The day when AI will be as intelligent as humans is coming, according to Sam Altman. And even faster than expected. While critical voices are beginning to anticipate a stagnation in AI, OpenAI, which celebrated the opening of its Paris office on the evening of November 14, is more confident than ever. “Progress in AI continues to accelerate,” assures Romain Huet, head of developer experience. Thanks to our latest o1 model, a new era will begin.” In less than a year, OpenAI has transformed itself, going from a headquarters in San Francisco to a network of nine offices around the world. “ChatGPT has 250 million active weekly users,” says OpenAI. And in France, the group already counts major players among its partners and clients, including Sanofi, Publicis Groupe, ESCP and even the unicorns. Pigment and Mirakl.
Sam Altman is not in the habit of listening to pessimists anyway. In a recent podcast from Y Combinatorhe reveals the hostility that OpenAI aroused in its beginnings: “Our initial postulate was that deep learning worked and that the more we practiced it on a large scale, the better the results would be. But this bet was almost heretical at the time. At the time, senior industry figures thought it would be a huge waste of resources, that it would trigger an AI winter. The OpenAI team persisted. Rightly so. By “scaling” its models, it has made giant leaps in the AI sector.
Will this recipe continue to work? OpenAI identifies five levels of artificial intelligence. ChatGPT corresponds to the first, that of chatbots, these conversational tools. The second is that of AIs that reason, like its OpenAI o1 model launched last September. Level 3 corresponds to “agents” capable of carrying out actions autonomously (making a reservation, placing an order). The fourth to “Innovators”, models that help make new discoveries. At the top, level 5 of “Organizations” designates AIs capable of carrying out the work of an entire company.
And several signs lead Sam Altman to believe that climbing the pyramid is possible. To begin with, “the cost of AI has dropped drastically,” points out OpenAI during a press presentation on November 14. AI players often charge company offers for “million tokens”, these elementary units of text or code into which user requests are broken down. And the bill has indeed reduced: from $36 in March 2023 to $4 last August.
Sam Altman’s latest AI “reasons”
But it is above all the perspectives opened up by the o1 model that make OpenAI confident. Romain Huet amazed the public last night, by showing how this latest generation AI could help people with no computer skills. Asking o1 for help on the microphone, he explains to the AI that he does not know how to fly the drone model he has, nor trigger his camera. A few moments later, the magic happens: o1 writes him from A to Z a program allowing him to take off the machine and direct it. And film the audience.
“o1 has helped certain start-ups to properly formulate patent applications or analyze financial data, others have tested it to build complex physical objects,” specifies OpenAI. A start-up which experimented with it during a hackathon organized by the group thus reduced the time to design a sail – an object which requires complex aerodynamic calculations – from three hours to a few minutes.
The secret of o1’s progress? This new AI takes the time to “think”. More concretely, to break down a complex request into series of simple steps in order to better understand it. “o1 also explores several response options and carries out a sort of critical examination of these options in order to choose the best one,” the OpenAI team explains. This whole process is called “chain of thought” (chain of thoughts, in English).
A way of doing things that contrasts with previous generations who instantly deliver the answer that their sophisticated probabilistic calculations have deemed relevant. This is why since o1, OpenAI has divided its range of AI into two families: “classic AI” and “reasoning AI”. The first, designed for conversations with AI and urgent requests. The others, for complex problems requiring patience.
“What price for a new cancer treatment?”
“At the moment, o1 thinks for several seconds. But we are considering future versions where it would think for hours, days or even weeks. The inference costs would be much higher. But what price would we be willing to pay for a new treatment against cancer? For revolutionary batteries? For proof of the Riemann hypothesis (Editor’s note one of the unsolved mathematical problems considered to be the most important of our time). chatbot”, underlined on X Noam Brownone of OpenAI’s leading researchers.
Sam Altman says it and repeats it on social networks : “there is no wall (of AI)”. And unlike many leaders in the sector, he clearly does not consider it necessary to invent fundamentally new approaches to achieve the holy grail of artificial general intelligence (AGI). “For the first time ever, I feel like we know what to do. Getting from where we are to AGI will take a huge amount of work. There are always some unknowns […] but overall we now know what to do […] And when you have this degree of visibility, it can go quite quickly,” he recently explained to Garry Tan, the CEO of the Y Combinator incubator.
Sam Altman is not the only one to believe it. In the last Lex Friedman podcastDario Amodei, CEO of Anthropic, also anticipates a rapid arrival of the AGI. “Many obstacles could delay its materialization […] But if you look at the rate at which AI capabilities are increasing, it suggests that we could achieve it by 2026 or 2027.”
What will super-powerful AI bring to our civilization? For the CEO of Anthropic, for sure, major progress in scientific research. “Humans have a hard time understanding the complexity of biology. At Stanford, Harvard, Berkeley, there are entire departments of people studying the immune system or metabolic pathways. Each of them specializes and only masters a fragment of the whole and they struggle to combine their knowledge. I have a feeling that AI would be more intelligent,” he explains to Lex Fridman.
The Elon Musk threat
In his essay The Age of Intelligencethe CEO of OpenAI, Sam Altman, also discusses the arrival of virtual educational tutors, caring medical assistants, and solutions to the climate crisis. “In the decades to come, we will be able to do things that would have seemed like magic to our grandparents,” he writes.
It remains to be seen if the political sphere does not put a spoke in Sam Altman’s wheels. Certainly, Donald Trump, who has just been re-elected president, seems inclined to strongly deregulate AI. He notably floated the idea of removing Biden’s Executive Order which nevertheless establishes a flexible framework for the sector. But Donald Trump also entrusted leading responsibilities to Elon Musk who actively campaigned in his favor. And the latter has a stubborn grudge against OpenAI, which he helped to co-found but left in 2018. He has since continued to criticize the lucrative turn taken by the entity which he judges to be a betrayal “of Shakespearean proportions” with its initial mission. And filed a complaint against them. Elon Musk is also determined to take advantage of the AI boom and has launched his own start-up in the field, xAI. Sam Altman’s biggest challenge on the road to AGI may not be taming the machine. But to agree with a man.
.