Silicon Valley loves to get excited about new trends in a cyclical way. Facebook Messenger’s chatbots were on everyone’s lips in 2015-2016: at the time, they were supposed to replace mobile applications. Then it was the turn of the “hype” around Google Assistant, Siri and Cortana, supposed to replace the Web, quite simply. Today, text-generating AIs are the buzzword.
In July 2002, we were enthusiastic about DALL-E mini (recently renamed Craiyon), the consumer version (free and open) of this system created by OpenAI in 2021, which allows images to be created from a description textual. In September-October, it was the turn of LaMDA, the latest AI specialized in automatic text generation, developed for the needs of Google’s next conversational robots. This conversational AI is a priori able to understand what is asked of it, to deduce the meaning of human language, and to generate the most “natural” language possible. To the point that some ended up believing that she was “conscious”. Note that LaMDA also allows Google to explore assisted writing, with Wordcraft, a text editing tool intended to help writers find story ideas, and even write them. The battle seemed declared between OpenAI, the (for-profit) AI research laboratory created by Elon Musk and Peter Thiel in 2015, and Alphabet / Google.
The new buzz around ChatGPT, OpenAI’s new chatbot, confirms this impression. For 2-3 years, in parallel with Dall-E, the company has been developing a system capable of generating press articles, works of fiction and computer code from text recognition. This engine, GPT, is now in version 3. It is a “large language model” (large language model, LMM), trained with 175 billion parameters, from more than 500 billion texts from the web, encyclopedias and books. The result, in the form of a chatbot open to the general public on December 5, 2022, is striking.
Already tested by more than a million users According to OpenAI President Greg Brockman, ChatGPT is talked about because it is able to generate written content “autonomously”. It ranges from articles to jokes, poems, math problems and computer programs. I personally wanted to get to the bottom of it: could he have helped me write my previous article, on “Buy now, pay later” (BNPL)? His answers (below) to my questions (“what is the BNPL”, “how do companies that offer such services earn money?”, “Can you give me examples of companies”) are impressively accurate and useful.
“No, I’m not going to replace Google”
We inevitably come to wonder: with such a chatbot, does Google still have a use? Asking him questions related to the news, ChatGPT’s response falls like a cleaver: “I’m sorry, but I am not able to browse the Internet or access external information other than that for which I was trained”. Information that he has brewed and with which he has been “trained” until 2021. Ditto with Wikipedia: unlike the online encyclopedia, ChatGPT only answers your questions from necessarily obsolete data, since dating from two years ago.
When asked directly if he will “replace Google”, here are his answers: “No, I will not replace Google. I am an OpenAl-trained language model, and I am designed to answer user questions in a automated. I am able to provide information on a wide range of topics, but I am not a search engine like Google. I do not have access to the Internet to search for information, and I cannot post results from different sources like search engines do. To become stronger than Google, I would need to have access to a large volume of data and the ability to process it to provide relevant results to users.”
And if he was connected to the Net, then? “It is possible that I could be connected to the Internet and become an improved search engine. As a large language model, I have been trained to understand and process natural language, which allows me to answer questions on a variety of topics. If I were connected to the Internet, I could use this ability to help people find the information they are looking for online. However, as I am a virtual assistant and not a search engine real, my answers would always be limited by my area of knowledge and my ability to process information,” he replies.
Never mind, for Paul Buchheit, the creator of Gmail: for the former engineer of Alphabet, who left the firm in 2006, Google “is perhaps only a year or two away from a total upheaval. AI will wipe out the search engine results pages, where he makes the most money.Even if he catches up with AI, he can’t fully deploy it without destroying the most valuable part of his activity”. Buchheit goes so far as to compare the future of Google with that of “directories” and other portals (like Yahoo! or Lycos) of the years 1990-2000: “It was also a phenomenal economic model, until Google arrived AI will now do the same for search engines,” he tweets.
ChatGPT is wrong (too) often
Difficult, however, to adhere to such a discourse: what “delay” in AI is therefore talking about Paul Buchheit? As we said above, Alphabet has been working for several years on its own text-generating AI, LaMDA, and the result is just as impressive as that of GPT3 and ChatGPT. The language model on which Google is working, intended to integrate its own chatbots, is increasingly advanced and aims to allow users to ask questions and receive the most reliable answers possible.
If ChatGPT were one day “connected” to the Internet (and in its case, to Bing, from Microsoft), it’s a safe bet that LaMDA, and therefore Google’s search engine, would also be connected to the Web. On the contrary, it seems that Google is quietly preparing to pass another stage of its development, and that the buzz around GPT3 and the OpenAI chatbot is more of a “jolt” intended to remind people of its existence.
A Google Assistant that would really understand what is asked of it and that would be linked to the Alphabet search engine: admit that you have chills thinking about it. It should also be noted that ChatGPT does not thrill everyone either: several researchers, including Nicholas Weaver, researcher in computer security and networks at the University of Berkeley, point out that he is often wrong in his answers, in addition to not source nothing. OpenAI itself warns: its AI has been trained to recognize patterns in large amounts of text harvested from the internet in order to provide “useful information”, but some answers “may sometimes seem plausible”, while being “incorrect or misleading”.
Google should be launching soon
In his “Big Technology” podcast, American journalist Alex Kantrowitz explains that LaMDA is quite capable of doing better than ChatGPT, but that Alphabet does not release it and continues to offer a less efficient Google Assistant, for business reasons. model and reputation. Indeed, if ChatGPT is wrong, so is LaMDA. Google claims that its research “is still in its infancy”, and that sometimes the answers are “absurd”.
However, the Mountain View firm, whose business is based on the sale of sponsored links, cannot afford to offer false answers; even temporarily. In addition, its business model is also not to give a single answer, next to which it is not possible to place multiple links. A niche that it could however be interesting to dig into, both for OpenAI, which plans to monetize ChatGPT in the long term, and for Alphabet with Google Assistant.
“Even if chatbots became more accurate, Google would still have a business model problem to solve. The company makes money when people click on ads next to search results, and it’s hard to integrate ads in conversational responses Imagine you receive a response and are immediately offered to go somewhere else – this seems oppressive and unnecessary, so Google has little incentive to take us beyond traditional search , at least not in a way that leads us to a paradigm shift, until he figures out how to make the financial aspect work,” observes Alex Kantrowitz.
However, he believes that Google, whose strike force is colossal, “will not remain on the sidelines for too long”, and will eventually offer a new generation Google Assistant, as well as an AI-powered Google Search. Especially that OpenAI would already prepare the GPT4 engineand therefore an even more powerful system, which could lead ChatGPT to deliver fewer errors.
It now remains to follow the news of Alphabet in the matter, which in 2022 drew more than 150 billion dollars from the activities of its search engine; more than half of its income.