---
type: "article"
title: "Shortwave: They Just Don’t Work (Yet)"
summary: "Karpathy reminds us why agentic AI still has a long way to go."
newsletter: "Signals from Our Futures Past"
newsletter_handle: "mamund"
newsletter_url: "https://usecommune.com/n/mamund"
author: "Mike Amundsen (@mamund)"
published: "2026-01-30T13:30:38.000Z"
canonical_url: "https://usecommune.com/n/mamund/a/shortwave-they-just-don-t-work-yet"
markdown_url: "https://usecommune.com/n/mamund/a/shortwave-they-just-don-t-work-yet.md"
chat_url: "https://usecommune.com/n/mamund/a/shortwave-they-just-don-t-work-yet/chat"
source_url: "https://mamund.substack.com/p/shortwave-they-just-dont-work-yet"
body_source: "imported"
likes: 0
replies: 0
body_words: 267
---

# Shortwave: They Just Don’t Work (Yet)

## Tuning In

AI researcher [Andrej Karpathy](https://en.wikipedia.org/wiki/Andrej_Karpathy), formerly of OpenAI and Tesla, thinks the hype around “agentic AI” is getting ahead of itself.

[![](https://substackcdn.com/image/fetch/$s_!90v2!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcc8baa3e-8a4a-4f28-9676-7eedeaf78d1a_6000x4000.jpeg)](https://substackcdn.com/image/fetch/$s_!90v2!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcc8baa3e-8a4a-4f28-9676-7eedeaf78d1a_6000x4000.jpeg)

In an October 2025 *[Decoder](https://the-decoder.com/)* interview, he told podcaster [Dwarkesh Patel](https://www.youtube.com/@DwarkeshPatel) we’re not entering the *year* of agents but the *decade* of agents. The reason? Today’s models still lack memory, multimodal reasoning, and reliable task execution. “They just don’t work,” he said flatly.

## Signal Strength

Karpathy’s view cuts through the optimism driving today’s “agent economy.” He sees value in narrow, autocomplete-style use cases but not yet in autonomous problem-solving. The training data itself, he argues, is “total garbage”; a web of fragments and noise that rewards memorization over understanding.

His proposed fix is evolutionary, not revolutionary: better curation, better data, better architectures, better hardware. No single breakthrough—just many small, coordinated ones. “I feel like the problems are tractable,” he says, “but if I average it out, it just feels like a decade to me.”

## Return Signal

It’s a timely reminder that the next leap won’t come from another model drop. Instead, it’ll come from re-examining what and how we teach these systems to learn. Progress may be steady, not spectacular. As Karpathy puts it, the real work ahead isn’t making AIs *act* like us—it’s teaching them to *understand* what they’re doing.

## Logging the Frequency

> “I feel like the problems are tractable, they’re surmountable, but they’re still difficult. If I just average it out, it just feels like a decade to me.” — Andrej Karpathy

Source: *[The Decoder: “Andrej Karpathy says agentic AI is years away from matching industry hype”](https://the-decoder.com/ai-researcher-andrej-karpathy-says-agentic-ai-is-years-away-from-matching-industry-hype/)*

***

## Discussion

No replies yet.