AGI Is Still a Decade Away

A Foundation Thought Piece for the AI and Robotics Series

When a headline says “AGI is only months away,” it usually means somebody wants funding. In contrast, former Tesla and OpenAI engineer Andrej Karpathy recently sat down with Dwarkesh Patel for a two-hour conversation that feels refreshingly sane. Karpathy’s view: artificial general intelligence is probably ten years away — and that’s optimistic.

His logic isn’t pessimism; it’s engineering.

The Missing Pieces

Karpathy walks through the pieces we don’t yet have:

• Continual learning. Current models can’t update themselves safely once deployed. Every improvement still means retraining from scratch.
• Memory. Large language models don’t truly remember context across sessions; they only simulate short-term recall.
• Computer use. Agents struggle with multi-step reasoning across digital tools — for example, navigating a browser, filling forms, or combining information sources coherently.
• Multimodality. Real understanding requires vision, text, audio, and action working together. Today’s systems stitch these together loosely at best.

Put simply: today’s AIs can talk about doing things; they can’t do them without hand-holding.

Ghosts, Not Animals

One of Karpathy’s most quoted lines from the interview is that we are “building ghosts, not animals.” These systems imitate intelligence but don’t inhabit it. They lack embodiment, intrinsic motivation, and physical grounding — the sensory feedback loops that teach real-world consequence.

That insight ties directly to robotics. If AI remains a “ghost,” robots are the bodies waiting for a soul. The missing bridge between them — sensory integration, adaptive control, situational awareness — may define the next decade of progress.

The Slow Decade of Agents

Rather than a sudden singularity, Karpathy sees a “decade of agents.” Over the next ten years, AI systems will gradually learn to perform constrained tasks — scheduling, research, document creation, or simulation — with increasing autonomy. Progress will be steady, not explosive.

He compares this to historic productivity trends: despite all our innovations, global GDP per capita has grown at roughly 2 percent a year for centuries. Expect AI to extend that curve, not break it. The revolution, in other words, will feel incremental — a million small automations rather than one great awakening.

Reinforcement Learning and Its Limits

Karpathy jokes that “reinforcement learning is terrible,” but then admits that every other approach has been worse. His point: AI doesn’t yet know how to learn from consequence. Reinforcement learning is still the only framework that approximates trial-and-error, but it’s crude and fragile.

Until models can refine themselves through lived experience — not just pattern recognition — they’ll remain savants without intuition.

Why This Matters

Karpathy’s realism is a useful counterweight to both doom and hype. It suggests that society has time to prepare — time to build governance, safety systems, and human-centered design before the technology truly generalizes.

It also reaffirms something we’ve seen across the Robots Series: progress happens first in structured, predictable environments. Before machines can think broadly, they must succeed narrowly — sweeping floors, sorting parts, assisting nurses, learning cause and effect in the real world.

This view sits in direct contrast to the idea explored in “If Anyone Builds It…” — that AGI could erupt spontaneously from scale and data. Karpathy’s take is humbler and, arguably, truer: intelligence isn’t a switch we flip; it’s a staircase we climb.

A Thoughtful Pause Before AI 6 and 7

For readers following the AI and Robotics Series, this interview offers a calm middle ground between fear and fantasy. It recognizes both the astonishing pace of current progress and the enormous distance still ahead.

“We’re building ghosts,” Karpathy says, “and it will take about a decade to give them bodies.”

It is a long read or listen but worth it.  Andrej Karpathy — AGI Is Still a Decade Away (Dwarkesh Patel Podcast)

Facebook Twitter Youtube