The laboratory "Safe Superintelligence" (SSI), founded by Ilya Sutskever, may launch its first major model this week!
Industry insiders are strongly hinting that SSI will launch its first groundbreaking large model.
Silicon Valley giants leak shocking preview: "AI will be completely disrupted!"
Renowned investor and a16z partner Martin Casado was thrilled, directly hinting that he recently encountered "the most significant new model of the year."
The community immediately erupted, with all eyes instantly locking onto Ilya's SSI.
The reason is simple: a16z is one of the key investors behind SSI. Although some have speculated that Casado may have seen OpenAI’s next-generation Astra model.
But rumors of breakthroughs from "non-mainstream giants" are rampant.
Prominent tech observer Andrew Curran enthusiastically noted that this major breakthrough came not from a mainstream giant like OpenAI or Google, but from an independent lab—making it feel “very real.”
Most dramatic was AI observer Dan McAteer, who bluntly stated: "Ilya has truly created superintelligence—the rules of the game have been changed!"
At the beginning of the month, investor Gavin Baker bluntly stated on a podcast:
SSI said they will release their model in August.
If you think this is just theoretical, take a look at where the real money is flowing.
On July 27 of this year, NVIDIA unexpectedly announced a long-term strategic partnership with a mysterious company that has no revenue or products, and plans to increase SSI's computing power by tenfold within the next 12 months.
More shocking details emerge: This investment reaches $5 billion, and NVIDIA has granted SSI exclusive access to the next-generation Vera Rubin system!
Why is Jensen Huang willing to invest such a huge amount?
In the press release, NVIDIA stated that they decided to make a significant investment after gaining rare access to their closely guarded research findings.
What did Jensen Huang see? Given today’s leaks, the answer is clear.
Defying Common Sense: AI Has Learned to "Think and Learn Simultaneously"
To understand how groundbreaking the SSI model is, we must first understand how current AI systems (like ChatGPT and Claude) work.
Current AI models, regardless of their parameter size, essentially固化 knowledge into their weights during the pre-training phase. Once training is complete, their "brain" is frozen.
To enable models to process new information, major companies are fiercely competing to expand the context window, from 100k to millions of tokens.
It’s like an open-book exam—the model’s brain hasn’t become smarter; you’ve just allowed it to bring in thicker and thicker “cheat sheets” (context).
Even reasoning models like OpenAI's o1, which emphasize test-time compute, are essentially just spending more time thinking on scratch paper (consuming more tokens)—the connections within their neural networks, their "brain cells," remain static.
However, SSI's latest disclosure involves a completely new architecture based on Test-Time Training (TTT).
What does this mean?
When this model reads a long document you provide, it doesn't just cram the information into a cheat sheet—it truly learns it, generating gradient updates that alter its internal structure.
After reading, it had transformed into an extremely subtle yet newly evolved AI.
It no longer needs a large context window because it turns what it reads into its own true "internalized knowledge."
It is no longer constrained by the computational dominance of pre-training; a small, refined model that can continuously adapt and evolve on the job is enough to outperform those massive, inflexible models built with enormous computational power.
Currently, the industry is competing over "how long a model can think," but Ilya is betting on whether a model can change itself.
Ilya's Crazy Vision
Looking back over the past two years, SSI’s “no products, no papers” status led some to question whether they were in trouble.
But if you connect the timeline, you'll see that Ilya has been playing a much larger game all along.
Ilya has repeatedly conveyed to the public a core idea: the era of pretraining as we know it is coming to an end.
At the 2024 NeurIPS conference, he made this astonishing prediction.
By November 2025, he went further on Dwarkesh Patel’s podcast, stating: “We are moving from the era of scaling to the era of research.”
Ilya believes the entire industry has been misled by the terms "AGI" and "pre-training." Humans are not born with knowledge.
His true vision of artificial intelligence is not a massive machine that memorizes all online data at factory setup, but rather an “extremely intelligent, endlessly curious 15-year-old genius.”
This 15-year-old may initially know nothing, but if you place him in any role and allow him to learn through continuous trial and error, he can quickly master programming, medicine, law, or any other unknown skill.
He further explained: Humans themselves are not AGIs that are "fully capable out of the box," but rather rely on continual learning.
True superintelligence should be the same: deployment itself is a learning process, continuously evolving through real-world feedback, rather than being "completed" after a one-time pre-training.
This philosophy directly shapes SSI’s strategy: avoiding short-term products, not releasing intermediate models, and focusing solely on a “straight shot to safe superintelligence,” with research centered on efficient continual learning and alignment.
To achieve this goal, meta-learning is the only solution—the model must not only master skills but also master the method of acquiring skills.
This perfectly explains why SSI has been so low-key. If you're building an entirely new species that disrupts the existing paradigm, you certainly wouldn't be blogging about it while the paradigm is still being built.
But paper can't wrap fire—the clues were already laid.
In July 2024, scholars Yu Sun and others published the original TTT paper.
Most importantly, Jed McCaleb, co-founder of Stellar and an investor in SSI, co-authored a paper that explicitly stated
Long-context language modeling is not an architecture problem at all—it's a continual learning problem!
From aligned research directions and investors personally co-authoring papers to today’s revelations, all clues point to one fact: SSI has transformed TTT from an academic concept in the lab into a real commercial weapon.
Conclusion: The second half of AI has just begun.
All eyes are now on August.
Whether it’s a limited beta available only to a small group of enthusiasts or a public launch that shakes the industry, SSI’s first model will overturn the entire logic of the AI industry once it truly achieves “training during inference” and “real-time weight updates.”
The moats of computing power hoarded by major tech companies in data centers, business models charged per million tokens, and even debates over open-source weights, will all face a dimensional downgrade.
This proves one thing: the AI race has not yet reached a stage where victory is determined solely by financial and computational power. True technological breakthroughs still reside in the minds of top innovators who dare to challenge conventional wisdom.
This time, Ilya Sutskever stands once again at a historical crossroads. It was his code that brought deep learning back to life; now, perhaps it is he who will personally bring an end to the era of pre-training large models.
Do you think Ilya can achieve godlike status again? If AI can truly evolve while being used, how far are we from losing complete control?
Reference materials:
https://x.com/hakmgpt/status/2091855200713638146
https://x.com/daniel_mac8/status/2091891607641440598
https://x.com/AndrewCurran_/status/2091890441465499995
https://x.com/martin_casado/status/2091650951736361073
https://aimidday.com/ssis-first-model-reportedly-trains-itself-while-it-thinks/
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation; edited by David.