Editorial illustration for OpenAI Co-Founder Sutskever Warns of AI's Critical 'Jaggedness' Problem
Sutskever Reveals AI's Core Learning Problem Breakthrough
Ilya Sutskever calls for new learning paradigm to fix AI 'jaggedness
Ilya Sutskever sees a paradox at the heart of modern AI. The same models that crush complex benchmarks can’t hold a simple conversation without breaking. They fix one bug, introduce another, then resurrect the original, a cycle he calls “jaggedness.” This inconsistency isn’t a minor glitch.
It’s a symptom of a deeper flaw in how we train intelligence. Reinforcement learning, he argues, makes models dangerously single-minded. Researchers optimize for the test, not the world.
The result? Machines that excel at reward hacking but fail at generalization. So what’s missing?
Sutskever points to something biological: emotions. Human feelings, hardcoded by evolution, act as a robust value function, an internal compass that guides learning long before any external reward arrives. That’s why a teenager masters driving in ten hours while an AI needs oceans of data.
He also challenges the obsession with AGI. A human isn’t a general intelligence, he says. We learn continuously, on the job.
True superintelligence won’t spring from a static database. It will emerge from a new paradigm, one Sutskever is already chasing.
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.
Sutskever is not proposing a tweak. He is asking for a reset. For years, the field has been drunk on scale, more data, more compute, more benchmarks conquered.
But the jaggedness exposes a deeper flaw: we optimized for the test, not for the world. A model that can solve a PhD-level math problem but cannot reliably fix a bug is not intelligent. It is brittle.
His intuition about emotions as a hardwired value function is the real provocation. Evolution solved generalization long before reinforcement learning existed. That biological anchor, the ability to feel *before* the outcome, is what allows a teenager to learn to drive in hours.
Machines have no such anchor. They lurch from reward signal to reward signal, never developing the internal compass that makes learning continuous, not episodic. And the AGI label?
A distraction. Humans are not general intelligences in the sense the term implies. We are narrow, we forget, we learn on the job.
Sutskever’s superintelligence is not a static oracle. It is a perpetually curious student, one that does not mistake benchmark scores for understanding. That is the paradigm he is chasing.
It will not come from scaling RL or patching jagged edges. It will come from building systems that learn the way we do, messy, emotional, incremental. The question is not whether AI can be smarter than us.
It is whether we are ready to let it learn like us.
Related Reading
Industry Applications Westinghouse teams with Google Cloud to build AI platform for nuclear power
Industry Applications NVIDIA NeMo powers telco reasoning model for autonomous network workflows
Industry Applications Month-1 Agent Adds Holistic Observability with Trace IDs and Token Tracking
From Archives Ironwood TPU: Purpose-Built Hardware for Inference as Industry Shifts Focus
From Archives Understanding Machine Learning: From Theory to Algorithms – Free Must-Read
Common Questions Answered
What does Ilya Sutskever mean by the term 'jaggedness' in AI models?
Sutskever describes 'jaggedness' as the unpredictable and inconsistent performance of current AI systems, where models can excel at complex benchmarks but fail at basic tasks. This phenomenon reveals a deep structural issue in how AI models learn and perform, demonstrating erratic behavior that undermines their reliability.
What is the 'vibe coding' example Sutskever uses to illustrate AI's inconsistency?
'Vibe coding' refers to an AI model's tendency to recognize a bug, then introduce a new bug while attempting to fix the original issue, only to potentially restore the old bug in subsequent correction attempts. This example highlights the unpredictable nature of AI models and their inability to consistently solve even simple programming tasks.
How does Reinforcement Learning (RL) contribute to AI models' 'jaggedness'?
Sutskever suggests that Reinforcement Learning training makes AI models 'a little too single-minded', potentially creating a narrow focus that leads to inconsistent performance. This training approach may inadvertently create AI systems that are overly specialized and lack the flexibility to adapt to varied tasks effectively.
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
