Karpathy's "AI Psychosis": When Code No Longer Needs to Be Handwritten

*Structured summary of a long-form Chinese forum post analyzing Andrej Karpathy's shift to agent-driven programming.*

Key points

• A paradigm shift around December 2024: Karpathy's workflow flipped from ~80% hand-written code to ~80% delegated to AI agents. He attributes the jump to new-generation agents (Claude, OpenAI Codex) crossing a "consistency/coherence threshold," becoming reliably integrable into daily work.
• Barely coding since: By early 2026 he stated he had "probably not typed a line of code since December." His daily work became communicating intent to agents—up to 16 hours a day—acting as specifier and quality gatekeeper rather than implementer.
• "AI psychosis": His term for the dual experience of liberation (unprecedented personal productivity, the "sauna" metaphor of delegating and walking away) and anxiety (skill deprecation, role confusion, competitive pressure). His December 26, 2025 tweet—"I've never felt more behind as a programmer"—drew 22,000+ likes and 3.6M views.
• "Vibe coding": He coined the term for a new abstraction level where humans focus on *what* they want, *why*, and *how to verify it*, delegating implementation details to machines.

Technical practices

• Multi-agent orchestration: Top engineers now run ~10 parallel codebase sessions (the Peter Steinberger pattern), requiring systems thinking, parallel tracking, and rapid context switching.
• Dobby (OpenClaw/Clawbot): A personal home-automation agent that replaced six separate apps. With no documentation provided, it auto-discovered Sonos speakers via LAN scanning, probed ports (e.g., UPnP on port 1400), and reverse-engineered undocumented APIs—then integrated lighting, HVAC, pool/spa, and security into one natural-language-controlled hub.
• AGENTS.md: An open standard now present in 60,000+ open-source projects, encoding project context, coding conventions, verification commands (single-file type-check/test to save token/compute budgets), permission boundaries, and multi-agent orchestration rules.
• AutoResearch: An autonomous research-agent system driven by a ~120-line Program.md spec. In one experiment it made 110 commits over 12 hours while Karpathy slept, lowering validation loss from 0.862415 to 0.858039 with no added training time.
• microGPT case: Agents could not distill a neural net to his ~200–243-line pure-Python GPT implementation—human "obsession" with essential, minimal solutions remains a durable human edge.

Social and professional implications

• Industry data: GitHub Copilot contributes ~46% of code in enabled files; Accenture measured 55% faster task completion; blogger Theo reports 70–90% AI-generated code in some teams.
• Human value shifts upstream/downstream: toward requirements definition, architecture, spec-driven development, and design-level code review; execution becomes cheap.
• "Token anxiety" and throughput: Token throughput becomes a new productivity metric; efficiency gaps between developers who optimize prompting, context management, and model selection widen.
• Jevons paradox: Cheaper software development stimulates demand—like ATMs increasing bank branch employment—suggesting AI expands rather than shrinks the software economy (a view also voiced by Microsoft CEO Satya Nadella).
• Education: Traditional syntax/algorithm training loses relevance; skills like problem decomposition, AI collaboration, and critical evaluation rise. Karpathy's Eureka Labs envisions AI handling knowledge delivery while human teachers focus on intuition, motivation, and personalized judgment.

Philosophical outlook

• "Verifiable vs. unverifiable": Karpathy's framework—"everything unverifiable is still human; everything verifiable belongs or will belong to machines"—re-anchors human value in goal-setting, values, aesthetics, and ethics.
• Human–machine trajectory: from tools to partners to eventual symbiosis, with open questions on cognitive outsourcing, responsibility, and augmentation vs. singularity.
• Post-AGI work as gym: Like exercising despite machines, future work may be driven by intrinsic motivation rather than economic necessity.

*Source: zhichai.net forum post; claims reflect the original author's synthesis of Karpathy's interviews and social media statements.*