Andrej Karpathy — AGI is still a decade away

Karpathy's phrase 'the decade of agents' is meant as a correction to claims that agents will arrive fully this year. He already uses tools such as Claude and Codex and finds them impressive, but says they still need substantial work before they can function like dependable employees. He lists missing intelligence, multimodal ability, computer use and continual learning. His estimate of a decade is an intuition drawn from roughly 15 years watching AI predictions and working in research and industry; he calls the problems tractable but difficult rather than deriving a formal schedule.

The timeline is connected to a sequence of shifts in the field. Karpathy entered deep learning when neural networks were still a niche; AlexNet then made neural networks central, though systems remained trained for individual tasks such as image classification or translation. The next aspiration was a complete agent that could perceive, act and interact with the world. Atari reinforcement-learning experiments were an early attempt, but Karpathy now regards the field's multi-year emphasis on game-playing as a misstep relative to real-world knowledge work.

Karpathy's own OpenAI project tried to build an agent that operated web pages with a keyboard and mouse. He says it was far too early: an agent that randomly clicks and types receives sparse rewards, making learning inefficient and consuming large amounts of compute. What was missing was the representational power of a pretrained language model. In his account, the sequence matters: first learn broad representations from data, then build agents on top. Current agents are more capable, but he thinks parts of the full stack are still absent.

Dwarkesh raises the animal-learning counterexample: animals enter the world without language labels and learn to act from sensory experience. Karpathy is cautious about treating animals as a direct blueprint because evolution produces them through a different optimization process and provides substantial built-in hardware. A newborn zebra can run within minutes, which is not something it learned through ordinary reinforcement learning. Karpathy instead describes today's language models as digital entities trained by imitation of human-created internet data, a different kind of intelligence from an evolved animal.

The disagreement turns on what evolution contributes. Dwarkesh suggests evolution may have discovered a learning algorithm that then operates over a lifetime, rather than encoding every skill directly in DNA. Karpathy accepts that genes cannot store the detailed weights of a developed brain and agrees that some learning algorithms must be encoded. But he takes a practical stance: evolution is not a process he knows how to run, whereas imitation-based pretraining is available now. He calls pretraining a crude, technologically feasible route to a starting point from which later learning methods can be applied.

Karpathy argues that internet pretraining does more than store facts: exposure to patterns also seems to induce circuits that support in-context learning. But a large store of knowledge may sometimes make agents over-rely on familiar patterns and struggle outside the internet's data manifold. He therefore proposes a research goal of separating a model's cognitive core—the algorithms and strategies for reasoning—from some of its stored knowledge. The idea is not that knowledge has no value, but that too much memorized content may distract from general problem-solving.

The exchange examines whether in-context learning is itself a kind of gradient descent. Karpathy describes it as pattern completion within a token window; the patterns discovered during training are encoded in the model's weights, while adaptation occurs as the model processes the current context. He mentions research showing that a neural network can perform linear regression through in-context examples, with internal mechanisms resembling gradient descent. He does not claim the mechanism is settled, only that a small internal optimization process may be involved.

A key difference may be how much information is accessible at once. Dwarkesh compares the information stored in a model's weights after pretraining with the much larger amount represented in the KV cache for each token in context. Karpathy calls pretrained knowledge a hazy recollection, because trillions of tokens are compressed into billions of parameters, while context-window content is directly accessible like working memory. This helps explain why a model may answer roughly correctly about a book from memory but perform much better when given the chapter itself.