Investment from Nvidia and estimated launch date: new details on Ilya Sutskever's venture
About three years after leaving OpenAI, Ilya Sutskever is nearing the launch of his startup SSI's product. The company, which has already raised $7 billion, is designed to solve the memory limitation of existing models and allow them to adapt personally to the user in real time.
Almost three years ago, Ilya Sutskever, OpenAI's chief scientist and the mastermind behind ChatGPT, clashed with the company's founder Sam Altman and nearly had him ousted. Behind the argument lay an ideology: Sutskever, who saw OpenAI as a non-profit organization, sought to limit language models to ensure safety and prevent rapid development that could endanger the public. Altman, who supported turning the organization into a for-profit company, held a different view.
In the end, Altman had the upper hand: investors returned the founder to the CEO's chair, approved the transformation into a commercial company, and Sutskever left to start a new venture that he claims will revolutionize the world of artificial intelligence: a language model with a conscience and a safety margin, aware of risks to human users and capable of learning from them.
SSI (Safe Superintelligence)
-
Field of activity: artificial intelligence, a language model that learns while operating.
-
History: founded in 2024 by Ilya Sutskever, Daniel Levy, and Daniel Gross (who left for Meta); raised $7 billion, with a valuation of $32 billion in the last round.
-
Employees: over 50, in San Francisco and Tel Aviv, including several Israelis.
-
Investors: Nvidia, Alphabet, and funds such as Sequoia, SV, and OurCrowd.
Safety, reliability, and alignment with human values
Over the past two years, Sutskever—a former Israeli who completed his bachelor's degree at the Open University—has raised $7 billion for Safe Superintelligence (SSI). The company will offer a language model to compete with giants like Claude from Anthropic or GPT from OpenAI. Sutskever secured an investment of about $5 billion from Nvidia, likely involving a deal that returns capital through the purchase of graphics processors.
Sutskever claims that while other companies are busy increasing computing power and adding more parameters, SSI's model will focus on safety, reliability, and alignment with human values. While the company's activity has remained a mystery, investors—including Nvidia, Alphabet, Sequoia, Atreides, Greenoaks, Yuri Milner, and Israeli firms OurCrowd and VentureIsrael—are reportedly pressuring Sutskever to show progress toward a product launch.
One investor, Atreides fund founder Gavin Baker, revealed that SSI intends to launch its first language model by the end of the current month. Some AI experts even claim the model has already been launched secretly and is currently in the testing phase.
Dr. Khademi Nouri, a senior researcher at the University of Toronto's AI lab, where Sutskever completed his doctorate, noted on X that he had received access to the model: «It seems they have solved the problem once and for all—the model has unlimited context—it knows about me and everything I have written in the last 15 years. How did they do it?».
Solving limitations in other models
Experts estimate that SSI's breakthrough involves more than just building a safer AI model. It appears they have solved the «knowledge freeze» limitation found in the smartest models from Anthropic or OpenAI. In existing models, knowledge is fixed once pre-training is complete; when new documents are uploaded, the model does not learn but holds information in a disconnected text window that vanishes when the conversation ends.
SSI's solution allows for updating model weights in real time while reading a document. According to experts, this should also result in a safer model: while previous attempts to teach models «on the fly» led to uncontrolled information loss or safety failures, SSI appears to have solved this. If successful, this would be a revolution, potentially reducing the number of processors AI giants need to purchase as models become more efficient.
«Constant learning may be an obstacle»
Uri Eliabayev, an AI consultant, believes SSI will create a new niche rather than immediately replacing advanced models from Anthropic or OpenAI: «SSI's model will focus on using lean models that do not require massive processing power, allowing them to improve over time. The fact that it adapts personally to users could allow it to reach the level of the most powerful models». However, Eliabayev notes there is no guarantee the model will be safer: «Its constant learning may be an obstacle and change its character in an unpredictable way», he says.
