It’s nearly impossible to escape the constant stream of AI hype. Daily announcements can make it feel like superintelligence is just around the corner. But for those in the trenches building these systems, the reality is far more complex. Andrej Karpathy, a renowned AI engineer who has led teams at both OpenAI and Tesla, approaches the field with an engineer’s “hard hat on,” offering a perspective that is deeply technical, refreshingly grounded, and often surprising.

In a recent conversation with Dwarkesh Patel, Karpathy broke down the practical realities of building AI today. This article distills four of his most counter-intuitive and impactful ideas—lessons learned from the front lines that cut through the hype and reveal the true state of artificial intelligence.

——————————————————————————–

1. We’re Summoning Ghosts, Not Building Animals

It’s common to hear AI models compared to human or animal brains, but Karpathy argues this analogy is fundamentally flawed. He proposes a different way to think about the intelligence we’re creating, one grounded in engineering reality.

Animals are products of a long, slow evolution that bakes immense capability directly into their hardware. A newborn zebra, for instance, can run and follow its mother minutes after birth—a feat of complexity that isn’t learned, but inherited. Karpathy notes that we simply don’t know how to run that optimization process.

Instead, we have what he calls a “crappy evolution”: pre-training. It’s the messy, imitation-based process we have to use because it’s the only practical version available to us. This results not in evolved creatures, but in what Karpathy calls “ghosts” or “spirits.” They are ethereal, purely digital entities whose entire nature is a compressed, “hazy recollection of the internet documents” they were trained on.

This distinction is crucial. It reframes our expectations and research, moving away from strict biomimicry and toward understanding the unique properties of an intelligence born from imitating a vast, static collection of human data.

In my post, I said we’re not building animals. We’re building ghosts or spirits or whatever people want to call it, because we’re not doing training by evolution. We’re doing training by imitation of humans and the data that they’ve put on the Internet.

——————————————————————————–

2. Today’s Reinforcement Learning Is “Terrible”

Reinforcement Learning (RL) is a key technique for improving AI models, but Karpathy offers a blunt critique of how it currently works, labeling the process “terrible,” “noisy,” and “stupid.”

The standard approach is outcome-based. A model attempts a problem (like a math equation) in hundreds of ways. It then looks at which attempts produced the correct answer and reinforces every single step taken in those successful paths.

Karpathy finds this incredibly inefficient because it incorrectly up-weights every step in a successful chain—including inefficient detours, lucky guesses, and outright mistakes—as long as the final outcome was correct. It rewards luck as much as skill.

A human, by contrast, engages in a “complicated process of review.” We reflect on our strategy, identifying which specific parts were effective and which were flawed, not just the final result. This flaw in AI learning reveals the urgent need for better supervision methods and is a major reason models still struggle with complex, multi-step reasoning.

The way I like to put it is you’re sucking supervision through a straw. You’ve done all this work that could be a minute of rollout, and you’re sucking the bits of supervision of the final reward signal through a straw… It’s just stupid and crazy. A human would never do this.

——————————————————————————–

3. AI Is Surprisingly Bad at Writing Novel Code

Coding is often hailed as AI’s biggest success story, but Karpathy’s recent experience building nanochat—a ChatGPT clone from scratch—reveals a more nuanced reality. He identifies three types of users today: those who reject LLMs, “vibe coders” who ask an agent to write entire features, and “intermediate” users like himself, who rely on autocomplete but remain the architect. From this pragmatic sweet spot, he identified a critical weakness.

LLMs excel at writing boilerplate code and implementing patterns common on the internet. However, they struggle profoundly with code that has “never been written before” or deviates from standard conventions. When Karpathy implemented a custom gradient synchronization, the models repeatedly failed to understand his intent. They kept trying to add defensive “try-catch statements” and turn his focused project into a bloated “production code base,” producing a “total mess.”

This firsthand experience directly informs his skepticism about the “year of agents.” If today’s agents, with their many “cognitive deficits,” produce “slop” when faced with a simple custom implementation, they are nowhere near ready to autonomously innovate on AI research itself. For true novelty, human architects remain essential.

They’re not very good at code that has never been written before, maybe it’s one way to put it, which is what we’re trying to achieve when we’re building these models.

——————————————————————————–

4. For True Intelligence, Perfect Memory Is a Bug, Not a Feature

One of an LLM’s most powerful capabilities is its ability to memorize and regurgitate vast amounts of training data verbatim. In a deeply counter-intuitive turn, Karpathy argues this is not a strength but a fundamental weakness—and it’s a direct consequence of their nature as digital ghosts.

Because their entire existence is based on pattern-matching a static dataset, this powerful memory distracts the model from its more important task: learning the generalizable, abstract patterns within the data. It’s a crutch that prevents the model from being forced to develop deeper reasoning.

This stands in stark contrast to human cognition. Our famously imperfect memory is a feature, not a bug. Because we can’t remember everything perfectly, our brains are forced to compress information, find underlying patterns, and “see the forest for the trees.” This compression is the foundation of true understanding.

The implication is profound. Karpathy suggests future research must find ways to strip away rote knowledge to isolate what he calls the “cognitive core”—the pure algorithms of thought. He speculates this core could be much smaller, potentially only a billion parameters, if it weren’t so burdened by the need to memorize the entire internet.

We’re not actually that good at memorization, which is actually a feature. Because we’re not that good at memorization, we’re forced to find patterns in a more general sense. LLMs in comparison are extremely good at memorization… and it’s probably very distracting to them in a certain sense.

——————————————————————————–

Conclusion: The Long March of the Builder

Andrej Karpathy’s insights reveal a coherent picture from the engineering front lines. We are building digital “ghosts” whose nature—a hazy recollection of the internet—makes them prone to a perfect-yet-distracting memory. We then try to improve them with “terrible” learning methods that reward luck as much as skill. It’s no surprise, then, that these systems falter at true novelty.

His perspective is that of a practical builder: deeply optimistic about what AI can become, but soberly realistic about the immense challenges. Getting from a cool demo to a reliable product is a “march of nines,” where every step of improvement requires monumental effort. Fundamental discoveries about learning, reasoning, and intelligence are yet to be made.

As we continue to build these powerful new forms of intelligence, Karpathy’s insights push us to ask a crucial question: Are we merely trying to build a better tool, or are we trying to create a better thinker?

Reference Link: https://www.youtube.com/watch?v=lXUZvyajciY