OpenAI, the company behind the GPT-3 language model and ChatGPT chatbot, is preparing GPT-4. The future artificial intelligence tool is already arousing many fantasies.
ChatGPT has been available for public testing for more than two and a half months. Since November, much has been said about this conversational agent capable of manipulating natural language with ease. As a result, even if it is sometimes found wanting, the tool is able to return structured and articulated responses to the requests made to it.
To work, ChatGPT was trained from very large textual resources. This preparation was carried out using methods found in artificial intelligence and more specifically in machine learning. The bottom line, essentially, is that ChatGPT is based on the GPT-3.5 language model.
Billions and billions of parameters for GPT-4?
However, GPT-4 will arrive soon. The American company OpenAI, which is behind ChatGPT, iterates regularly to provide an increasingly powerful tool. GPT-2, released in 2019, for example, had 1.5 billion parameters. It was followed in 2020 by GPT-3, which has 175 billion – an artificial neural network 116 times more developed.
On GPT-4, OpenAI has remained very discreet so far – leaving Internet users to their fantasies. So, as of September 2021, GPT-4 is rumored to have one hundred trillion (100,000,000,000,000) parameters. That is a network 571 times more developed than GPT-3. This leap was mentioned in a Wired article in August 2021, on a subject not directly concerning OpenAI.
Since then, this number has been coming back, whether on the OpenAI Discussion Forumin blog posts Where in enthusiastic tweets. It turns out that Sam Altman, the founder of OpenAI, denied the rumor of a GPT-4 with as many parameters. On the contrary, GPT-4 would actually be equivalent in size to GPT-3. Progress would rather be on the side of data processing.
Video, image, sound…
Since then, the hall noises around GPT-4 suggest that this language model would be multimodal, i.e. it could not only accept text, but also image, audio and the video. Today, ChatGPT (which relies on GPT-3.5) only returns text. But today there are tools capable of generating images. OpenAI also provides DALL-E.
GPT-4 seems obviously operational since 2022, but access is restricted to a small number of Internet users. ” A friend has access to GPT-4 and cannot talk about it due to confidentiality agreements.[…] Her emotion alone told me it was next level », according to Robert Scoblea well-known American blogger, in August 2022.
That same Robert Scoble then wrote in November : ” Disruption is coming. GPT-4 is better than expected. And it’s one of many such AIs that will be delivered next year “. Difficult to do better in terms of teasing. In any case, it seems that the arrival of GPT-4 would be good for 2023 — OpenAI does not say anything about it on its website.
What does Sam Altman say, anyway? The person concerned tends to blow hot and cold. On Twitter, he indulged in an enigmatic publication, hijacking a quote from Darth Vader in the very first film of Star Wars. The opportunity to talk quickly about the Turing test, without saying more.
The original quote from the film reads: Don’t be too proud of this technological terror you’ve built. The ability to destroy a planet is insignificant next to the power of the Force… But the second part has been adapted to: “ The ability to pass the Turing test is insignificant next to the power of the Force. »
A way to say that the interest of GPT-4 will not be in the validation of this test? The mystery remains: we wanted to do the Turing test with ChatGPT, but it didn’t go as planned. The concept ? A person exchanges blindly with a machine and human, without knowing who is who. If he fails to locate the machine, the test is validated.
Since then, Sam Altman even seems to be seeking to mitigate any overly high expectations around GPT-4. In an exchange with TechCrunch, he initially declined to say when exactly this language model will be released. He ” will come out at some point, when we are sure we can publish it in a safe and responsible way “. That is, without harmful diversion.
As for those who already dream that GPT-4 is a general AI, that is to say that can approach human cognitive abilities, Sam Altman tempers. ” We don’t have AI of this type, he warned. Here, GPT-4 goes “ disappoint “. It remains to be seen where the exceptional disruption evoked by Robert Scoble and on which Sam Altman jokes with his meme will nestle.
Help us build the future of Numerama by answering this survey!
