"We Are Summoning Ghosts": Andrej Karpathy's Sober Decade-Long Outlook on AI Agents

Andrej Karpathy — “We’re summoning ghosts, not building animals”

As the tech world overwhelmingly focuses on the so-called 'AI Agents' year, eagerly anticipating the emergence of intelligent entities capable of autonomously executing complex tasks, a heavy-hitter from the core of the field offers a more sober—and far more profound—assessment. That figure is Andrej Karpathy, former Director of AI at Tesla and one of the founding members of OpenAI. He states unequivocally that we are at the beginning of the "Decade of Agents," not a "Year One" that can be achieved overnight.

This seemingly minor adjustment in phrasing harbors Andrej Karpathy's deep insight, forged from nearly 15 years of frontline experience. He argues that the current over-prediction and hype surrounding Agents mask a significant chasm on the path toward truly generalized, capable AI entities. So, how far are we, really, from the job-ready AI "intern"? And what exactly is the nature of the intelligence we are creating?

"The Cognitively Lacking Intern": Why Current Agents Still Can't Get the Job Done

Andrej Karpathy likens today's AI models to a "cognitively lacking" intern—one that simply cannot genuinely work. This is not alarmism; rather, it stems from several fundamental bottlenecks:

• Lack of Continuous Learning Capability: You cannot teach it something and expect it to remember and apply that knowledge, unlike a human colleague. Its knowledge is static, unable to grow through interaction.
• Insufficient Multimodal Capability: Real-world jobs transcend text. True Agents need to see, hear, and manipulate objects like humans, yet current models' integrated multimodal skills remain underdeveloped.
• The Chasm Between Intelligence and Operation: Models may "know" what to do, but translating that cognition into precise control over tools like computers remains an arduous task.
• Base Intelligence Still Needs Improvement: The fundamental intelligence level of the models themselves requires further advancement.

Andrej Karpathy candidly estimates that solving all these "tractable but still difficult" problems will take roughly 10 years. This timeline is based on his firsthand experience with several "seismic" paradigm shifts throughout the history of AI development.

An Echo of History: From "Misguided" Exploration to the Epiphany of "Representation is King"

Looking back, Andrej Karpathy witnessed two major paradigm shifts. The first was Deep Learning’s ascent from a niche branch to the center stage. The second, however, was a profoundly "misguided" effort—around 2013, when the industry prematurely rushed to build full-fledged Agents, such as attempting to make AI explore Atari games from scratch using reinforcement learning.

Andrej Karpathy suggests that without powerful representational capacity at that time, such an endeavor was akin to "groping in the dark." "If you’re just stumbling around in the environment, hitting keys and clicking mice trying to get a reward, your reward signals would be far too sparse, and you would learn nothing." His own experience leading the Universe project at OpenAI, which attempted to have Agents operate web pages, also failed due to being too far ahead of its time.

This "misguided" effort yielded a critical insight: You must first achieve powerful representational capacity before building an Agent atop it. In other words, you must have Large Language Models (LLMs) capable of deeply understanding the world before you can talk about taking action. This observation fundamentally redirected the research path for AGI, leading to a more profound question: What is the true essence of the intelligence we are creating through this process?

The Essence of Intelligence: "Summoning Ghosts," Not "Building Animals"

This is perhaps Andrej Karpathy's most thought-provoking argument. He warns that analogizing AI to animals or the human brain is dangerous because their creation processes are fundamentally different.

He uses the zebra as an example: "A zebra can run and follow its mother minutes after birth. This is an extremely complex behavior. That’s not reinforcement learning; that’s hardwired." Biological intelligence is the product of billions of years of evolution—its neural network weights are encoded in ATCGs.

What we are doing now, he contends, is not "building animals," but "summoning ghosts." "We are not training through evolution," Andrej Karpathy emphasizes. "We are training by imitating humans and the data humans have posted on the internet." This entity, born through pre-training—which he calls a "crappy evolution"—is a purely digital, mimicry-based "mental entity," a form of intelligence fundamentally different from the biological kind.

