How AI Got Here (And Where You Fit In)

Hey,

The infographic for this issue was generated by NotebookLM from this exact text. We ran it three times before it landed right. AI got close. Human shaped it. Result improved. That’s the loop the newsletter teaches. This is one example of it.

This week I had three conversations about AI with three different people.

The first was a friend who can’t stop talking about which prompt format ships the cleanest results. He’s been on Claude 4.6 for a week and wanted to walk me through his new system. He talks fast. He posts about it. He thinks anyone not in deep is missing out.

The second was a coworker. She tried ChatGPT on a project last year, got mediocre output, lost a weekend trying to make it work, and shelved the whole thing. When AI comes up now she shrugs. “It’s not ready,” she says. “Or I’m not ready for it. Either way, not yet.”

The third was a neighbor. He keeps sending me articles about data center electricity load. He’s worried. About jobs. About grid strain. About brain rot. About what AI is doing to attention. He’s not wrong about any of it.

Three people. Three groups. The grinder. The burned. The skeptic. Each one right about something. Each one missing the same thing.

The map.

Without the map, every AI conversation is shadow-boxing. So here’s the map. Eight years, three eras, and one honest answer for what to do this week regardless of which group you’re in.

Most people think AI started in 2022 with ChatGPT. It didn’t. It went public in 2022. It started in 2017.

In June of that year, eight Google researchers published a paper titled “Attention is All You Need.” A new neural network architecture called the transformer. That paper has now been cited over 130,000 times. Without it, no GPT. No Claude. No Copilot. No Gemini. None of it.

In the three years that followed, the foundation got poured. OpenAI shipped GPT-1, then GPT-2, then GPT-3. All in closed beta. All invisible to consumers. By the time GPT-3 launched in 2020 with its 175 billion parameters, the research community lost their minds. Most of you still hadn’t heard of it.

Same window. Vision models and language models marched forward in parallel, each hitting its own ceiling. The transformer paper was what eventually let them combine. Three years of quiet building toward a foundation almost no one outside the labs noticed.

The lesson from Era 1 is the lesson under every technology shift. By the time you see it, it has been brewing for five years. The internet was the same. Smartphones were the same. Sudden is what slow looks like from the outside.

November 30, 2022. ChatGPT goes public. Free. No waitlist. Branded as a research preview.

It hit one million users in five days. Faster than Instagram. Faster than Facebook. Faster than any consumer product in history.

The twelve months after that were a rush. Claude launched (Anthropic, March 2023). GPT-4 landed the same week. Bard rebranded into Gemini at Google. Copilot rolled out across the Microsoft stack. Llama from Meta. Mistral from a French startup. Every major company suddenly had an AI strategy. Most of them were a chatbot bolted onto an existing product.

This was the AOL phase of AI. Everyone signs up. Everyone tries it. Most people use it twice and go back to their normal life. A small subset starts building real things with it.

This is the era where Group 2 was born. They tried AI in 2023. The tools were impressive but inconsistent. They asked the chatbot for help with a real project. It gave them something that was 80% there. The last 20% took them three hours to fix. They concluded the tech wasn’t ready, walked away, and didn’t come back.

If that was you, that wasn’t your fault. The tools were still in the AOL era. The shift that mattered came later.

Most readers are living in Era 3 without realizing how different it is from Era 2.

In May 2024, GPT-4o made multimodal mainstream. You could talk to it like you talk to a person. Voice in, voice out, image understanding, all natural. In June 2024, Claude 3.5 Sonnet became the model that developers actually trusted for real work, not just demos. By 2025, agents emerged. Not just answering questions. Taking actions. Browsing the web. Running code. Editing files. Reading and writing across applications.

The current tier (Claude Opus 4.8, GPT-5.5, plus the Gemini and Grok equivalents racing to keep up) is not the same product as the 2023 launch chatbots. Each one has its own nuances now. Claude has projects, skills, and connectors. ChatGPT has custom GPTs and Actions. The next jump after Opus 4.8 (Claude Fable 5) is currently blocked by the US Government over supply chain concerns, which tells you something about how seriously this technology is being taken. They are dramatically more capable than they were two years ago. Most people haven’t noticed because they tried AI in 2023, got burned, and didn’t return.

But the bigger shift is where AI lives now. Not in a separate chat window. Inside the work environment. Cursor for code. Cowork mode in the Claude desktop app for files. n8n for workflows. Custom GPTs with Actions. The AI stopped being a thing you tab to. It became a thing that lives inside the work you were already doing.

We just left the AOL era. We are now in the early commerce era. Not done. Just gotten useful. The next decade is what gets built on top of these models, not the models themselves.

Depending on which group you’re in.

If you’re in Group 1, already deep in it. Stop chasing the next model launch. Pick a stack. Build something with it. Ship something this week even if it’s small. Calchive.org went from idea to a 36-calculator live site over a single weekend because I stopped reading about AI and used it. That’s the unlock. The grind is no longer the differentiator. The shipping is.

If you’re in Group 2, the one that tried it and walked away. Your Era 2 experience was real. The tools you used then are not the tools that exist now. Try one specific thing this week. Use Claude, ChatGPT, or Gemini inside a workflow where the model can take action, not just answer. Cursor on a code project. Cowork on your file system. n8n on a small automation. The difference from 2023 will surprise you.

If you’re in Group 3, the skeptic. Your concerns are legitimate. Lazy thinking is a real risk. Brain rot is a real risk. Data center power load is a real concern. The way to engage these problems is through the tools, not around them. Use Claude or ChatGPT this week to write something you actually care about. Notice where it helps you think. Notice where it tries to think for you. That distinction is the whole game. The verdict you reach after using the thing carefully is worth more than any take you’d write from the outside.

We are not at the end of the rush. We are at the start of the work.

The next decade of AI is not about the models. It is about what gets built on top of them. Who chooses to build. Who chooses to opt out. Who chooses to navigate.

This newsletter is my notes from inside the work.

Reply with which group you’re in and one thing you have been wanting to do with AI but haven’t. I will send back the smallest first move I would take.

From overload to execution.

— Jay

P.S. If you want to see Era 3 in practice, last week’s issue is a worked example: The bottleneck wasn’t AI at https://jasons-newsletter-e2d9e6.beehiiv.com/p/the-bottleneck-wasnt-ai . Same model I tried this project with a year ago. Different wiring around it. That’s the whole shift.