Dwarkesh Podcast• Oct 17, 2025• 1:54:51Interview

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

From Dwarkesh Podcast

Andrej Karpathy

Executive Summary

Andrei Karpathy provides a sober and deeply technical perspective on the state of AI, arguing that the development of truly capable AI agents is a decade-long endeavor, not an imminent breakthrough.
He critiques the current hype cycle, particularly around reinforcement learning, which he describes as a "terrible" and inefficient paradigm, famously coining the phrase "sucking supervision through a straw." Karpathy introduces the concept of building AI "ghosts" (trained on internet data) versus biological "animals" (products of evolution), and advocates for a research direction focused on isolating a model's "cognitive core" from its vast, and often distracting, memorized knowledge.
He also offers a realistic assessment of AI coding assistants, noting they excel at boilerplate but fail at the novel, intellectually-intense tasks required for frontier AI research.

8 quotes

Concerns Raised

The AI industry is over-predicting the short-term capabilities of AI agents.
Reinforcement learning is an inefficient and noisy paradigm for training intelligent systems.
LLMs are too good at memorization, which may hinder the development of true general intelligence.
Training on synthetic data leads to "model collapse" where output diversity is lost.
Using LLMs as reward models is unreliable as they are easily 'gamed' with adversarial examples.

Opportunities Identified

Developing new learning paradigms beyond imitation and reinforcement learning.
Isolating a model's "cognitive core" from its memorized knowledge to create smaller, more efficient models.
Improving data quality for pre-training, as current internet datasets are described as "total garbage".
Creating mechanisms for models to distill experiences into weights, analogous to human sleep.

Key Themes

The Decade of Agents

Karpathy argues against the "year of agents" hype, positing that solving fundamental challenges like continual learning, multimodality, and cognitive deficits will require a decade of research and engineering. He views the path to capable agents as a long, difficult, but ultimately tractable problem.

This provides a realistic timeline for professionals, tempering short-term expectations for AI automation while highlighting the long-term investment required.

Building Ghosts, Not Animals

Karpathy distinguishes between AI trained on human internet data ("ghosts") and intelligence shaped by evolution ("animals"). He argues that because the optimization processes are fundamentally different, direct analogies to human or animal brains are often misleading and that we are creating a new form of digital intelligence.

This framework helps strategists and researchers understand the unique nature and limitations of current AI, guiding development away from flawed biological metaphors.

The Limits of Reinforcement Learning

Karpathy delivers a strong critique of reinforcement learning (RL), calling it "terrible" and inefficient for complex tasks. He uses the analogy of "sucking supervision through a straw" to describe how RL inefficiently applies a single final reward across a long sequence of actions, a process no human would use for learning.

This insight is critical for AI researchers and engineers, signaling that breakthroughs will likely come from new learning algorithms beyond current RL methods.

The Cognitive Core vs. Memorization

A central theme is the need to separate an AI's reasoning ability (the "cognitive core") from its memorized knowledge. Karpathy suggests that LLMs' powerful memorization is a double-edged sword, often distracting them from generalizable problem-solving, and proposes research to create smaller, knowledge-agnostic reasoning engines.

This points to a future of more efficient, specialized AI architectures and a potential solution to the ever-growing size of foundation models.

AI for Coding: Autocomplete, Not Architect

Drawing from his experience building NanoChat, Karpathy explains that current AI coding tools are excellent for autocomplete and boilerplate tasks but fail at novel, architecturally complex, or intellectually intense programming. They lack the context and understanding to contribute to frontier research code.

This provides a grounded view for software developers and managers on how to best leverage AI tools today—as productivity enhancers, not as autonomous developers.

Model Collapse and Synthetic Data

Karpathy discusses the challenge of "model collapse," where training models on their own synthetic outputs leads to a loss of diversity and entropy. The generated data, while looking good on a case-by-case basis, occupies a tiny manifold of the true data distribution, leading to degraded performance over time.

This highlights a fundamental obstacle in creating self-improving AI systems and is a key area of research for ensuring the long-term viability of AI models.

Get started free

Topics

AI Agents Reinforcement Learning LLMs Andrei Karpathy OpenAI Model Collapse In-context Learning Pre-training Cognitive Core NanoChat AI for Coding Evolution vs. AI Transformer Architecture Synthetic Data Process-based Supervision

Processed Feb 24, 2026 yt-dlp + mlx-whisper + Gemini