Ilya Sutskever calls for new learning paradigm to fix AI 'jaggedness
Ilya Sutskever has been sounding the alarm for a while now, and it’s not just about the usual hype around ever-bigger models. While the industry proudly posts record scores on high-profile benchmarks, the OpenAI co-founder points out a nagging mismatch between those headline numbers and the day-to-day reliability most users actually need. He notes a pattern that feels almost like a mood swing: a system can crush a complex challenge one minute and then fumble a simple, routine task the next.
That wobbliness, he says, isn’t a random glitch, it looks more like a built-in flaw that chips away at trust in AI-driven products. To fix it, Sutskever isn’t just fine-tuning the usual pipelines; he’s pushing for a completely different learning approach, something he’s already testing in his own labs. The stakes feel real: without a shift, developers will keep wrestling with models that shine in the lab but stumble in real-world codebases.
Here’s the quote that captures his concern.
AI models suffer from "Jaggedness" A central problem with current models, according to Sutskever, is their inconsistency or "jaggedness." Models might perform excellently on difficult benchmarks, but often fail at basic tasks. He cites "vibe coding" as an example: A model recognizes a bug, introduces a new one while fixing it, only to restore the old bug on the next correction attempt. Sutskever suspects that Reinforcement Learning (RL) training makes models "a little too single-minded." Unlike pre-training, where one simply used "all the data," one has to be selective with RL.
This leads researchers--often unintentionally--to optimize models for specific benchmarks ("reward hacking"), which impairs generalization capabilities in the real world. Human emotions as a biological "Value Function" To reach the next level of intelligence, AI systems need to learn to generalize as efficiently as humans. A teenager learns to drive in about 10 hours, a fraction of the data an AI requires.
Sutskever theorizes that human emotions play a crucial role here by serving as a kind of robust "value function." These biologically anchored assessments help humans make decisions and learn from experiences long before an external result (as in classical RL) is available. "Maybe it suggests that the value function of humans is modulated by emotions in some important way that's hardcoded by evolution," says Sutskever. AGI is the wrong goal - Superintelligence is created on the job Sutskever also fundamentally questions the established term AGI.
The success of pre-training created the false expectation that an AI must be able to do everything immediately ("General AI"). However, this overshoots the target: "A human being is not an AGI," says Sutskever. Humans lack enormous amounts of prior knowledge; instead, they rely on continual learning.
His vision of a superintelligence, therefore, resembles an extremely gifted student rather than an all-knowing database.
Will the shift Sutskever talks about actually close the gap he describes? He says scaling has run into a ceiling, and that today’s models show a kind of “jaggedness” - they crush hard benchmarks but still trip over simple tasks. In his chat with Dwarkesh Patel, the SSI co-founder points to “vibe coding” as an illustration: a model can spot a bug but then fumble the fix.
He argues we need a learning style that mirrors human efficiency, and claims he’s already experimenting with it. Still, the road ahead is hazy; he mentions that open discussion of these ideas is getting tighter. The notion that fundamental research, not bigger models, will drive the next leap sounds plausible, yet we haven’t seen a working alternative yet.
It’s reasonable to stay skeptical until we get solid results. For now the AI community is watching what could be a turning point, aware that the promised efficiency gains haven’t materialized.
Further Reading
- Ilya Sutskever says a new learning paradigm is necessary and is already chasing it - The Decoder
- Ilya Sutskever breaks silence on AI's future - The Rundown AI
- OpenAI cofounder says scaling compute is not enough to advance AI - Business Insider
- Highlights from Ilya Sutskever's November 2025 interview - Effective Altruism Forum
- Ilya Sutskever - We're moving from the age of scaling to the age of research - Dwarkesh Podcast
More in Industry Applications
Ironwood TPU: Purpose‑Built Hardware for Inference as Industry Shifts Focus
It feels like the AI hardware market is quietly shifting. The big headlines still shout about...
Understanding Machine Learning: From Theory to Algorithms – Free Must‑Read
Scrolling through the endless stream of AI recommendations, the “5 FREE Must-Read Books for Every...
Westinghouse teams with Google Cloud to build AI platform for nuclear power
When Westinghouse announced its newest partnership, the headline was simple: Google Cloud is now...
Gemini for Education boosts learning for 1 M+ Italian university students
It might sound odd that one AI tool reaches over a million students across Italy, but that’s what’s...
Google's AI plan aims to overhaul a pressured global education system
When Google unveiled its new AI roadmap, the headline promised a kind of “cure-all” for schools...
Common Questions Answered
What does Ilya Sutskever mean by AI "jaggedness"?
Sutskever uses the term "jaggedness" to describe the inconsistency of current AI models, where they achieve high scores on complex benchmarks but frequently fail on simple, everyday tasks. This uneven performance, he argues, reveals a structural flaw rather than a random glitch.
How does the "vibe coding" example illustrate the problem of jaggedness?
In the "vibe coding" scenario, a model identifies a bug, attempts a fix, but then introduces a new bug and eventually reverts to the original error on subsequent attempts. This cycle shows how models can excel at spotting issues yet stumble when applying corrections, highlighting their unreliable behavior.
Why does Sutskever suspect Reinforcement Learning (RL) training contributes to jaggedness?
Sutskever believes RL training may make models overly specialized to specific reward signals, causing them to perform well on targeted benchmarks while neglecting broader, basic competencies. This over‑optimization can lead to the erratic performance patterns he describes as jaggedness.
What new learning paradigm does Sutskever propose to address the limitations of scaling?
He advocates for a learning approach that mimics human efficiency, focusing on consistent, reliable performance across both difficult and elementary tasks rather than merely increasing model size. Such a paradigm would aim to close the gap between headline benchmark scores and everyday reliability.