73M Views on the Warning. Nobody Got the Next Step.

Subscribing to ChatGPT for $20/month is step zero. The real divide is between people who type prompts and people who engineer context.

Matt Shumer's post hit 73 million views in two days. His diagnosis is accurate. His prescription will leave most people stuck.

Shumer, the CEO of OthersideAI, published a 5,000-word essay called "Something Big Is Happening" that became the most-shared piece of writing about AI this year. He argued that current models crossed a real threshold, that most people have no idea what AI can do now, and that the disruption coming will dwarf anything we've seen. He compared this moment to February 2020, before COVID hit.

The diagnosis landed because it's true. GPT-5.3 Codex and Opus 4.6 are a genuine capability jump. People who tried ChatGPT in 2023 and dismissed it were right to. They'd be wrong to dismiss what exists now.

But Shumer's prescription, the part where he tells 73 million viewers what to actually do, fits in a single paragraph: subscribe to Claude or ChatGPT for $20/month, select the best model available, and spend an hour a day experimenting.

That advice was useful in February 2024. In February 2026, it's a starting line disguised as a finish line.

The Gym Membership Problem

Telling someone to "subscribe to ChatGPT and experiment" is like telling someone to buy a gym membership and try different machines. Technically correct. Practically useless for most people.

Gyms don't transform bodies. Training programs do. The membership gets you through the door. The program tells you what to do when you're inside, which muscles to target, in what order, at what weight, on what schedule.

AI subscriptions work the same way. The tool isn't the bottleneck. The system is.

I know because I lived this. I've been a productivity obsessive since 2017, always trying to systematize my work and the projects I run. When AI tools started shipping in 2022, I went all in. ChatGPT, Claude, Gemini, Perplexity, vertical agents, workflow automators. I kept hitting the same wall: constant copy-pasting between apps, re-explaining my context in every conversation, output that sounded generic no matter how specific my prompts got, and none of the true AI autonomy that everyone in the space kept promising. In May 2025, I stopped looking for a better tool and started building a system. That changed everything.

In the ten weeks since I launched that system to the world, 79 clients have signed up. The pattern repeats: every one of them had a ChatGPT subscription before they came to me. Most had tried Claude too. They weren't lacking access. They were lacking architecture.

Sure, some had tried ChatGPT Projects or Claude Projects, and those help. You get persistent instructions, some file uploads, a bit of memory. But the context is still siloed in one chat window, disconnected from your actual work infrastructure. It's a better gym locker, not a training program.

The people getting real results from AI aren't the ones who experiment the most. They're the ones who built a system once and then ran it daily.

What the Divide Actually Looks Like

Shumer frames the coming divide as "people who use AI" vs. "people who don't." That framing was accurate a year ago. The divide has already shifted.

The real gap is between people who type prompts into a generic chatbot and people who taught AI their context.

Here's what the difference looks like in practice:

With context architecture: "Draft a follow-up for the prospect I spoke with Tuesday." AI checks my CRM, reads our email thread, applies my communication style guide, references our actual conversation, and produces a draft that sounds like me with a specific next step. Time: 2 minutes.

Without context architecture: Same request. I paste in the prospect's details, explain my communication style, provide conversation context, and specify what I want the follow-up to achieve. Time: 12 minutes. And tomorrow I start over.

The first version compounds. The second resets to zero every session. Over weeks and months, that gap becomes the difference between someone who needs AI and someone AI needs.

Experimentation Is How You Start, Systems Are How You Win

Shumer's advice to "try to get AI to do something new every day" is solid onboarding guidance. But it keeps you in tourist mode. You sample the local cuisine without ever learning to cook.

The people I've watched get the most from AI followed a different arc. They experimented briefly, found one workflow where AI saved real time, then engineered that workflow into something repeatable. They stopped asking "what can AI do?" and started asking "what does AI need to know about MY work to do this reliably?"

It shifts the focus from prompting to context engineering: deciding what information the AI receives, from which sources, in what order, at what level of detail.

One client, a management consultant, used to spend 90 minutes preparing for each client call. Reviewing notes, scanning emails, pulling relevant frameworks. After we built her system, that prep takes under 10 minutes. The AI reads her meeting notes, cross-references her CRM, pulls the client's recent communications, and produces a briefing document in her preferred format. She didn't get there by experimenting. She got there by building.

Another client runs a small agency. His team generates content across 12 accounts. Before, every piece started from scratch because the context lived in people's heads. Now, the AI knows each client's brand voice, past content, performance data, and editorial calendar. It drafts at 80% quality. His team edits to 100%. Output tripled. Experimentation didn't produce that. Architecture did.

