Ilya Sutskever's Safe Superintelligence (SSI) lab may release its first groundbreaking model as early as this week!
Industry insiders are heavily hinting that SSI will launch its first disruptive large model.
Silicon Valley Bigwigs Crazy Teasers: "AI Will Be Completely Upended!"
Well-known investor and a16z partner Martin Casado was so excited he directly hinted that he recently saw "the most important new model of the year."
The industry immediately erupted, with attention instantly focusing on Ilya's SSI.
The reason is simple: a16z is one of the core investors behind SSI. Although some speculate that what Casado saw might be OpenAI's next-generation Astra model.
But rumors of a breakthrough from a "non-mainstream" player are rampant.
Notable tech observer Andrew Curran was even more excited to point out that this major breakthrough doesn't seem to come from a mainstream giant (like OpenAI or Google), but from an independent lab, which feels "very real."
The most exaggerated claim comes from AI watcher Dan McAteer, who outright stated: "Ilya has truly created superintelligence, the game has changed!"
Earlier this month, investor Gavin Baker flatly stated on a podcast:
SSI says they will release their model in August.
If you think this is just talk, look at where the real money is flowing.
On July 27th of this year, NVIDIA suddenly announced a long-term strategic partnership with this mysterious company with no revenue and no product, and will insanely increase SSI's computing power scale by 10 times in the next 12 months.
There's even hotter gossip: the investment is as high as $50 billion, and NVIDIA has granted SSI exclusive access to the next-generation Vera Rubin system!
Why is Huang Jen-Hsun willing to bet such a huge sum?
In the press release, NVIDIA stated that they decided to make this heavy bet only after "gaining rare access to its closely guarded research results."
What did Huang Jen-Hsun see? Combined with today's leaks, the answer seems obvious.
Upending Common Sense: AI Learned to "Think and Learn On the Fly"
To understand how shocking SSI's new model is, we must first understand how current AI (like ChatGPT, Claude) works.
Current AI models, regardless of how many parameters they have, essentially solidify knowledge into their weights during the "pretraining" phase. Once training is complete, its "brain" is frozen.
To allow models to process new information, major companies are frantically expanding the "Context Window," from 100k to millions of tokens.
This is like an open-book exam—the model's brain doesn't get smarter; you're just allowing it to bring a thicker and thicker "cheat sheet" (context) into the exam.
Even reasoning models like OpenAI's o1, which focus on "Test-time Compute," are essentially just spending more time thinking on scratch paper (consuming more tokens). The connections in its neural network, its "brain cells," are still static.
But what SSI has reportedly achieved is a new architecture based on Test-Time Training (TTT).
What does this mean?
When this model reads a long document you provide, it doesn't stuff the document into a "cheat sheet." Instead, it truly "learns" it, generates gradient updates, and changes the structure of its own brain.
After finishing reading, it has already become a minutely but entirely new, evolved AI.
It no longer needs a massive context window because it has turned what it read into its true "internalized knowledge."
It is no longer limited by the computing power monopoly of pretraining. A small, refined model that can continuously adapt and evolve on the job could easily surpass those massive, inflexible behemoths built with vast computing power.
While the industry currently competes on "how long a model can think," Ilya is betting on "whether a model can change itself."
Ilya's Crazy Foresight
Looking back over the past two years, SSI's "zero products, zero papers" status once made outsiders wonder if they were stuck.
But if you connect the timeline, you'll find Ilya has been playing a long game.
Ilya has repeatedly conveyed a core idea to the public: 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, on Dwarkesh Patel's podcast, he went further: "We are moving from the Scaling era to the Research era."
Ilya believes the entire industry has been misled by the terms "AGI" and "pretraining." Humans are not born omniscient.
In his mind, true superintelligence is not a massive machine that memorized the entire internet at birth; it is an "extremely smart, infinitely curious 15-year-old genius."
This 15-year-old might initially know nothing, but if you place them in any role, through continuous trial, error, and learning, they can quickly master programming, medicine, law, or even any unknown skill.
He further explained: Humans themselves are not "AGI out of the box," but rely on continual learning.
True superintelligence should be the same: deployment itself is a learning process, evolving through real-world feedback, not being "finished" after a one-time pretraining.
This philosophy directly determines SSI's strategy: no pursuit of short-term products, no release of intermediate models, only aiming for a "straight shot to safe superintelligence," focusing research on efficient continual learning and alignment.
To achieve this, "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 unusually low-key. If you're building a new species that completely overturns the existing paradigm, you wouldn't be blogging about it before the paradigm is even fully built.
But the truth can't stay hidden forever; clues were already planted.
In July 2024, scholar Yu Sun et al. published the original TTT paper.
Most crucially, Stellar co-founder and SSI investor Jed McCaleb co-authored a paper, bluntly stating:
Long-context language modeling is not an architectural problem at all, but a continual learning problem!
Consistent research direction, investors personally co-authoring papers, and now today's leaks—all clues point to the same fact: SSI has turned TTT from an academic concept in the lab into a real commercial weapon.
Conclusion: The Second Half of AI Has Just Begun
Now, all eyes are on this August.
Whether it's a test version accessible only to a small group of geeks or a stunning public release, if SSI's first model truly possesses "Test-Time Training" and "real-time weight updating" capabilities, the entire logic of the AI industry will be overturned.
The computing power moats piled up in data centers by major companies, the business model charging by the million tokens, even the debate over open-sourcing weights—all will face a dimensional reduction strike.
This proves one thing: the AI race is far from the final stage where "money and computing power can guarantee victory." Real technological leaps are still hidden in the top minds daring to break common sense.
This time, Ilya Sutskever stands again at history's crossroads. Back then, it was his line of code that brought deep learning back into the light; today, perhaps it is him again who will personally end the "Pretraining era" of large models.
Do you think Ilya can ascend to legend status again this time? If AI can truly "evolve as it's used," how far is humanity from completely losing control?
References:
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 Zhiyuan," author: ASI Revelation; editor: David
