For the first time, in the Censis Report, there is data regarding artificial intelligence. They are found in the “Communication and Media” section and describe the AI ​​situation well two years after the launch of ChatGPT.

“Today – we read in the press release announcing the report – image generator software is used by 8.4% of Italians, text generators (ChatGPT and similar) by 8.2%. But for 65.5% of Italians the effects on employment will be disastrous due to the replacement of human beings with computers and chatbots. 37.4%, however, think that thanks to artificial intelligence we will be able to free ourselves from repetitive and boring jobs, thus encouraging the exercise of creative activities”.

Translated: artificial intelligence has been and continues to be talked about a lot, imagining extremely close futures and perspectives. To date, however, the announced revolution has not yet arrived; or, in any case, it was adopted by a fairly minority share of the population.

Artificial intelligence is slowing down

It is, indeed, a time of slowdown in the world of artificial intelligence. Which, ever since that now famous November 30, 2022 launch of ChatGPT, had accustomed us to continuous developments, to very fast changes, to an almost frightening evolution. And instead, in the last year, something seems to have stopped: the level of chatbots has stabilized and no particularly surprising news has emerged. “Only” the costs and energy consumption increased. OpenAI, the leading company of this new wave, will lose 5 billion dollars in 2024; Google, another of the protagonists in the development of AI systems, admitted in a report that its emissions have grown by almost 50% in the last 5 years.

As an article appearing on Vox suggests, at the heart of the slowdown there could be the questioning of a strategy that has always been considered the basis of generative artificial intelligence. The adage, in English, goes: “Scaling is all you need”. In other words, all you need is to scale, to grow. To understand this, you need to quickly remember how generative AI works. Which is a system trained on an enormous amount of data – lots of texts, in the case of ChatGPT – and which on the basis of that material learns to find relationships between words and generate text.

Until now, the idea was that as that training material increased, the system’s capabilities automatically improved as well. The news, the feeling circulating in the AI ​​world is that this may not be the case. It is not enough to scale, to use a somewhat literal translation of the English term: new technologies, new ways of thinking about artificial intelligence itself, could be needed.

OpenAI’s answer: artificial intelligence that reasons

In the Vox piece by Kelsey Piper, the journalist highlights precisely this point: a possible paradigm shift must not lead us to underestimate the extent of the impact of generative artificial intelligence. Paradigm shift that OpenAI seems to be clear about. In recent days, the company announced the arrival of the full version of o1, a new artificial intelligence model quite different from those we have had to deal with in the last 2 years. For the occasion, Sam Altman’s company has launched a Pro subscription plan, from 200 euros per month, compared to 20 for the Plus one, which allows you to make the best use of this new system. While previous models were designed to be excellent at predicting words based on massive amounts of textual data (an approach known as word prediction), o1 is notable for its focus on another goal: reasoning.

A complex yet fascinating concept. Mark Chen, one of OpenAI’s scientists, explained in an interview with The Atlantic that GPT models were trained to mimic what humans write by looking for patterns in textual data. But this approach has a flaw: Humans are not word prediction machines. o1, on the other hand, was designed to produce responses that simulate independent thinking. He doesn’t just imitate; explores, evaluates and chooses the best path to solve a problem.

Although the technical details of o1 are secret, some clues emerge from statements and research. OpenAI appears to have used a method similar to that of gaming algorithms: subjecting the model to thousands of problems and providing continuous feedback on its solutions. This approach is reminiscent of an artificial intelligence that plays millions of games of chess to perfect its strategy or a rat that, after exploring thousands of mazes, learns to navigate them efficiently.

A concrete example of the differences between the two approaches was given by the expert Vincenzo Cosenza, in a video published on his YouTube channel. In the test, Cosenza asks both models, GPT 4o (the default one on ChatGPT today) and o1, to find the solution to a game that is very reminiscent of The Guillotine from L’Eredità. The response of the classic model is disappointing; that of o1, with a long reasoning and a series of tests on the most suitable answers, seems to be the correct one.

https://www.youtube.com/watch?v=ID2Fhf2QwPMt=1275

Despite the progress, o1 is not without its flaws. As noted by several researchers, the model works best when addressing problems for which there is a lot of training data or testable solutions, such as programming or solving mathematical equations. François Chollet, an artificial intelligence expert, pointed out that despite the ability to “reflect” on its answers, o1 remains bound to what it already knows. In other words, he does not invent new ideas, but optimizes the use of the information at his disposal.