Andrej Karpathy just dropped one of the most important observations about AI in 2026: there are now two fundamentally different groups of AI users, and they’re talking past each other completely.

His tweet (3.6M views and counting) describes a growing chasm that anyone working with AI needs to understand.

Group 1: “AI Is a Toy”

This group tried ChatGPT --- usually the free tier --- sometime last year. They saw hallucinations, laughed at viral videos of voice mode fumbling simple questions, and concluded that AI is overhyped.

They’re not wrong about what they experienced. Free-tier models from 2025 are genuinely limited. But here’s the problem: they froze their mental model of AI at that point and never updated it.

As Karpathy puts it, these old and deprecated models don’t reflect the capabilities of current frontier models.

Group 2: “AI Psychosis”

This group uses state-of-the-art agentic models --- Claude Code, OpenAI Codex --- professionally, in technical domains. They pay $200/month. They hand these models a computer terminal and watch them solve problems that would normally take days or weeks of work.

Karpathy says this group is experiencing what he calls “AI Psychosis” --- not because they’re delusional, but because the improvements have been so staggering that it’s hard to communicate what they’re seeing to anyone who hasn’t experienced it firsthand.

Why the Gap Exists

Karpathy identifies two structural reasons why AI capability is advancing unevenly:

1. Reinforcement learning works best with verifiable rewards.

Tasks like coding, math, and research have clear success criteria --- does the code compile? Do the tests pass? Is the proof correct? These domains are naturally suited to RL training, where the model gets concrete feedback on whether it succeeded.

Tasks like writing, advice, and conversation are much harder to evaluate objectively, so they improve more slowly.

2. B2B value drives resource allocation.

The biggest revenue opportunities are in technical/professional domains. That’s where AI companies focus their best teams. Consumer-facing features like voice mode get less investment relative to the B2B products that generate the most revenue.

The Translation Problem

The result is a bizarre disconnect. It is simultaneously true that:

A free AI voice assistant will fumble basic questions in your Instagram reels
A paid agentic AI will spend an hour coherently restructuring an entire codebase

Both of these things are happening in 2026. But people in Group 1 only see the first. People in Group 2 only see the second. And when they try to talk to each other about “AI,” they’re describing completely different technologies.

How to Cross the Divide

If you’re in Group 1, the path to Group 2 isn’t about spending $200/month on a different chatbot. It’s about changing how you use AI --- from casual Q&A to structured knowledge work.

Here’s what separates the two groups in practice:

Group 1 behavior:

Opens ChatGPT, asks a question, reads the answer
No persistent context between sessions
AI starts from zero every time
Judges AI by its worst failure

Group 2 behavior:

Builds knowledge bases that AI can reference
Gives AI access to project files, documentation, and research
Uses AI as a colleague with memory, not a search engine
Judges AI by its best capability

The key insight: the quality of AI output is directly proportional to the quality of context you give it.

Building Context: The Missing Layer

This is where most people get stuck. They hear “give AI better context” and think it means writing longer prompts or spending more time explaining things. That’s the brute-force approach. The smarter approach is building a persistent knowledge layer that AI can access automatically.

Here’s what that looks like in practice:

Capture web research as structured Markdown --- Instead of bookmarking pages or copying snippets into a doc, convert the full content to clean Markdown. This preserves the information in a format any AI tool can ingest.
Organize into searchable knowledge bases --- Group your saved content by project, topic, or research area. This gives AI the ability to find relevant context without you having to remember what you saved.
Connect AI to your knowledge --- Tools like MCP (Model Context Protocol) let Claude search and reference your saved content directly. When you ask a question, Claude checks your knowledge base first, grounding its answers in your curated research rather than generic training data.

This is the workflow that Save enables. Every webpage you save becomes a Markdown file in your local knowledge base. Save Vault’s built-in MCP server connects it to Claude. Your AI assistant now has access to everything you’ve read and researched.

The Compound Effect

The gap between Group 1 and Group 2 isn’t just about which model you use. It’s about the accumulated context you’ve built over time.

A developer who has saved six months of documentation, Stack Overflow answers, and architecture articles has an AI assistant that understands their codebase, their stack, and their specific problems. A casual user who opens ChatGPT fresh every time has a generic tool that knows nothing about their work.

This is why Karpathy’s observation matters beyond the AI hype cycle. The capability gap isn’t closing --- it’s widening. And the dividing line isn’t technical skill or budget. It’s whether you’re building a structured knowledge practice around AI or treating it as a glorified search engine.

Starting Today

You don’t need to spend $200/month to start crossing the divide. You need to start capturing and organizing the knowledge you already encounter every day:

Install Save and start converting useful web pages to Markdown as you browse
Create knowledge bases for your main areas of work or interest
Connect Claude via Save Vault’s MCP server so your AI assistant can reference your saved content
Be consistent --- save 2-3 pages per day and watch your knowledge base compound

Within a month, you’ll have a personal knowledge layer that makes every AI interaction fundamentally better. That’s the difference between Group 1 and Group 2 --- and it starts with how you capture knowledge, not which model you pay for.

Save converts any webpage to clean Markdown with one click and stores it in your local knowledge base. With Save Vault’s MCP server, Claude searches your saved content before answering. Try Save for free.