AI researcher Andrej Karpathy says agentic AI is years away from matching industry hype

Andrej Karpathy, former AI researcher at OpenAI and Tesla, is skeptical of the current hype surrounding agent-based AI and large language models.

AI researcher Andrej Karpathy believes that the current euphoria surrounding so-called agent AI is premature. In an interview with podcaster Dwarkesh Patel, Karpathy emphasizes that we should not be talking about a "year of agents" but, more realistically, about a "decade of agents".

Karpathy sees some practical benefits from systems like Codex and Claude Code, but says these models are nowhere near behaving like human interns or skilled employees—the kind of capability many labs are aiming for.

"They just don't work"

The issues, he argues, are fundamental: the models lack core cognitive abilities, aren't truly multimodal, have no reliable memory, and can't consistently handle complex computer tasks. "They just don't work," Karpathy says.

He estimates it will take around ten years for the field to solve these deep-seated problems, based on his "extrapolation with respect to my own experience in the field."

"I feel like the problems are tractable, they’re surmountable, but they’re still difficult. If I just average it out, it just feels like a decade to me," Karpathy says.

For now, he sees a disconnect between industry hype and technical reality: "There’s some over-prediction going on in the industry." Karpathy, who left OpenAI in early 2024, is part of a growing group of critics who view today's large language models as useful but fundamentally limited.

Karpathy prefers autocomplete over agent intelligence

Karpathy thinks even advanced models like GPT-5 Pro are only useful in narrow roles, such as serving as an "oracle" for code analysis. "Often it’s not too bad and surprisingly good compared to what existed a year ago," he says.

But that's not enough for real software integration. The models regularly struggle with project-specific styles, dependencies, and assumptions, and Karpathy doesn't see them solving new problems outside the data they were trained on.

"Overall, the models are not there," Karpathy says. For now, he sees their real value in simple autocomplete tasks: "For now, autocomplete is my sweet spot." Still, he points out, the industry is overstating what's possible: "I feel like the industry is making too big of a jump and is trying to pretend like this is amazing, and it’s not. It’s slop."

"The internet is terrible"

Karpathy points to poor training data as a major issue. Most language models today are pretrained almost exclusively on online data, which he calls useless. Instead of high-quality journalism like the Wall Street Journal, most of what ends up in training datasets is just fragments, symbols, and noise. "It's total garbage. I don’t even know how this works at all."

Because of this, much of what these models do is just memorization, not real understanding. "There's a huge amount of gains to be made because the internet is terrible," Karpathy adds. His solution is to use intelligent models to curate future training data by filtering out junk and keeping only meaningful content. The aim is smaller, more efficient models with a real cognitive core - the ability to learn from condensed, high-quality information instead of just memorizing huge amounts of random web data.

This argument strengthens the position of content creators in ongoing fair use debates. By emphasizing the value of top-tier sources like the Wall Street Journal, Karpathy gives publishers leverage. If the best models require carefully curated, professional data, the case for compensating content providers—rather than just scraping the web—gets much stronger.

Progress without a breakthrough

Karpathy doesn't believe there will be a single breakthrough that changes everything in AI. Instead, he sees progress coming from lots of small, coordinated steps: better training data, stronger model architectures, improved learning processes, and faster hardware. He frames progress toward agentic AI as a longer, incremental effort.

Back in August, Karpathy was already skeptical about the potential of reinforcement learning—the technique many labs use to optimize so-called language reasoning models through reward functions. He thinks these reward signals are too unreliable and easy to game for complex reasoning. Still, Karpathy sees reinforcement learning as a step up from plain imitation learning, with much room to improve.

Long term, he argues for a new approach: language models should learn by interacting with and experiencing environments, not just by mimicking human text. This puts him in line with DeepMind researchers like Richard Sutton and David Silver, who have been pushing for similar ideas.