Karpathy Stirs Up a Storm Internally: Forcing Agents to Work is AI's Biggest Mistake! The Cutting Edge Isn't at OpenAI, It's in Your Hands.
One sentence poured a bucket of cold water over the entire Agent community.
Andrej Karpathy—now a core researcher on the pre-training team at Anthropic—recently dropped a bombshell during a live presentation for Agent developers, leaving the audience in stunned silence:
The biggest mistake in the current AI field is that people are in such a hurry to force Agents to work, without first truly understanding the underlying large models.
The clip was edited and thrown onto X, where it spread like wildfire in just a few days.
Because it precisely hit the nerve of the hottest, most crowded track that everyone is rushing into right now.
And the person saying this isn't an outsider throwing cold water; it's someone who has been through the trenches, reflecting on their own hard-earned lessons.
Lessons Learned with Real Money on the Line
Let's rewind to 2016.
Back then, Karpathy was working on a project at OpenAI called World of Bits. The goal sounded very "2026": Teach an Agent to use a keyboard and mouse to operate a computer, book flights, order food—basically, do your chores.
Sound familiar? That's almost the exact picture on the first slide of every Agent startup's pitch deck today.
The result? It didn't work out.
Karpathy put it bluntly: He, along with Tianlin Shi and Jim Fan, worked on it, frantically clicking on a few simple web pages, trying to book a flight or order some food. They even managed to publish a paper at ICML 2017.
The paper was titled "World of Bits: An Open-Domain Platform for Web-Based Agents"—a grand vision of a "world of bits" ultimately trapped on a few clunky web pages.
The technology wasn't ready. The only tool they had was reinforcement learning, and no matter how hard they hammered, it just wouldn't work.
Looking back, the truly correct approach back then would have been to completely forget about Agents and turn to working on language models instead.
Five years later, the toolbox is completely different—you folks building Agents today hardly use reinforcement learning at all. Back then, that was unimaginable.
Interestingly, Jim Fan, his co-author on that paper, is now a Senior Research Scientist at NVIDIA, creating explosive projects like Voyager and MineDojo, and winning an Outstanding Paper Award at NeurIPS.
A young intern from a 2016 "failed project" has become a top player in the AI Agent field a decade later.
But the path taken wasn't the one from 2016.
Demos are Easy; Building a Product Takes a Decade
Following this lesson, Karpathy offered three pieces of advice, each directly contradicting the current hype.
First, stop forcing your Agent to do everything. First, get the underlying model right.
When he joined Anthropic's pre-training team this past May, the first thing he wrote on X was: I believe the work on the LLM frontier will be particularly critical in the coming years.
The person who "invented" vibe coding, which Collins Dictionary named its Word of the Year, is now choosing to return to the most foundational pre-training research—this in itself is a "behavioral vote" against the Agent hype.
Second, Demos are easy. Turning it into a product takes a decade.
He cited two examples everyone knows: Self-driving cars. Anyone can make a demo of a car driving around the block, but turning it into a real product took a full decade. He personally experienced that marathon at Tesla.
The same goes for VR. Impressive demos were everywhere, yet taking it to a product was also a decade-long journey.
Agents fall into this exact category.
Extremely easy to imagine, extremely easy to demo, but extremely difficult to turn into a real product.
If you're really getting into this field, you need to be prepared to work on it for ten years, not think you've made it after a flashy demo.
Third, an Agent is not the product. Foundational capabilities are the product. Build a solid foundation, and Agents will naturally emerge.
These three sentences almost completely negate the current playbook of "slap a shell on it, stack an Agent, and release it ASAP."
Karpathy's message is clear: If the foundation isn't solid, the faster you build the building, the harder it will collapse.
Self-driving cars have already validated this for everyone over the past decade. There's no reason Agents should skip this lesson.
Learning from the Brain
After the lesson, Karpathy pivoted, diving headfirst into neuroscience for inspiration.
On stage, he fired off a series of questions: In an Agent, what is equivalent to the hippocampus, responsible for memory, indexing, and retrieval?
What corresponds to the basal ganglia, controlling action selection and execution? What is the thalamus, that "seat of consciousness" where multiple thoughts fight for the microphone?
A top AI researcher is saying: What we lack most right now in building digital life isn't fancier features, but reverence for the root question of "what intelligence even is."
He even specifically brought a copy of David Eagleman's neuroscience book, "Brain and Behavior: A Cognitive Neuroscience Perspective," and recommended it to everyone in attendance.
In his view, building Agents today deserves the same approach as in the early days of deep learning—back then, we stole the inspiration for artificial neural networks from the structure of a single neuron; now, we can absolutely go steal from the brain again.
The Real Bombshell was This Last Sentence
If the earlier part was throwing cold water, Karpathy ended by lighting a fire for the audience.
He said to the room full of independent developers and entrepreneurs:
Those truly at the cutting edge of Agent capabilities are you. Not OpenAI, not DeepMind. It's you.
This wasn't just polite talk. He gave a particularly poignant explanation:
For a big lab like OpenAI, training large-scale Transformer language models is indeed unparalleled—when a new Transformer training paper comes out, the reaction inside their Slack is often, "Oh, someone tried this two and a half years ago, we know exactly why it didn't work."
But when a new Agent paper pops up, everyone's reaction is: "Oh, that's really cool, really novel."
Why? Because no major lab has five years of accumulated experience in the Agent field.
The big labs are not at the edge of capability here. You—entrepreneurs, hackers—you are the ones standing at that edge.
The logic isn't hard to grasp.
The big labs have been running on the language model track for so many years, they've already stepped in every pothole and marked every detour. But Agents are a newly opened frontier; no one has a five-year head start. Everyone is almost on the same starting line.
At this point, flexible, daring, and fast-moving independent developers actually have a better chance of stumbling upon something new than the giant, hard-to-turn ships of the major players.
Back to That Initial Bombshell
The cold water Karpathy wants to pour isn't "don't build Agents," but "don't skip the fundamentals to build Agents."
He himself is the best footnote—the person who invented vibe coding and used Agents to great effect made his most important career move in 2026: Returning to pre-training, returning to the most fundamental lab of large models.
The fire he wants to light isn't to make people anxious, but to tell everyone struggling on the front lines: In this battle, you're not behind. You're right at the front.
The hype will always fade, and demos will eventually lose their luster.
But those who truly master the underlying models and are willing to dive deep into something for ten years are the ones who deserve to stand on the shore a decade from now.
References: https://x.com/0xCodila/status/2073544407643496771
This article is from the WeChat public account "新智元" (Xin Zhi Yuan), author: ASI启示录, editor: Solomon