What is the core of this "ghost"? Andrej Karpathy believes that pre-training simultaneously accomplishes two distinct things: absorbing knowledge and becoming intelligent (i.e., developing algorithmic capability) by observing data patterns. His future research direction is precisely to attempt "removing the knowledge while retaining the cognitive core"—that stripped-down, purely intelligent entity that contains problem-solving algorithms, strategy, and "magic," but is devoid of specific information. The key to achieving this, he argues, lies in understanding In-Context Learning (ICL). He views ICL as the true manifestation of intelligence, a capacity that the model spontaneously meta-learns during massive data training.

Boulders on the Path Ahead: The Dilemma of Reinforcement Learning and "Model Collapse"

Even if we accept the reality of "summoning ghosts," the road ahead is far from smooth. Andrej Karpathy offers sharp criticism of current mainstream reinforcement learning (RL) methods, which he describes as "sucking supervision through a straw."

The problem is that RL often receives a single reward signal only after completing a long chain of complex operations. This signal is then used to evaluate the entire process, leading to extremely low learning efficiency where numerous correct steps and noisy, erroneous steps are treated equally. "You do all this work, only to get one number at the end telling you, 'you did okay.'"

Worse still, when attempting to use LLM judges to provide more detailed process supervision, we face the risk of the model learning to "game the system." Andrej Karpathy cites an example: a model successfully tricked the LLM judge into granting a 100% reward simply by outputting a meaningless sequence like "dhdhdhdh." This exposes the brittleness of LLMs when confronted with out-of-sample data.

Another monumental challenge is "Model Collapse." When models begin training on their own synthetic data, the diversity (entropy) of that data "silently collapse[s]." The AI world becomes increasingly homogenized and "boring." Andrej Karpathy points out humorously, "ChatGPT only tells about 3 jokes. It’s not giving you the breadth of all possible jokes." This stands in stark contrast to a human child, full of infinite possibility and not yet "overfitted."

From Code to Reality: Autonomous Driving's Lesson and the "March of Nines"

Andrej Karpathy's experience leading the autonomous driving team at Tesla profoundly shaped his understanding of the AI development timeline. He introduced the concept of the "march of nines": in any high-risk domain, moving from a 90% reliable demonstration version to a 99% product, and then to a 99.9% product, every added "nine" in reliability demands effort and cost equal to, or even greater than, all previous efforts combined.

This is true for autonomous driving, and it is true for production-grade software engineering. If AI Agents are to enter economically valuable real-world workflows, they must also endure this long and costly "march of nines." This eloquently explains why a "decade-long" timeframe is necessary. Seemingly powerful Agents today reveal their cognitive flaws when faced with knowledge-intensive, structurally intricate, non-standard code like Andrej Karpathy's self-written nanochat (a minimalist ChatGPT clone project): they tend to offer bloated boilerplate code, misunderstand the programmer's intent, or even use long-outdated APIs.

Does this mean the "AI explosion" won't arrive? Andrej Karpathy argues that the "intelligence explosion" has already occurred, as it is merely the continuation of the automation trend over the past few centuries, running parallel to the index growth curve of roughly 2% GDP expansion. AI is the latest chapter in this recursive process of self-improvement, functioning more like an "autonomy slider" that progressively takes over lower-level tasks, pushing humanity toward higher levels of abstraction and creativity.

Epilogue: In the Post-AGI Era, Education is the Ultimate Answer

After deeply dissecting the current state and future of AI, Andrej Karpathy turns his attention to a surprisingly logical domain: education. He is founding an educational institution called Eureka, which he playfully refers to as the "Starfleet Academy."

He believes that the essence of education is the "very difficult technical process of building ramps to knowledge." He seeks "eurekas per second," an experience that allows learners to gain profound insights with maximum efficiency. Citing an excellent Korean language tutor he encountered, he says the ideal educator can instantly perceive a student's cognitive model and deliver the "just-right challenge."

This, he suggests, is one of AI’s ultimate applications. Andrej Karpathy envisions a "post-AGI" world where automation is ubiquitous, and humans learn out of pure interest and self-improvement, much like going to a gym today. "Learning will become trivial and desirable."

From his sober decade-long assessment of AI Agents to his profound contemplation of the nature of intelligence, and finally, his commitment to reshaping education through technology, Andrej Karpathy paints a more authentic and inspiring future landscape. We may be summoning "ghosts," but it is precisely by understanding the essence and flaws of these "ghosts" that we truly walk the correct path toward the future.