SSI Ilya Sutskever: the first model who completes her studies directly at work

Briefly about the main thing (BLUF)

U SSI Ilya Sutskever's first model has leaked online - and she is completing her studies while working. We explain how the TTT approach differs from everything that exists now, why this can be serious, and what the main security issue has to do with it.

Two years ago, Ilya Sutskever, the man behind AlexNet, GPT and o1, left OpenAI and founded the company SSI (Safe Superintelligence). Since then, not a single model, not a single article, or a single public product has come out of it. All this time, the answer to the question “what does Ilya do” was silence.

Now it seems the silence is ending.

What was leaked online?

If this is true, we are not talking about another “bigger GPT”, but about a model that knows how to learn how to learn.

How is TTT different from everything that exists now?

Modern models work like this: first they are trained once, then they respond based on what is loaded into the context. In order for the model to “remember” something new, it must either be retrained again, or the information must be slipped into the context window - like a cheat sheet.

TTT changes the very principle. The model reads incoming information not as a cheat sheet, but as educational material: while working, it continues to predict the next token and “presses” what it read into its own weights. Context ceases to be a side note - it becomes part of the model itself.

This is not the same as agent-harness and context engineering that are now fashionable: they build a forest around the model from external tools and extended context, but the model itself does not change. TTT goes one step further - it is the model that changes.

Why could this be serious?

For now this is an anonymous leak and should be treated with caution. But it rests on several already known facts.

Firstly, Ilya himself has said more than once that the terms “AGI” and “pre-training” have taken the industry in the wrong direction. People are strong not because they know everything from birth, but because they learn throughout their lives. His ideal is superintelligence, like a very smart fifteen-year-old teenager who, after starting, continues to master programming, medicine, any job.

Secondly, in July 2026, Nvidia announced a long-term partnership with SSI: An investment of approximately $5 billion and next-generation computing power that will increase the company's resources by approximately tenfold. In the release, Nvidia specifically noted that SSI For the last two years I have been following a “new research line”.

The main issue is safety

Letter S in the title SSI means "safe". A model that changes itself after launch is both a huge opportunity and a huge risk: it can not only learn new things, but also learn the wrong things, or forget what it already knows. In 2016, the Tay chatbot from Microsoft picked up insults from users within a day and was disabled.

Anonymous states that the problem of “what can be taught and what cannot” SSI I've already decided. But these are still words, not proof.

What to expect

According to rumors, the current version is already ready, and the next one will be ten times larger. The first release could take place as early as this August and will initially open to a small circle of users. However, the opposite information immediately appeared: “this month may not happen.”

While there is no official confirmation, all this remains rumors. But if SSI will actually release a model that is completing its training directly at work, this will be a shift not in the size of the model, but in the very way of its existence. And such shifts are worth keeping an eye on.

Modern neural networks can already do a lot - from generating pictures And video to chat with AI. All this is available in one window on NeuralSpace — without VPN and with payment in rubles.

Frequently asked questions (FAQ)

Question: How to get started? Answer: Register on the platform and receive free tokens for testing.

Question: Are the generated materials suitable for commercial use? Answer: Yes, you receive full commercial rights to all content created.