If your feed is anything like mine, it’s drowning in AI hype. The noise is deafening.

But when Andrej Karpathy- one of the most important voices in AI- offers a “reality check,” it’s time to listen. He just shared a fascinating perspective that cuts right through the hype, explaining why he’s extremely optimistic about AI but 5-10 times more pessimistic than the current “Twitter timeline.”

Here’s the breakdown of his core insights.

1. 🤖 Goodbye “Year of Agents,” Hello “Decade of Agents” (2025–2035)

We all agree: AI agents are the next big thing. But Karpathy believes the window for their real, valuable, and widespread proliferation isn’t a year.1 It’s a decade.

Why the slower timeline? The “Scaffolding Problem.”

The core intelligence of today’s models is incredible. But the infrastructure, tooling, and memory required to actually get value from that intelligence are lagging far behind. We still have immense “grunt work,” integration, and safety challenges to solve before agents are running our lives.

2. 👻 LLMs Are “Ghosts,” Not “Animals”

This is a brilliant analogy.

• An animal (like a zebra) is born prepackaged with intelligence by evolution. It can walk almost immediately.

• An LLM prepackages its intelligence by predicting the next token over the entire internet. This makes it more like a “ghost or spirit.”

The critical takeaway: LLMs rely too much on memorization.2 True Artificial General Intelligence (AGI) requires generalization – the ability to learn new things on the fly, not just recall what it has already seen.3

3. 🧠 The Push for a Smarter “Cognitive Core”

To beat this memorization trap, Karpathy is interested in the “cognitive core.”4

This means intentionally stripping away memory from models to force them to reason better. Imagine a highly capable, smaller model that sacrifices encyclopedic knowledge to boost its core problem-solving ability.

He also highlighted better learning methods, like “System Prompt Learning,” where a model “takes notes” on successful problem-solving strategies to reuse later- learning how to think, not just what to remember.

4. 🤝 Collaboration Over Full Autonomy

If you dream of handing an agent a huge task and checking back in a week, Karpathy advises caution.

He critiques the agent industry for “overshooting the tooling” and warns that full autonomy could create “mountains of slop” (e.g., a thousand lines of code you can’t possibly review).5

What he wants instead is collaboration:

• Working in smaller, manageable chunks he can verify.

• Having the LLM explain the code it’s writing and prove its work is correct.

• Focusing on “agentic interactions” (letting agents experiment in digital playgrounds) to learn, rather than relying on noisy and inefficient Reinforcement Learning (RL).

###

We’ve made incredible strides. But the path to AGI isn’t about one magic algorithm. It requires massive integration work, a fundamental shift from memorization to generalization, and new ways for us to work withAI.

It’s time to roll up our sleeves for the decade ahead!

What’s your take on this timeline? Is the industry getting ahead of itself, or is Karpathy being too cautious? Let’s discuss in the comments!