OpenAI co-founder says he is in a ‘state of psychosis’ — Here's why Andrej Karpathy is nervous about the future

Andrej Karpathy says he has not written a line of code since December and now delegates roughly 80% of coding tasks to AI agents versus previously writing ~80% himself (a reversal to ~20% coding). He describes broad adoption of agentic AI (referred to as OpenClaw) for tasks from calendar management to emails and home automation—his ‘Dobby’ agent controls sound, lighting, security, shades, HVAC, pool and spa via WhatsApp. Karpathy reports being in a “state of psychosis” trying to grasp the technology’s limits and is anxious about not being at the forefront of the AI revolution.

Analysis

Agentic AI reaching the point where senior engineers delegate the majority of routine coding is a structural productivity inflection, not a one-off convenience. Over 12–36 months this should compress marginal developer hours per feature while raising the value of systems integration, prompt engineering, and model orchestration skills — shifting compensation and hiring from bulk code production to high-signal oversight and productization roles.

Immediate winners are firms that sell the heavy-lift: data-center GPUs, inference hooks, and managed orchestration (cloud + model-tooling). Second-order beneficiaries include cybersecurity vendors and workflow automation platforms because agentic stacks increase privileged API surface area and enterprise automation spend; losers in the medium term are low-margin outsourced dev services and training-heavy junior hiring pipelines that rely on scale rather than expertise.

Risk profile: the tail risks that could reverse adoption include a high-profile data breach or regulatory clampdown on agentic access to third-party systems, and a plateau in model capabilities that forces a reversion to hand-coded solutions. Watch catalysts over weeks (usage metrics, early enterprise Copilot/agent ARR), months (data-center booking trends, job-posting composition), and quarters (guidance changes that reprice cloud/compute spend).

Tactically, trade around measurable signals: 1) NVDA/AMD server booking cadence and channel inventory; 2) cloud vendors’ Copilot/agent monetization metrics; 3) security telemetry showing agent-driven incidents. These will provide concrete 1–3 quarter entry points for both equity and options exposure rather than posture based on hype cycles alone.

Key Decisions for Investors

• Long NVDA (9–12 months): buy a defined-risk call spread (e.g., buy 12-month IV and sell a higher strike) sized to 2–3% of portfolio. Rationale: capture accelerated data-center GPU demand if enterprise agent adoption continues; target 2.5–4x upside vs premium, max loss = premium paid. Enter on <10% pullback or on next strong server-booking print.
• Long MSFT (6–12 months): accumulate 1.5–2.5% position ahead of Copilot/agent monetization cadence. Rationale: cloud + tooling capture; set stop-loss at 12% below entry or trim on revenue guide miss. Reward: asymmetric 30–50% upside if enterprise ARR acceleration shows sustained spending shift.
• Long cybersecurity (PANW or CRWD, 3–9 months): initiate 1–2% positions — add if telemetry shows agent-driven breaches/incidents. Rationale: agents expand attack surface; expect 20–40% re-rating on visible security spend. Risk: 100–150bps drawdown if incident count remains low and sentiment fades.
• Pair trade (12–24 months): long cloud/tooling (MSFT/NVDA) vs short traditional outsourcing (example: CTSH) sized dollar-neutral 2–3% net exposure. Rationale: secular margin divergence as automation reduces low-skill dev demand. Risk: outsourcing firms pivot or win large-ticket managed-AI contracts; hedge by sizing conservatively and monitoring contract announcements.