About three years ago Ilya Sutskever, then OpenAI chief scientist and the brains behind ChatGPT, fell out with the company’s founder Sam Altman, and almost ousted him. Behind the dispute was ideology. Sutskever, who saw OpenAI as a non-profit organization, sought to limit the language models in order to make them as safe as possible, and to prevent too rapid development that would endanger the public. Altman, who supported turning the enterprise into a profit-oriented company, thought otherwise.
Ultimately, Altman had the upper hand and the investors reinstated him as CEO's, and approved the process of turning OpenAI into a company. Sutskever left to found a new startup that he claims will revolutionize the world of AI: a language model with a conscience and a margin of safety, one that will be aware of the dangers to human users, learn from them, and behave cautiously towards them in the future.
Safety, reliability, and humanity
Over the past two years, Sutskever, an Israeli who graduated from the Open University - has raised $7 billion for his company Safe Superintelligence (SSI), which will offer a language model that rivals giants like Anthropic’s Claude Fable and OpenAI’s GPT. He has hired over 50 employees for offices in Palo Alto and Tel Aviv and has brought in Nvidia to invest about $5 billion in the company, likely an investment that will be repaid through the purchase of graphics processors.
Sutskever claims that while other companies are busy increasing the computing power of their models and adding more and more AI processing units (parameters), SSI’s model will focus on safety, reliability, and humanity. But so far, the company's operations have remained a mystery. Moreover, it is estimated that some of the investors behind it - giant investors like Nvidia, Alphabet, Sequoia, Atreides, Greenoaks and Yuri Milner, as well as Israeli investors such as OurCrowd and VentureIsrael, are pressuring Sutskever to show some progress towards launching a product.
Sutskever himself remains silent, but one of the investors in the company, Atreides Fund founder Gavin Baker, revealed that SSI intends to launch its first language model by early next week. Some experts in the field of AI even claim that the model has already been launched secretly and that it is in the testing stages of those close to Sutskever.
A senior researcher at the Artificial Intelligence Laboratory at the University of Toronto, where Sutzkever completed his doctorate, admitted in a tweet on the X network that he had gained access to the model from an anonymous company that keeps a low profile, saying that "It seems they have solved the problem once and for all - the model has unlimited contact - it knows everything about me and everything I have written in the last 15 years. How did they do it?"
Overcoming limitations of other models
Despite speculation, AI experts believe that SSI’s breakthrough involves more than building a safer AI model - on the contrary, it is not at all certain that the company has succeeded in this task. Instead, however, it seems to have solved several limitations that exist even in the smartest models of Anthropic or OpenAI - the most important of which is the "knowledge stagnation" of language models - that is, the knowledge that existing models demonstrate is fixed as soon as their pre-training is complete, and each time they are fed a document or given new instructions, they do not learn the new information but rather hold it in a kind of disconnected text window that disappears when the conversation ends.
To solve complex problems, the models think longer but the structure of their "brain" - the model’s weights, which determine how it will respond to a question - does not change. SSI’s apparent solution, on the other hand, allows the models’ weights to be updated in real time while reading a document.
According to experts, the model should also be safer as a result. If previous attempts to train the models on the move resulted in uncontrolled loss of information or loss of safety restraint, then, according to estimates, SSI has managed to solve this problem. If SSI succeeds in its mission, it is a revolution that also has a possible impact on the number of processors that the AI giants will be required to purchase. Language models will become more efficient and the advantage of scale will be reduced.
<b."Continuous learning may be an obstacle"
AI consultant Uri Eliabayev does not believe that SSI will promote its model as a replacement for the advanced models of Anthropic or OpenAI in the first stage, but will create a new niche.
He says, "SSI's model will not necessarily be better and will not necessarily be able to serve as a replacement - it will focus on examining the use of lean models that do not require strong processing power and will allow them to improve over time. However, the fact that it adapts itself personally to users could allow it to reach the level of the most powerful models. This is because it will learn each of us dynamically.."
However, Eliabayev points out, there is no guarantee or proof that Sutzkever's model will be more secure: "Its constant learning may be an obstacle and change its nature in an unexpected way," he says.
Published by Globes, Israel business news - en.globes.co.il - on August 26, 2026.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2026.