Andrej Karpathy: Ten More Years to Artificial General Intelligence
Andrej Karpathy — AGI Is Still a Decade Away
Full Interview: Andrej Karpathy & Dwarkesh Patel (via Hacker News)
A deeply insightful 2 hour 25 minute discussion covering the future of AI, the definition of “agents,” and why AGI might be a decade away.
---
Karpathy’s Timeline on “The Year of Agents”
The conversation begins with Karpathy’s claim that “the year of agents” is more likely a decade away — in contrast to some near-term predictions, including my own acceptance of 2025 as the year of agents just yesterday.
Karpathy uses a different definition of “agents” than my preferred one:
> Think of it almost like hiring an employee or intern to work with you.
> Currently, they can’t do this kind of work. The blockers include:
> - Insufficient intelligence
> - Lack of multimodal capabilities
> - Inability to handle computer use autonomously
> - No continual learning — they can’t remember what you teach them
>
> Resolving these will take about a decade.
---
Key Challenges Before “True Agents” Arrive
Karpathy’s comments highlight the gulf between today’s AI assistants and the fully autonomous, adaptable agents envisioned by labs and researchers.
The necessary breakthroughs include:
- Reliable memory and continual learning
- Robust multimodal processing (vision, audio, text, etc.)
- Independent computer use
- Overcoming fundamental cognitive limitations
---
Practical AI Use Today — AiToEarn Example
Even if AGI-level agents are a decade out, platforms like AiToEarn官网 are enabling creators to leverage current AI capabilities today.
AiToEarn is an open-source, global AI content monetization platform that integrates:
- AI content generation
- Multi-platform publishing (Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X)
- Analytics
- Model ranking
This allows creators to:
- Publish across channels in one workflow
- Turn “imperfect” AI assistants into productive, revenue-generating partners
- Build audiences and monetize content right now
---
Ghosts vs Animals — A Useful Analogy for LLMs
Karpathy offers an evocative framing for LLMs: they are ghosts or spirits, not brains like animals or humans.
> Brains come from evolution. LLMs are trained by imitation of human data on the Internet.
>
> They are fully digital, ethereal entities — a different kind of intelligence.
>
> While we start from a different point than animals, it’s possible to make them more “animal-like” over time.
Read Karpathy’s blog post: Animals vs Ghosts.
---
When AI Agents Fall Short — The Nanochat Example
Dwarkesh asked about Karpathy’s tweet on Claude Code and Codex CLI shortcomings while building his nanochat project.
> AI models excel at boilerplate and common patterns found across the Internet.
>
> Nanochat required unique, nonstandard code architecture, where precision was vital.
>
> The models repeatedly misinterpreted the code, carrying excessive bias from typical coding patterns — which did not apply here.
---
Lessons for AI Deployment
Specialized, non-boilerplate projects magnify the weaknesses of current agents.
The key takeaway:
- AI performs best in structured, repetitive contexts
- Unique, complex workflows still require human oversight
Platforms like AiToEarn can help bridge this gap by offering:
- AI-assisted content creation for standardized tasks
- Centralized analytics
- Cross-platform publishing tools
- Monetization pathways without sacrificing creative control
---
Bottom line:
- AGI-level “employee-like” agents may be 10 years away
- Ghost-like LLMs are powerful now when embedded in well-designed workflows
- The winning strategy today is combining AI’s strengths with targeted human guidance
---
Do you want me to also create a summary table comparing Karpathy’s "true agents" vs current AI assistants? That could make the differences crystal clear.
