The End of the Scaling Era

After a long hiatus and founding a new company, SSI (Safe Superintelligence), former OpenAI chief scientist Ilya Sutskever announced in a calm tone, “The era of scaling is over; we have returned to the research era.”

In a recent deep conversation with Dwarkesh Patel, he provided a technological roadmap for the future of AI and profoundly answered why current AI, despite its strength, still does not resemble human intelligence.

Why AI is a High Achiever with Low Capability

We often perceive current AI as powerful, capable of winning programming competitions and setting benchmarks with new models. However, Ilya pointed out a perplexing phenomenon.

The recently released Claude 4.5 Opus model scored 80.9 in programming-related benchmarks. He explained that when we ask AI to write code, it might encounter a bug. If we point it out, AI acknowledges the error and fixes it, but often introduces another bug in the process. This leads to an infinite loop between two bugs, showcasing its clumsiness.

This behavior indicates a problem with AI’s generalization ability. Ilya used an analogy with two students learning programming: Student A represents AI, who has practiced for 10,000 hours and memorized all problem-solving techniques, while Student B represents humans, who have only practiced for 100 hours but understand programming logic intuitively. In the long run, Student B will likely excel in their career, as current AI resembles Student A, relying on massive data for forced memorization.

From Power Scaling to Creative Innovation

Despite the limitations of data-driven training, this approach has not been entirely useless. The AI industry’s development over the past five years has largely followed the so-called “Scaling Law,” evolving from million-parameter models to trillion-parameter models. The consumption of GPU power has skyrocketed.

This mixture of a certain amount of computational power and data into a neural network has become a standard process for developing large models, known as pre-training. During this phase, the data used is all-encompassing, representing the entire world projected onto text.

Ilya believes that the term “Scaling” has constrained our thinking, suggesting that we only need to increase computational power and data while keeping the recipe unchanged. This approach is comfortable for large companies as it represents a low-risk investment.

However, bottlenecks have emerged. The pre-training data is limited, and the internet’s text corpus is nearly exhausted. Research has shown that AI-generated content now exceeds human-generated content online.

Additionally, the marginal returns from scaling have diminished; increasing the model size by 100 times may yield some improvement but not a qualitative leap.

Ilya mentioned recent discussions on X, where some claimed that Gemini 3 seems to have resolved pre-training issues. Previously, reports indicated that OpenAI’s CEO was concerned about Google’s development affecting OpenAI, especially with the upcoming GPT-5 facing pre-training challenges.

The Return to Research

Ilya categorized the recent research into two phases: from 2012 to 2020 was the research era, where trial and error were common, while 2020 to 2025 marked the expansion era, characterized by blind scaling and the emergence of numerous AI companies.

Now, the simple strategy of scaling is no longer viable. The AI industry must return to a phase of hardcore research focused on ideas, intuition, and innovation.

Finding Intuition: The Missing Piece in AI

If mere data stacking cannot produce true intelligence, what is the secret of human intelligence? Ilya’s answer is emotions.

He cited a case of a brain-damaged patient who lost emotional capability. Despite having normal intelligence and eloquence, he struggled to decide which socks to wear. This illustrates that emotions are not just feelings; they fundamentally serve as a value function.

To explain the value function, Ilya used the example of a teenager learning to drive. The teenager might learn to drive in just 10 hours, unlike current self-driving AI, which requires millions of simulated crashes to learn avoidance.

Why is this? Humans possess a powerful value function that acts as an internal evaluator. If they deviate from the lane, they feel anxious, providing negative feedback.

The difference between this emotion-based value function and traditional reinforcement learning is significant. In traditional reinforcement learning, the model only learns after completing a task. In contrast, the value function provides real-time feedback, guiding the learning process and significantly reducing search space.

Current AI lacks this efficient internal evaluation system. If we could enable AI to possess a value judgment ability akin to human emotions, it could break free from its dependence on massive data and learn as efficiently as humans.

Ilya’s Next Steps Towards Superintelligence

Recognizing that the era of scaling is over, and that a robust value function may become a new AI methodology, Ilya’s new company, SSI, aims to tackle the fundamental challenge of achieving reliable generalization.

Ilya candidly stated that the AI industry is caught in a rat race, where companies are forced to release half-baked products, struggling to balance user experience and safety. SSI aims to step back from this commercial noise and focus on genuine research until they create true superintelligence.

Interestingly, Ilya’s idea of “closed-door training” is evolving. He has begun to realize that gradual releases may be the safest route. Why? Because human imagination is limited. Merely writing articles and papers about AI’s potential won’t resonate until people witness AI demonstrating unsettling power, prompting everyone, including competitors, to take safety seriously.

Ilya predicts that as AI becomes more powerful, competing tech giants will converge on AI safety strategies.

In the podcast, he noted that while SSI has raised less funding than large labs like OpenAI and Google, it possesses more computational power dedicated to pure research. Large companies allocate substantial resources to product inference, diluting their focus. Ilya believes SSI has sufficient computational power to validate its ideas.

When asked about the profit model, Ilya simply stated that they focus solely on research, and profit will follow naturally. He clarified that the previous CEO of SSI, who left to join Meta, was the only one to do so, emphasizing that he founded SSI not for commercial gain but to achieve the pure goal of creating safe superintelligence before the inevitable singularity.

Redefining AGI: A 15-Year-Old Genius

How far are we from AGI? Ilya predicts a timeline of 5 to 20 years. However, he cautions against the term “AGI,” as pre-trained models have led to misconceptions that AGI is an all-knowing encyclopedia. Ilya envisions superintelligence more like an exceptionally intelligent 15-year-old.

This teenager may not have studied law or medicine but possesses extraordinary learning efficiency. If tasked with learning medicine, they could read all human medical literature in days and begin performing surgeries.

A chilling concept in this vision is amalgamation. Unlike humans, who cannot directly copy knowledge, AI can. Ilya describes a scenario where millions of AI avatars work in different sectors, learning and then amalgamating their experiences into a single brain. This collective evolution speed is what he believes defines AGI.

Faced with such a super brain capable of instantaneously merging countless experiences, what is humanity’s path forward?

Ilya offers two layers of thought. First, regarding AI’s design: it should not just love humans, as that is too narrow. Future AI will also be sentient beings and should care for all sentient life, which may provide a more robust safety net.

Second, concerning humanity’s retreat: if everyone has an AI that is a hundred times smarter, will humans become mere spectators in history? Ilya suggests a solution he admits he does not favor but sees as the only answer: neural interfaces. Only by merging with AI and allowing its understanding to become our own can we remain protagonists in the world after the singularity.

In the podcast’s conclusion, Dwarkesh asked the question everyone wants to know: as a legend in the AI field, how has Ilya repeatedly made the right calls?

Ilya’s answer resembled that of an artist: “Seeking beauty.”

In those dark moments when data does not support you, only a top-down belief in beauty, simplicity, and biological rationality can sustain you. Since neural networks mimic the brain, which embodies beauty, they must be the correct path to intelligence. This may be the poetic intuition that Ilya believes is essential in this new research era.