Andrej Karpathy says humans are now the bottleneck in AI research with easy-to-measure results

2026-03-22 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Andrej Karpathy asserts that humans are now the bottleneck in AI research, particularly in areas with objective metrics, advocating for autonomous agents to take over tasks like hyperparameter tuning, as demonstrated by an agent outperforming his manual GPT-2 optimizations. He describes a "state of AI psychosis" where agents enable "macro actions" and significantly increase "token throughput" in software development, transforming workflows to delegation rather than direct coding. Beyond coding, Karpathy implemented a home automation system, "Dobby the elf claw," showcasing agents' ability to unify disparate smart home devices through natural language and APIs, suggesting a future where agents drive tool usage and refactor industries towards API-first designs. He envisions "auto research" where agents autonomously optimize models and even meta-optimize research processes, but cautions about the current "jaggedness" of AI capabilities, which excel in verifiable tasks but struggle with nuanced, less measurable domains. Karpathy also discusses the future of open-source AI, its increasing proximity to frontier models, and the long-term potential of robotics, emphasizing the interface between digital and physical worlds.

Key takeaway

Andrej Karpathy argues human intuition is now the bottleneck in AI research, demonstrating autonomous "auto research" agents can find overlooked GPT-2 hyperparameter tunings (e.g., weight decay, Adam betas) that he, with decades of experience, missed. This enables significantly more efficient model optimization and broader agentic applications like home automation. However, it also highlights challenges in non-verifiable domains and necessitates rethinking research abstractions for an "agent-first" web.

Topics

• AI Agents
• Autonomous Research
• Large Language Models
• AI/ML Optimization
• Home Automation

Code references

• karpathy/minGPT

Best for: AI Scientist, Research Scientist, AI Researcher, Machine Learning Engineer, AI Architect

Related on AIssential

• AgentFloor: How Far Up the tool use Ladder Can Small Open-Weight Models Go? — Small open-weight models can handle routine agentic tool use, reserving large models for complex planning.
• Researchers let AI Agents Optimize LLM Reasoning and Cut Tokens by 70% — AI agents can autonomously optimize LLM reasoning strategies, significantly reducing token usage while maintaining accuracy.
• 😸 Anthropic and OpenAI Both Dropped Their Best AI Models On The Same Day — Leading AI developers are advancing models towards autonomous job execution, emphasizing either broad utility or deep specialization.
• Evidence on AI R&D Progress from NanoGPT — Cumulative progress on public AI R&D challenges offers insights into human and agent contributions and their evolving nature.
• GPT-5.2 Can't Identify a Serial Killer & Was The Year of Agents A Lie? EP99.28-5.2 — GPT-5.2's verbose, over-tuned responses and agentic failures highlight a gap in practical AI application compared to competitors.
• The Sequence AI of the Week #826: Sleep While it Computes: Inside Karpathy’s AutoResearch — Human limitations, like sleep and slow communication, bottleneck AI research progress.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.