The Implementation Chasm

Shumer's post paints a seductive picture: "I describe what I want built, in plain English, and it just appears." For code, that's increasingly accurate. OpenAI said GPT-5.3 Codex was "instrumental in creating itself," used by the team to debug its own training and manage its own deployment. That's a real milestone.

But for knowledge work, generation was never the bottleneck. The bottleneck was always knowing what to generate, for whom, against what standards.

AI can write a proposal in 30 seconds. But which proposal? For which client? In whose voice? Referencing which past work? Meeting which specific requirements the client mentioned in an email three weeks ago?

Without context, AI generates plausible-sounding output that requires heavy editing. With context, it generates useful output that requires light refinement. The gap between "plausible" and "useful" is where most people give up. They see the plausible output, compare it to the hype, and conclude AI doesn't work for their job.

It works. It just doesn't work without context.

This is the implementation chasm Shumer's post doesn't address. The gap between "AI can do impressive things in demos" and "AI does useful things in my daily work." Crossing that chasm requires engineering, not experimentation.

What Context Architecture Actually Looks Like

The concept that made the biggest difference in my work isn't a fancy prompt technique or a specific tool. It's what I've started calling context architecture: treating the entire information environment around your AI as the thing you design, not just the question you ask.

Most people think context means "paste more stuff into the chat." It doesn't. Context architecture has four layers, each operating on a different timescale:

Layer	What It Includes	How Often It Changes
Knowledge	SOPs, style guides, decision frameworks, domain terminology	Write once, compounds forever
Project	Client briefs, architecture decisions, ongoing documentation	Weekly
Task	Immediate context for the current request	Every interaction
Tool	Live data connections (CRM, email, calendar via MCP servers)	As needed

Most people skip the knowledge layer entirely. That's why their output feels generic. A consultant's methodology framework. An agency's brand voice guidelines. A manager's decision-making criteria. These documents are the foundation everything else builds on, and they're the layer that compounds the most.

The tool layer matters too. I run 12 MCP servers that connect Claude to my live work infrastructure: email, calendar, CRM, payments, LinkedIn. But MCP is one layer, not the whole architecture. Plenty of my clients get massive results before they ever touch MCP, just by building their knowledge and project layers properly.

Prompt engineering asks "how do I phrase this?" Context engineering asks "what does the model need to know?"

You don't need 12 MCP servers to start. You don't even need to code. You need a document that captures your professional standards. A folder structure that feeds your templates and SOPs into the conversation. Start with the knowledge layer. Everything else builds on it.

Fear Without a Flashlight

Seventy-three million people viewed Shumer's post. Most of them closed the tab feeling more anxious than before. The ones who already use AI felt validated. The ones who don't felt afraid. Almost nobody got a concrete path forward.

This is the pattern with viral AI content: accurate alarm, vague prescription. The diagnosis generates engagement. The prescription generates dependency. "Follow me, I'll tell you which model to use" is a content strategy, not a transformation plan.

The real fear shouldn't be about AI itself. It should be about building your work life on a foundation AI can replicate without knowing anything about you. If a generic chatbot with zero context about your clients, your standards, and your domain can produce work indistinguishable from yours, the problem isn't AI. The problem is that your work didn't have enough context-dependent value in it.

The people who benefit most are the ones whose work requires deep context: knowledge of specific clients, institutional memory, professional judgment built over years. The person who encodes their own context into an AI system doesn't get replaced by it. They get amplified by it.

What to Do This Week

Shumer is right that something big is happening. He's right that the capability jump is real and that most people underestimate it. He's right that early movers have an advantage.

Where I part ways is on what "moving early" means.

Moving early means picking your highest-value repeating task and building a system where AI handles it with your context, your standards, your data. Not experimenting in a generic chatbox. Not following someone's model recommendations. Building.

If you write proposals, create a template with your voice, your case studies, and your pricing structure, then teach AI to draft against it.

If you manage clients, set up a system where AI can access your notes, your communications, and your project history.

If you create content, build a context file that captures your style, your audience, your past work, and your editorial standards.

One system. One workflow. Run it for 30 days. That's worth more than a year of casual experimentation.

The people who will own the next two years aren't the ones who subscribed first. They're the ones who built first.

Iwo Szapar is a context architect who helps knowledge workers build AI systems that compound instead of reset. After years of systematizing his own workflows and testing dozens of AI tools since 2022, he built the Second Brain architecture described in this essay. Seventy-nine clients have deployed it in the ten weeks since launch.

73 Million Views on the Warning. Almost Nobody Got the Next Step.