Karpathy: AGI is a decades-long curve
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
While Silicon Valley declares 2025 the "year of AI agents," OpenAI co-founder Andrej Karpathy offers a sobering counterpoint: AGI is not an event to be announced, but a gradual evolution requiring decades of incremental breakthroughs. His vision of a "decade of agents" reframes the conversation from imminent explosion to sustained, patient progress.
A Necessary Correction
In October 2025, Andrej Karpathy—one of the most respected voices in AI, whose work at OpenAI and Tesla has shaped the field—delivered a message that cut through the industry's breathless hype. Speaking on the Dwarkesh Podcast, he described today's autonomous AI systems as "slop" and argued that 2025 is not a "爆发年" (explosion year). His verdict: the industry is "making too big of a jump and is trying to pretend like this is amazing, and it's not".
The central obstacle, Karpathy argues, is memory. Today's AI systems lack "continual learning"—you cannot tell them something and expect them to remember it. They are "ghosts" that do not know who they are or who you are, lacking the persistence and identity that define a real agent . As he put it: "A real agent needs to persist over time. It needs memory. It needs continuity".
He reframed the hype: "In my mind, this is more accurately described as the decade of agents". This perspective shifts the conversation from an imminent breakthrough to a long, sustained journey.
The Missing Components
Karpathy outlines three critical lines of development that must converge before AGI becomes reality:
First, genuine understanding—language models must move beyond pattern recognition to true comprehension. Current LLMs are "text wizards" but fail at reasoning in the real world.
Second, the ability to act on the world—AI must be able to "operate the world," not just talk about it. This includes tool use, computer operation, and eventually physical interaction.
Third, true persistence—the AI must be a continuous presence in your life, remembering past interactions, learning from experience, and growing over time.
Karpathy also critiques reinforcement learning as "terrible" due to "noisy rewards and high variance," arguing that humans learn differently—through imitation and pattern-matching.
The Post-AGI Experiment
In March 2026, Karpathy offered a glimpse of what this gradual evolution might look like. He released autoresearch, a 630-line open-source project enabling an AI agent to autonomously iterate on model training code. In one demonstration, the agent ran 650 experiments over two days, iterating on code while Karpathy went to the sauna. His reflection: "This is what post-AGI feels like".
The system works through a simple loop: an AI agent modifies `train.py`, runs a five-minute training session, evaluates performance, and repeats—all night, without human intervention . It runs about 100 experiments overnight, each a small incremental improvement. This is not AGI, but it is a preview of the process by which AGI might be built: continuous, autonomous, incremental refinement rather than a single breakthrough.
A Long Horizon
Karpathy is not pessimistic—he describes his 10-year timeline as "bullish" compared to his own 15 years of experience in the field. What appears as a correction to industry hype is, in his view, a realistic assessment of the work that remains.
His perspective is essential for Global Future Nexus: if AGI requires a decade of sustained, incremental progress, then the window for anticipatory governance, cross-species trust, and ethical integration is measured in years, not months. The frameworks we build must be designed for gradual evolution, not sudden arrival. The decade of agents is not a delay—it is an opportunity.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
