I read Matt Shumer’s “Something Big Is Happening,” and one line stuck with me: the labs made a deliberate choice to make AI great at writing code first.
Source: https://shumer.dev/something-big-is-happening
The “coding-first” bet wasn’t about engineers. It was about compounding.
Why coding came first (and why that matters to you)
Yes, building AI requires a lot of code. If your models can help you write and maintain that code, you move faster.
But it’s deeper than that.
Coding is measurable. It’s full of tight feedback loops. It’s an endless stream of real tasks with clear “works or doesn’t work” outcomes. That makes it the perfect place to push capability forward quickly.
And in the real world, that bet shows up in usage. In OpenRouter’s State of AI 2025 study (based on OpenRouter platform traffic, which is naturally developer-skewed), programming became the “most consistently expanding” category. They report programming queries going from roughly ~11% of token volume in early 2025 to over 50% in late 2025 (Figure 19).
Source (PDF): https://openrouter.ai/assets/State-of-AI.pdf
Once AI gets good at turning language into working systems, it stops being “a tool for engineers.” It becomes a tool for anyone who uses a computer to think, write, decide, analyze, plan, communicate, and ship work.
If your job happens on a screen, you’re in the blast radius of the upside.
“But I tried AI and it wasn’t that good”
I hear this constantly, and I get it. It used to be true.
The bigger truth now is that most people are judging today’s AI based on yesterday’s AI. Or on a free tier. Or on one attempt with a vague prompt.
That’s not a moral failing. It’s just what happens when tools improve faster than our mental models update.
The models and workflows people are using daily at the sharp end feel different. Not because they’re magical. Because the cost of iteration has collapsed.
The real shift: become AI-native (role doesn’t matter)
I think there’s a simple baseline habit that’s going to separate people:
Before you start anything, ask:
How can AI help me do this better, faster, or differently?
That applies to:
- engineers writing code
- product people writing specs
- managers writing feedback and plans
- sales writing outreach and proposals
- analysts turning messy data into a narrative
- marketers iterating on positioning
- support teams writing macros and troubleshooting steps
- anyone staring at a blank page, a spreadsheet, or a backlog
This isn’t about using AI sometimes. It’s about treating it like a default collaborator.
A practical way to work (tomorrow morning)
If you want a simple playbook, try this.
1) Start with context and intent, not output
Start in your AI tool of choice (ChatGPT, Claude, Gemini) and write 5–10 sentences that capture:
- the goal
- the constraints
- the audience
- what good looks like
2) Ask for 2–3 options
Ask for three approaches with tradeoffs.
This is how you get leverage without handing over judgment.
3) Make it do the first draft
Have it draft the email, outline the doc, propose the plan, generate the checklist, write the code, create the slide structure.
The win is not that the first draft is perfect.
The win is that you can produce and iterate far more than you could before.
4) You become the editor and decider
Your job shifts from producing raw material to shaping it: priorities, tone, correctness, risk, alignment.
5) Lock it down with verification
The output is only useful if you can trust it.
- For writing: ask it to self-critique, find weak claims, and tighten structure.
- For analysis: ask it to show assumptions and sensitivity.
- For work decisions: ask for risks, failure modes, and what you’d measure.
A note on longer tasks
Right now, AI can feel almost unreal on short, well-scoped tasks.
Longer tasks are still harder. The failure mode I see is less about raw intelligence and more about stamina: maintaining context, staying aligned, not wandering, not making a single subtle mistake that cascades.
But it’s also getting better fast. In my own recent experiments, I’ve had great outcomes with longer tasks too. The trick is usually to structure the work into checkpoints: define “done,” require intermediate artifacts, and force verification along the way. Most importantly pay for the upgraded accounts and use the latest models.
If you want a good external reference point on how reliability changes as tasks get longer, METR’s time-horizon work is worth reading:
https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
The ego trap
I’m still talking to smart people who believe: I can do this better myself.
Sometimes that’s true. But increasingly, it’s the wrong comparison.
The comparison isn’t AI versus me.
It’s me with AI versus me without AI.
And once you accept that, you stop asking whether AI can do everything. You start asking what you can do when the boring parts get cheap.
Pay for the leverage
If AI saves you even one hour a month, it’s already worth more than the subscription for most knowledge work.
Also, the best model changes quickly. The best tool yesterday may not be the best tool today. The most valuable skill is the ability to evaluate and switch without drama.
What happens next
I’m not going to make a bunch of predictions here. I do feel confident saying this:
The window where most people aren’t using this seriously yet is still open.
That window will close.
If you want an unfair advantage, build the habit now. If you manage a team, normalize it now. Make it safe to experiment. Share workflows. Teach people how to verify and review.
The only constant is change. The future is going to be wild.