First they gave us a chatbot. Then they gave it eyes, ears, and a terminal. Now it opens PRs while we sleep.
I got into AI right as the chaos started — self-taught, refreshing the OpenAI blog like it was a live sports score. I watched every era of this ride in real time: the hype, the crashes, the benchmarks that flipped overnight, the "this changes everything" takes that aged worse than milk.
And the wildest part isn't any single model. It's that we went from "ask it a question" to "hand it the keyboard" in under four years. Most technologies take a generation to go from demo to infrastructure. This one did it between two iPhone releases.
Here's the whole timeline, era by era. Spoiler: it gets crazier as it goes.
Era 1 — The Shock: November 30, 2022
A random Wednesday. OpenAI drops a free chatbot with no product page, no press tour, just a text box.
ChatGPT hit 100 million users in two months — the fastest-growing consumer app in history at that point. Instagram took two and a half years to do the same.
What made it different? It wasn't just autocomplete. It was autocomplete with a personality — fine-tuned with reinforcement learning to sound like a helpful human instead of a search index. People used it to write poems, debug code, and argue about whether it had feelings.
Devs had two reactions, simultaneously:
- This is a toy.
- This is the end of my career.
Both were wrong. The real change was quieter: the interface won. Chat was the killer app, not the model. Everyone with a GPU and a research paper suddenly understood what "good" felt like.
2022 taught us to talk to AI. It took two more years to learn how to hand it the keyboard.
Era 2 — The Arms Race: 2023
Now the floodgates. GPT-4 arrived in March 2023 and passed the bar exam at the 90th percentile. It could read images, write legal briefs, and absolutely humiliate your code review.
Google finally shipped — first Bard (we don't talk about Bard), then Gemini, then a 1-million-token context window that made reading your entire codebase in one shot a real thing.
Anthropic countered with Claude 3, the model that devs actually started trusting with their repos.
The leak that changed everything
But the biggest story of 2023 wasn't a company. It was a leak.
Meta's LLaMA weights got out in February 2023, and the open-source world lost its mind. Within weeks, the community had fine-tuned it into Alpaca, Vicuna, and a hundred cousins. By July, Llama 2 was officially open for commercial use. By September, Mistral 7B was beating models 10x its size.
The MoE architecture (Mixture of Experts) made it worse — Mixtral proved you could get frontier-adjacent quality from a fraction of the parameters. Suddenly, running a capable LLM on your laptop wasn't a dream. It was a Saturday project.
The lesson of 2023: closed models race, open models catch up, and they always will.
Era 3 — The Reasoning Era: late 2024
Here's where things get genuinely weird.
Every model before this was a sprinter: ask, answer, done. Then OpenAI shipped o1-preview in September 2024 — a model that thinks before it answers. Chain-of-thought at inference time. Test-time compute.
The numbers were absurd. On the AIME math competition, GPT-4o solved 12% of problems. o1 solved 74% on the first try, 93% with sampling. It wasn't a better sprinter. It was a different sport.
Google answered with Gemini 2.0 Flash Thinking. Anthropic shipped Claude 3.7 with hybrid "extended thinking." The entire industry pivoted from "bigger models" to "models that think longer."
The DeepSeek grenade
And then DeepSeek dropped the grenade.
DeepSeek V3 trained a 671B-parameter MoE model for ~$5.5 million — a number that made every lab with a $100M training run choke on their coffee. Then R1 in January 2025 matched OpenAI's reasoning models with open weights, triggering one of the biggest single-day market crashes in tech history.
The narrative died overnight: "reasoning is a closed-lab secret." Nope. It's an RL recipe, and now everyone has the recipe.
Era 4 — The Agent Era: 2024 → now
This is the era you're living in.
It started in November 2024, when Anthropic open-sourced MCP — the Model Context Protocol. Think of it as USB-C for AI: one standard plug so any model can reach your files, your database, your browser, your tools. It exploded faster than anyone expected. By 2026 it's the de facto interface — Claude Code, Cursor, Codex CLI, Cline, CrewAI all speak it natively.
The model isn't the product. The harness is.
Then came the agent harnesses — and this is the part most people still don't fully get.
The model isn't the product anymore. The harness is: the loop that gives a model a terminal, a plan, a way to run tests, and permission to iterate until the job is done. Claude Code. OpenAI Codex. Cursor. These aren't chatbots with extra steps. They're junior engineers that don't sleep, don't complain, and open a PR while you're in a meeting.
Here's the mental model: a raw model is a brain in a jar. A harness gives it hands, eyes, and a memory — and suddenly the brain can actually do things. Same brain, wildly different outcome. That's why two teams using the same model can produce completely different results. One is prompting a chatbot. The other is running a harness.
The software stack changed to match. AGENTS.md files now sit next to your README so agents know your conventions. CI pipelines assume agents will push to them. Entire companies run on agents doing the grunt work and humans doing the review.
The best model stopped mattering. The best harness won.
Where We Are Now (August 2026)
Multi-polar, and moving fast.
OpenAI's GPT-5.x line leads on reasoning-heavy benchmarks (GPT-5.4-Pro tops GPQA Diamond at 94.4%). Claude Opus 4.7 owns software engineering — best in the world on SWE-Bench Verified (87.6%), which is why it powers most serious agent workflows. Gemini 3.1 Pro doubled its predecessor's reasoning score, and OpenAI reportedly declared an internal "Code Red" when Gemini 3 landed in late 2025.
The open side isn't folding either. DeepSeek, Qwen, Kimi, GLM — open-weight models keep matching frontier quality at a fraction of the cost. OpenAI even shipped its first open-weight models since GPT-2.
The takeaway of 2026: the frontier is crowded, the harness is the moat, and the rate of change is still accelerating.
What This Means for You (the dev)
Every era created a new skill gap. Here's the pattern: the people who learned the new interface early printed, everyone who dismissed it as a toy got left behind.
• Era 1 — learning to prompt well mattered. It still does, but it's table stakes now.
• Era 2 — knowing which model for which job became a real skill. Still true.
• Era 3 — understanding reasoning modes, thinking budgets, and when to let the model think longer. This one's underrated.
• Era 4 — the big one: knowing how to wire agents. MCP servers, harness config, context management, review loops. That's the skill with the salary attached to it.
The devs winning right now aren't the ones who memorized the best model. They're the ones who built the best system around a model. That's the whole game in 2026.
The One-Line History
The Takeaway
Every era of this ride looked like a toy on day one and infrastructure by day ninety. ChatGPT was "just a chatbot." o1 was "just a math tutor." Claude Code was "just a CLI wrapper."
None of them were.
We're not in an AI bubble. We're in the middle of the fastest platform shift in software history — and the people winning are the ones who stopped asking "what can it do?" and started asking "what can I build with it?"
The chatbot era is over. The keyboard's been handed over. The question now is what you're going to build with the other 10 hours a day.
Follow me on X (@aj_livess) for build-in-public content, agent workflows, and the next post in this timeline — and drop your take in the comments. Which era broke your brain?
Top comments (0)