The AI Bubble: Echoes of the Dot-Com Era

Published: August 5, 2026 Category: ai

Andrej Karpathy—a titan in the field who helped launch OpenAI, directed AI efforts at Tesla, and recently transitioned to Anthropic to refine Claude—tends to approach Large Language Models (LLMs) with the same pragmatic simplicity that most developers apply to basic CRUD applications.

He is the architect behind nanoGPT and llm.c, the streamlined, transparent codebases that serve as the gold standard for anyone trying to grasp the end-to-end mechanics of a language model.

The "Paradigm" Conflict

When Anthropic integrated Claude into Slack, Karpathy hailed it as a "new paradigm," marking the third fundamental shift in human-AI interaction. He argued that AI is no longer just a destination (a website) or a tool (an app), but rather:

"a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans."

While Karpathy concluded that the system is "awesome" and functional, the public reaction was far more cynical. One specific response from @MatthewFoxAF on June 24, 2026, encapsulated the divide:

"This is a new paradigm brother it's a slack integration I miss the old andrej"

A Divided Landscape

This friction—between the visionary and the skeptic—is the defining characteristic of our current moment. Everything is in flux.

Executive Level: Leadership mandates "more AI" without providing a framework for responsible implementation.
Engineering Level: Developers rely on an old-school intuition: $\text{Risk} \propto \text{Lack of Understanding}$ . They are hesitant to deploy systems they cannot fully explain.
Corporate Level: Companies announce mass layoffs citing AI efficiency in the same press releases where they celebrate record earnings.

Beneath this lies a visceral anxiety: a widening chasm between those who own the intelligence and those whose livelihoods are being automated by it, with no clear bridge between the two.

The Velocity of Hype

The pace of change is dizzying. In a single week, OpenClaw garnered 100,000 GitHub stars and triggered a global shortage of Mac Minis. We see a rapid succession of "correct" methodologies:

~~Context Engineering~~ $\rightarrow$ ~~Harness Engineering~~ $\rightarrow$ Graph Engineering

Teams are shipping code at a velocity that outpaces their own comprehension. Even as someone who reviews all code before deployment, I can feel the erosion of understanding. Writing the code used to be the process of understanding it; now, that link is broken.

Lessons from the 1990s

When faced with these contradictions, our instinct is to label one side as "deluded." The critique "I miss the old andrej" suggests that if Karpathy sees a paradigm shift in a Slack bot, he must have lost his way. It assumes a binary: one perspective is true, and the other is false.

However, history suggests a more complex reality. Consider the dot-com crash:

Company	Short-term Result	Long-term Outcome
Pets.com	Total Collapse	Bankruptcy / Failure
Amazon	Stock dropped $>90\%$	Redefined global retail & logistics

Both the claim that the internet was "wildly overhyped" and the claim that it "changes everything" were simultaneously correct. The difference between the failure of Pets.com and the triumph of Amazon wasn't the technology—it was understanding.

Navigating the Noise

We are witnessing a transformational technology—an event that happens perhaps twice in a professional lifetime. I was a student during the last one, distracted by $\text{South Park}$ and $\text{Perfect Dark}$ . This time, I intend to understand the machine while it is still in motion.

My current objectives:

Test hypotheses against real-world outcomes.
Define my professional workflow while the choice still matters.
Resist the algorithmic push to be either "amazed" or "afraid."

The current discourse is a mess of debates:

Which model is superior?
Do open weights trump closed systems?
Panic: "We don't know what's happening, but we must act."

Some argue that model quality will eventually reach a point where these debates are moot because the AI will be perfect.

The Core Hypothesis

Here is my primary thesis, which I invite you to challenge:

$\text{Model Dependability} < \text{Required Threshold for Resolution}$

In plain English: Models will not become reliable enough, quickly enough, to solve these structural and professional dilemmas for us.

If you believe that near-flawless output is arriving within the next 12–24 months, you can stop reading now. The answers will reveal themselves, and this exploration will be unnecessary.

But if you agree that intelligent people can look at the same tool and see different worlds, then we must realize that our debates are based on unexamined assumptions. We are facing the same question that separated the survivors of the dot-com crash from the casualties—and this time, it applies to our own work.

Acknowledgments: Special thanks to Emma Bukacek and Tim Pote for their critical feedback on early drafts.