It's the first time in the history of the world.
Is Ilya's first model finally coming?
Just this morning, the SSI reported another big news!
Ilya is building a small reasoning engine, the core of which is TTT (test training).
Ilya is about to take the first step of experimentation, going out of the age of research and back to the age of scale.
After Ilya left OpenAI, the entire AI world was waiting for an answer.Ilya Sutskever，He did.What are we going to do?
This is the man who created AlexNet, GPT-2, GPT-4, and RLHF, leading the wave of AI on several occasions.
But OpenAI is still racking up crazy calculations, and Ilya predicted the dead end behind smoking:
There are security flaws, efficiency cliffs, and unobtainable smart ceilings.
Large models of static weights, but giant flowers sprinkled in greenhouses.
Once they hit the real world's "extra-distributed data", these giant objects instantly turn into a legitimate bullshit machine.
Because their brains, shaped by the past, faced with the future, they're like,They can't breathe, they can't save themselves.。
So the SSI came.
They trained TTT from the concept in the paper, a hard-boiled blade capable of industrial landing. According to front-line sentinels, the SSI's secretly sharpened reasoning engine, with its three sharp knives, killed overhead:
The first knife: data efficiency, fear of failure to communicate.
Because it can adjust weight in real time while reasoning, it does not need to swallow the nonsense of all humanity at the pre-training stage at all.
It only learns one thing.How to LearnSo, a model with a much smaller parameter, in real business, might be able to keep up with the huge unstable front wrists of those who burned tens of thousands of H100 tanks.
The second knife, the 10-fold jump, is just getting started.
The first generation of small engines is ready, and the team is directly enlarging the next generation by 10 times.The explosion。
The third knife: Your digital employee will never be forgotten.
What's the craziest thing about the traditional Agent? Every morning, it flips over and over the RAG database like a person with dementia, responding slowly until you want to hit the keyboard.
The SSI's engine, which is able to safely submerge the "fast weights" produced during the test by some mechanism. So when you interact with it, it has a "physical synapse" in its brain that belongs exclusively to your team. It really remembers you, it doesn't need to restart every day, it does.It's your man.。
Ilya, this knife slashes the gates of life in ancient times.
Why Ilya? Why now?Last year, Ilya himself repeatedly publicly proclaimed that "the age of pre-training is over, and we are entering the age of research", emphasizing great mathematics.
Last month, SSI partnered with Nvidia to raise at least $5 billion in investments.
The deal, which gives SSI access to the latest generation of Nvidia's Vera Rubin platform, is enough to expand its computing capabilities 10 times over the next 12 months.
Nvidia stated: "We have been granted a rare access to the company's rigorously protected research.
The SSI said more frankly: "Our research is at a time when it is worthwhile to scale it up".
Let's take a quick look at what the strangest company in AI has come up with:
June 2024: Ilya founded SSI -- a goal, a future product, a pen directed to "security superintelligence".
September 2024: $1 billion in financing valued at $5 billion.
April 2025: A further 2 billion were added and valued at up to 32 billion.
June 2025: Meta tries to swallow SSI. Ilya retreats. Co-founder Daniel Gross nods his head and turns to Meta.
November 2025: In a podcast, Ilya once again proclaims "the end of scaling" and says that we are "back to the age of research, but this time with a big computer".
On July 27, 2026, he quietly shouted: "It's time to scale up SSI too".
The "big computer" in Ilya's mouth may be ready.
On July 4, investor Gavin Baker revealed that SSI will release a new model this month that may have addressed continuous learning and sample efficient learning.
The marginal returns of pre-training are falling, and the real bottleneck is generalization and continuing learning.
Humanity is powerful not because it consumes the entire Internet all at once, but because we can continue to improve in limited experience.
Today, he is said to have re-entered the scaling orbit with his team, but not with simple stacks of parameters. The new expansion axis is the real-time learning capability of testing.
Ilya is redefining what is called "scaling" with the results of his study.
If the early results were Scaling, all current routes that rely on freezing weight + the above context window would appear obsolete, as in a hard drive simulation using tape.
The real difference is no longer who is more trained, but who can continue to renew themselves in the real world.
Human beings are able to organize and accumulate civilization because we have the ability to learn and remember continuously.
If AI really writes the experience itself, the boundaries of human-machine collaboration will be redrawn.
Of course, this also brings new problems. Real-time weight updates mean that risk control, security boundaries, and explainability will all face new challenges. A system that continues to change its own system, which is stronger than a frozen system and more difficult to fully predict.
This is a transition from "dead note hardback" to "really learn".
Ilya Sutskever once said,AGI's ultimate mission is to create a "gentle and noble mind".
If the mind can't leave a mark in the long rivers of time, if it can't physically reshape itself through trial, pain, awareness and success, then it will always be a tool, not a mind.
The SSI uses the TTT (test-training) paradigm to awaken life in the "time dimension".