Ilya Sutskever launches SSI to build safe superintelligence

On June 19, 2024, Ilya Sutskever, Daniel Gross and Daniel Levy introduced Safe Superintelligence Inc. as a lab with one goal and one product. The SSI manifesto promises to advance capability and safety together but publishes no method, metric or experimental evidence: the announcement defines a mission; it does not show that the mission has been solved.

The announcement comes just over a month after Sutskever left OpenAI, the company he co-founded and where he served as chief scientist. His new venture puts the safety of highly advanced AI systems at the heart of the business—not as a complementary team or an added condition for a commercial product.

A company with a single product

SSI defines its mission unusually narrowly. The company says it has one goal and one product: developing safe superintelligence.

In this context, superintelligence refers to an artificial intelligence that would surpass human capabilities in virtually every relevant cognitive task. It is a more ambitious goal than artificial general intelligence, or AGI: a machine capable of performing a wide range of intellectual work at a human level. Neither exists today, but both serve as reference points guiding much of the research and investment at major labs.

SSI’s promise is notable because, at least in its initial conception, it rules out launching assistants, enterprise models or consumer tools while working toward that goal. Sutskever, Gross and Levy argue that this approach will allow them to avoid the distractions of product cycles and shield safety, research and technical progress from short-term commercial pressures.

This is more than a statement of principles. Building frontier systems requires enormous computing resources, scarce specialists and sustained funding. A company with no plans for interim products will have to convince investors that long-term research merits capital even without offering immediate revenue.

Sutskever’s return after OpenAI

Sutskever is one of the most influential figures in modern AI. Before co-founding OpenAI in 2015, he worked with Geoffrey Hinton at the University of Toronto and contributed to decisive advances in deep learning, the technique that enables models to identify patterns in vast amounts of data.

At OpenAI, he was a prominent voice in debates over how to deploy increasingly capable models. He also served on the board that temporarily ousted Sam Altman in November 2023, a crisis that ended with Altman’s return and Sutskever’s subsequent departure from the board. The researcher announced on May 14 that he was leaving the company and said at the time that he was confident OpenAI would develop beneficial AGI.

His partners bring different backgrounds. Daniel Gross led artificial intelligence and machine learning initiatives at Apple and co-founded Cue, a company later acquired by the Cupertino-based firm. Daniel Levy worked at OpenAI and contributed to the organization’s product and model development.

SSI will have teams in Palo Alto, California, and Tel Aviv, Israel. The two locations connect the company to two of the sector’s main technology talent hubs.

Safety shifts from department to raison d’être

The announcement reflects a tension running through the entire industry: labs are competing to train more powerful models while trying to show they can control them before bringing them to market.

AI safety covers a range of specific challenges. One is alignment—ensuring that a system pursues human goals and follows human instructions even in unfamiliar situations. Another is the ability to supervise models that could reason or carry out tasks more complex than those their evaluators can handle. Protection against malicious use, control over access to advanced models and the ability to shut down a system if it behaves unexpectedly also matter.

So far, the leading companies have combined those research efforts with rapid product launches. OpenAI sells ChatGPT and access to its models; Google, Anthropic, Meta and Microsoft are also competing for users, developers and enterprise contracts. SSI is proposing an alternative: a lab that does not have to choose between delaying a capability for safety reasons and integrating it into the next version of a product.

The proposal, however, leaves the hardest questions unanswered. The company still has to show which technical methods it will use to verify that a superintelligence is safe and who will be able to audit those assurances. Declaring safety a goal does not, by itself, solve the problem of measuring it—especially when the capabilities in question do not yet exist.

SSI is therefore entering a field where its credibility will depend less on its manifesto than on its future research results. Its first test will be proving that it can turn concern about safety into a genuine technical advantage before the race for more capable models imposes its own deadlines.

“Safe” needs a failure criterion

Safety can refer to very different behaviours: resistance to harmful instructions, tool control, protection of weights, robustness to errors or oversight of new capabilities. Without a defined threat, nobody knows what is being prevented; without a test and threshold, nobody knows whether the system improves. A broad mission must become hypotheses that can fail.

The useful order is asset, threat, control and evidence. Identify what must be protected and from whom. Then design an intervention and adversarial evaluation. The report should include cases the system failed, test conditions and who was allowed to review it. Saying safety stays “ahead” of capability requires a common scale—or at least explicit reasons not to train or deploy.

The business model is part of the risk

Forgoing intermediate products may reduce pressure from a commercial release calendar, but it increases dependence on patient funding. Investors, voting rights, access to compute and the power to replace management affect whether a pause can be sustained. A promise of economic insulation needs governance mechanisms, not merely intention.

Focusing on one goal can produce a coherent agenda while reducing opportunities to learn from real use. Laboratory tests cannot fully reproduce users, incentives or attacks. A responsible programme therefore needs outside evaluators, reporting channels and publication criteria that permit scrutiny without releasing dangerous material.

The transferable skill is to read a safety manifesto as the beginning of a specification. Underline the threat, metric, threshold, stop authority and accessible evidence. Missing elements do not prove bad faith; they identify exactly what the laboratory must demonstrate before asking for trust.

Publishing evidence without publishing dangerous capability

Transparency does not require releasing weights, attack code or details that enable abuse. A laboratory can describe the threat, evaluation design, reviewers, aggregate results and corrected failures; it can also give controlled access to auditors. The question is not “does it publish everything?” but “does it provide enough evidence for an independent party to test the claim?”

A public record can be updated without revealing secrets: objective, threat, owner, test, threshold, date and status. If the method changes, preserve the previous version. That memory prevents “safety” becoming a moving target that always appears satisfied because its definition follows the result.

Evidence turns intention into a testable obligation.