neither source nor context; it affirms, perfectly simulates expertise and knowledge. ChatGPT covers a wide spectrum: it is capable of writing a cover letter or a complete diet plan. But it is never possible to understand how he proceeds, nor to reconstruct the origin of his knowledge.
In early January, Microsoft made the strategic decision to connect ChatGPT to its own search engine, Bing, which caps at 3% market share, compared to 92% for Google. If Bing became a natural language dialog interface, its adoption would be guaranteed. The Seattle company therefore took virtual control of OpenAI, the company that created GPT.
But, as always, the devil is in the execution. And, fascinating as it is, ChatGPT remains a prototype with many uncertainties. Its size, first. To generate a response, the robot requires phenomenal computing power – Microsoft knows something about this since it is the one that provides the thousands of machines. Researchers have estimated the cost of processing a request at 5 cents. However, Google carries out 8.5 billion searches per day, which would make 425 million dollars in daily expenditure, 155 billion per year, just to amortize the computer time, to which must be added the vast technical, commercial infrastructure, etc. . By way of comparison, Google achieves 260 billion dollars in annual turnover, with 150,000 employees. So ChatGPT clearly has a problem moving from a promising testbed to a mainstream product working on a planetary scale.
Next, Google has no intention of being left behind in the field of artificial intelligence. The firm has developed numerous tools for creating algorithms for machine learning which have been put into the public domain – just like those of Microsoft and Amazon, the challenge being to create global standards. Quite discreetly, Google Search already integrates AI into the heart of its search engine, for example to reformulate a query so that it generates the most relevant answers, and this is only the beginning. Google is finally building its own LLM called Pathways, even more powerful than ChatGPT. No one really knows how it will be deployed for Internet search, whether it will supplement or replace traditional Google results lists.
But this transition to information without source or context is a dangerous drift towards the absolute reign of the informational black box, with its biases, its internal logic, the uncertainty about the data that was used to learn it. Problem: a search engine stripped of all complexity is likely to seduce a large audience that does not care about the source of information. If there is one constant in tech, it is that the appeal of simplicity is universal.