It’s on everyone’s lips, yet its six months of life seem enough to make ChatGpt seem already obsolete. The issue was raised by the American magazine The Atlantic. In the service, signed by Matteo Wong, it is underlined that both the Bard chatbot, just presented by Google, and OpenAI’s Gpt-4They will soon be able to use artificial intelligence not only to produce text but also to describe images. It would be a step forward towards what Wong defines “a silent revolution” how companies, researchers and consumers «develop and use artificial intelligence to gain a richer and more complete understanding of the world». The protagonists of this revolution will be the so-called «multimodal models».
The most important models
To understand what it is, The Atlantic remember that ChatGpt (of which the iPhone app is already ready) and similar programs they are based on highly advanced language models and have been trained to produce accurate written texts quickly. Their fate, however, is of give way to machines capable of processing images, audio and even obtaining sensory data from robots. Multimodal models, in fact. It would be a matter of replicating the way children learn as they grow up and observe the world. This would lead to an Ai capable of performing multiple functions and therefore being adopted on multiple occasions and included in multiple products. In addition to Gpt-4 and Bard, the program is also part of the multimodal models ImageBind Meta (processes text, images, audio, motion and location information), Google’s PaLM-E (already trained as a language model but also to obtain sensory data from robots) e Dall-E 2 by Microsoft, text-to-image generator.
La spinta del deep learning
Many researchers hope that these multimodal models can taking AI to new heights, like doing internet research without making things up, animating a video, driving a robot, or building a website yourself. “When they talk about a traffic jam, current speech patterns have no experience of traffic jams,” he explained Melanie Mitchell, an artificial intelligence researcher and cognitive scientist at the Santa Fe Institute “But if an AI’s training data can include traffic jam videos, they can glean a lot more information.”
The push for multimodal models isn’t entirely new, but while fields of computing such as natural language processing, computer vision, and robotics have used wildly different methods for decades, now they are all using deep learning. This allows their products to complement each other. While Big Techs can handle ever-increasing amounts of data. It is no coincidence that it was the founder of OpenAI, Sam Altman, in his hearing in the US Senate who stated that the era of text-based templates is probably over.
Dangers and rich gains
The debate is open, also because – as the American magazine points out – a multimodal AI could unfortunately also be more susceptible to certain types of manipulations, such as altering key pixels in an image (think of the effect of fake images on Donald Trump’s arrest). But the economic aspect should not be underestimated. And Wong’s conclusion is illuminating in this regard: «Language models are already a gold rush for Silicon Valley: prior to the multimodality corporate boom, OpenAI projected $1 billion in revenue by 2024. Several recent analyzes have calculated that, in a few years, ChatGpt will add tens of billions of dollars to Microsoft’s annual revenue».