When a Nobel Prize winner in economics sounds the alarm, it is prudent to listen. What concerns Daron Acemoglu, recent winner of the prestigious title, can be summed up in two letters: AI. Not that he takes a dim view of the development of this technology. But the crazy enthusiasm it arouses among investors worries him. Only a small portion – 5% – of jobs can be largely automated and therefore replaceable by AI, he said in October in the columns of Bloomberg. Not enough to justify in his eyes the colossal investments that are made in it.
In the second quarter of the year alone, Microsoft, Alphabet, Amazon and Meta made 50 billion in investment spending, largely due to generative AI, as it requires expensive computing capacity. . So much so that barely two years after receiving 10 billion from Microsoft, OpenAI, which will record losses of 5 billion this year, had to carry out a new large-scale fundraising ($6.6 billion).
Daron Acemoglu is not the only one to fear that an AI bubble is forming. In a noted post, David Cahn, a partner at Sequoia Capitalwarns against “the illusion that the IAG [NDLR : intelligence artificielle générale] will arrive tomorrow and that in the meantime, we should all hoard the only resource that matters: graphics processors.
The statements of AI professionals, it is true, blithely maintain confusion about what generative AI can – or will soon be able to – do. “We have been bombarded with talk about transhumanism and singularity which has nothing rational about it. The problem cannot be formulated like that, in my opinion,” confides Jérôme Monceaux, co-creator of the famous robots Nao and Pepper, now founder and CEO of Enchanted Tools between two conferences at the Transatlantic Leaders Forum which was held on October 18 in New York.
The FTC at war against “AI washing”
The enthusiasm around AI is such that it encourages opportunists to claim it even when they are not doing it. So much so that the Federal Trade Commission, the American competition authority, ended up attacking several companies engaging in shameless “AI washing”. The big names in generative AI themselves play with words, extolling the merits of artificial intelligences that reason and are supposedly on the verge of surpassing humans. “The problem is that their training corpus is now so vast that we don’t know whether these AIs are able to generalize latent skills from this data or whether they are repeating, like parrots, things they have seen passing by,” confides an AI researcher.
The Achilles heel of these new generative AI? Reliability. A problem linked to their intrinsic nature. These AIs operate probabilistically and not deterministically. This is what gives them tremendous creative potential. But rather reserve them for uses where the Internet user does not need a 100% reliable response or for those where verification can be carried out quickly. The Internet user should not pass ultimately as much time controlling his AI as the task would have taken him without this technology.
In this uncertain climate, “the uses of AI for which customers will be willing to pay are not yet clear. The general public is unlikely to purchase AI. It is at the corporate account level that everything will play out,” says Alexis Deladerrière, global equity portfolio manager and head of developed markets at Goldman Sachs Assets Management.
For this expert, the next few quarters will be decisive. “The AI market is entering a second phase, quite different from the first, that of the development of concrete applications,” he points out. AI companies will have to show tangible elements, monetizable uses, and increasing revenues. “Because the market can quickly turn around, warns the expert met during the Transatlantic Leaders Forum. If necessary, financing will slow down. Investors will wait until clear uses are established with current models, before investing more in the development of new, even more sophisticated models.”
The AI market enters its phase 2
The needs for semiconductors will also change according to him. Because AIs do not need as powerful chips to respond to user queries – so-called inference – as during the training phase. Less expensive but specialized chips will enter the scene.
“The companies best positioned to benefit from this second phase of the market are those that already have a large customer base and a lot of data. They will be able to develop applications for their customers. Here again, small players who want to establish themselves will face facing significant challenges during this second phase”, underlines Alexis Deladerrière.
While some of the hopes placed in generative AI are excessive, we should not throw the baby out with the bathwater. “The dreams of the entrepreneurs of the 2000s did not materialize as quickly as they hoped. But most of the developments they predicted have largely come to fruition in the decades that followed,” underlines the head of developed markets at Goldman Sachs Assets Management.
And the disruptions that generative AI introduces are very real. An example given by Jérôme Monceaux, from Enchanted Tools, is enough to demonstrate this. In the past, this robotics and AI specialist has long tried with his teams to equip machines with the ability to detect a person lying on the ground so that in the event of an accident, they can raise the alarm. Without success. “With generative AI, it worked right away without us even having to code precisely for this purpose. The way AI is changing human-machine interactions is dizzying. For developers like me, it’s disconcerting We are so used to machines doing precisely what we want, now they surprise us every day,” he confides with a laugh.
If o1, the last major language model from OpenAI, still makes errors, it manages to significantly reduce their frequency by breaking down Internet users’ requests into multiple simple steps and allowing them to follow its “path”. of thought.”
“A real fundamental movement”
The fact that only 5% of jobs are largely automatable is not necessarily a bad thing. The calculations of economist Philippe Aghion lead him to the same estimate as Daron Acemoglu. But the latter draws very different conclusions. As he confided to L’Express, this removes the specter of mass unemployment without thwarting potential productivity gains. “Jobs will be enhanced, because only part of the tasks that make them up – the most boring – will be automated, which will allow employees to be more creative and efficient on other tasks,” he explained in our columns. By helping us produce goods and services more efficiently but also to design new ideas, AI could, according to him, boost France’s GDP from 250 to 420 billion in ten years.
“Generative AI has breathed new life into Silicon Valley, it’s a real fundamental movement that is only just beginning,” observes Benoît Buridant, CEO and co-founder of Frenchfounders, a network of international French-speaking entrepreneurs who organized the Transatlantic Leaders Forum 2024 in partnership with Sequoia and Goldman Sachs. The Nobel Prize selection committee made no mistake. If Daron Acemoglu won the Nobel in economics, big names in AI – Demis Hassabis, John Jumper, Geoffrey Hinton and John Hopfield – won those in chemistry and physics.
.