Andrej Karpathy’s sober AGI timeline: why 2025–2035 will be the decade of agents, not the year

AGI in 10 years and the decade of agents

Andrej Karpathy tempers AI hype with a 10-plus-year AGI horizon and argues 2025–2035 will be the decade, not the year, of agents. He highlights a gap between model capability and real-world tooling—scaffolding, memory, safety, and integration—that must close before agents truly scale. His roadmap favors agentic interaction, system prompt learning, and a “cognitive core” that prizes generalization over memorization.

Points clés

After a widely discussed podcast appearance, Andrej Karpathy clarified on X that AGI is likely 10+ years away, urging a more measured outlook.
He frames 2025–2035 as the “decade of agents,” countering claims that 2025 alone will deliver fully useful, economy-wide agent proliferation.
Citing OpenAI’s Operator, he argues digital agents will mature faster than physical robots because flipping bits is ~1,000x cheaper than moving atoms, with humans acting as high-level supervisors.
Despite rapid LLM progress since late 2022, he says significant grunt work remains: integration, scaffolding, sensors/actuators, societal alignment, safety, and security.
He describes a “model overhang” where core capabilities outpace the tooling, memory, and infrastructure needed to reliably extract value.
On learning, he contrasts “animals vs ghosts”: LLMs prepackage intelligence via next-token prediction, but true generalization (not memorization) is essential for AGI.
He critiques reinforcement learning as low signal-per-flop and noisy with outcome-based rewards, and instead is “long” on agentic interaction and alternative learning paradigms.
He promotes “system prompt learning,” where durable problem-solving notes live in prompts/memory; Anthropic’s Claude uses a ~17,000-word system message to guide stepwise reasoning.
His “cognitive core” vision favors smaller, always-on, multimodal models that trade encyclopedic memory for reasoning, tool use, personalization, and on-device adaptation—after a phase of scaling up, then pruning down.
He warns current agents overshoot model capability, prefers collaborative, auditable chunks of work, and built “Nano Chat” largely by hand; Elon Musk proposed a Grok 5 coding duel, which Karpathy declined in favor of collaboration.

À retenir

Practical takeaway: enjoy the demos, but don’t hand your job to an agent just yet unless you fancy “mountains of slop.” Start small—use agents in supervised, auditable chunks, keep receipts (system prompts and memory), and prioritize tools that boost your judgment rather than replace it. And if anyone promises AGI next Tuesday, smile politely, set a reminder for 2035, and carry on building.

Sources