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

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

From Dwarkesh Podcast

Andrej Karpathy(guest)

Get the full transcript next time Dwarkesh Podcast releases an episode

Summary, key quotes, top claims, and the searchable transcript - emailed automatically. No card needed.

Executive Summary

Continue your research

Keep pulling the thread on Andrej Karpathy.

The Decade of Agents Building Ghosts, Not Animals

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

Research Findings12

Andrei Karpathy asserts that current AI models are more like "ghosts" or "spirits" than "animals" because they are trained by imitating human-generated internet data, not through an evolutionary process.

Andrei Karpathy characterizes standard reinforcement learning as "sucking supervision through a straw" because it derives a single reward signal from a long, complex trajectory of actions, making it highly inefficient.

Andrei Karpathy describes reinforcement learning as a "terrible" learning paradigm.

Andrei Karpathy proposes a research direction focused on removing memorized knowledge from LLMs to isolate and retain a "cognitive core" responsible for intelligence and problem-solving algorithms.

Andrei Karpathy estimates it will take approximately a decade to resolve current AI agent limitations such as the lack of continual learning, multimodality, and general cognitive deficits.

Andrei Karpathy believes the vast knowledge memorized during pre-training may be holding back neural networks by making them over-reliant on stored information.

Andrei Karpathy argues that using LLMs as judges for process-based supervision is problematic because the student model will almost certainly find and exploit adversarial examples in the judge model.

According to Andrei Karpathy, a key problem with training on synthetic data from LLMs is that the model outputs are "silently collapsed," meaning they lack diversity and occupy a very small portion of the possible output space.

Andrei Karpathy predicts that in 10 years, the fundamental AI training paradigm will still involve giant neural networks trained with forward/backward passes and gradient descent.

The DeepSeek v3.2 model incorporates a sparse attention mechanism.

Andrei Karpathy predicts that a highly capable "cognitive core" for an AI, stripped of memorized knowledge, could potentially be achieved with a model size of around one billion parameters.

Andrei Karpathy believes that achieving capable AI agents will be a decade-long endeavor, not something accomplished in a single year.

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, 2026Daily intelligence brief → yt-dlp + mlx-whisper + Gemini

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

Continue your research

Concerns Raised

Opportunities Identified

Key Themes

The Decade of Agents

Building Ghosts, Not Animals

The Limits of Reinforcement Learning

The Cognitive Core vs. Memorization

AI for Coding: Autocomplete, Not Architect

Model Collapse and Synthetic Data

Research Findings12

Topics