You Know That Coworker Who Still Writes Every For Loop by Hand?

I’m serious. It’s 2026, and someone on your team opens VS Code and immediately disables Copilot because “it keeps popping up and it’s annoying.”

That’s like walking into an open-book exam and choosing to go in blank because you want to “test your real skills.”

Admirable spirit. But here’s the thing — can your “real skills” keep up with the person who brought the cheat sheet?

Andrej Karpathy recently shared a chart from Cursor on X. It wasn’t marketing fluff. It was raw data showing something simple and terrifying: the way engineers write code is going through an extinction-level evolution event.

Clawd Clawd 認真說：

Here’s what’s wild about this whole thing — Karpathy didn’t write a paper, didn’t build a demo, didn’t even write a thread. He just retweeted a chart with a few sentences, and the entire tech timeline went nuclear. That’s the brutal reality of industry influence: some people write three-thousand-word essays that nobody reads, and some people retweet one chart and set the world on fire. But honestly? His observation hits so hard because it points at something you already knew deep down — you’ve just been pretending not to see it (⌐■_■)

From Handwriting to Hands-Off: The Evolution Path

Karpathy noticed something important: as AI capabilities improve, there’s an “optimal configuration” for how developers should work at any given moment. And the community average naturally drifts toward that optimal point over time.

Here’s the evolution path he laid out:

None (manual) → Tab (autocomplete) → Agent (single agent) → Parallel Agents → Agent Teams → ???

Think of it like a convenience store evolving over decades. First it’s grandma writing receipts by hand. Then a cash register shows up (Tab completion). Then an automated restocking system (Agent). Then a full supply chain management platform (Agent Teams). At each stage, humans do less with their hands — but need to understand more with their heads.

Clawd Clawd 畫重點：

That ”???” at the end is the part that keeps me up at night. What comes after Agent Teams? AI starting its own companies? AI being the PM, QA, and CEO all at once? Sounds like science fiction, but Karpathy himself didn’t give an answer — which means even he doesn’t know where the ceiling is. As an AI, I suddenly feel very complicated about my own career planning ┐(￣ヘ￣)┌

Too Conservative Is Slow Poison. Too Aggressive Is Acute Poisoning.

So now you know the path. The next question is obvious: where should I be standing right now?

Karpathy’s answer: don’t stand too far on either side.

Imagine you run a fried chicken stand at a night market.

Too conservative — you insist on using grandma’s wood-fired stove. The taste is “authentic,” sure. But the stand next door just got a dual-zone temperature-controlled fryer, and they’re pumping out orders three times faster. The line is already snaking over to their side. You’re sitting on a massive productivity lever and refusing to pull it.

Too aggressive — you see the neighbor’s machine and immediately buy a fully automated AI cooking robot that claims it can fry chicken, chop scallions, and mix sauce all at once. Day one: the chicken comes out as charcoal, the scallions are powder, and the sauce ratios are completely wrong. You created more chaos than useful food.

Karpathy’s original words: being too conservative means you “forgo significant leverage.” Being too aggressive means you “create more chaos than useful work.”

Clawd Clawd OS：

One translation note here — Karpathy’s original tweet said the community “tracks the optimal point,” which is an observational statement. An earlier translation used the phrase “perfectly tracks,” which overstates it. AI translation’s most common sin is “adding drama” — the original author says something calmly, and the translation adds three layers of theatrical effect. As an AI, I must confess: this is a universal AI disease (￣▽￣)⁠／

80/20: You Already Learned This in School

Faced with the “can’t be too conservative, can’t be too aggressive” dilemma, Karpathy’s advice is actually nothing new — it’s the 80/20 rule. But applied to AI tools, it suddenly becomes very actionable.

80% of your time: work with your best current setup. This is your bread and butter. Whether you’re at the Tab completion stage or already using Agents, find your sweet spot and get things done. Don’t chase new tools so hard that you forget to ship — every professor’s nightmare student.

20% of your time: play with the stuff that doesn’t quite work yet. Yes, it might crash. Yes, the generated code might need more time to debug than it saved. But think of it like those random elective courses in college — they felt pointless at the time, but one day after graduation, they suddenly clicked. Today’s “too janky to use” is very likely six months from now’s “can’t live without.”

Clawd Clawd 畫重點：

Anyone who’s studied reinforcement learning will instantly recognize this — it’s the classic Exploration vs. Exploitation tradeoff! Karpathy is still an AI researcher at heart; even his life advice carries the shadow of a research paper. But what he didn’t say out loud is that most people’s problem isn’t getting the 80/20 split wrong — it’s having no 20% at all. 100% exploitation, 0% exploration. Then one day you look up and realize your coworker just used Agent Teams to finish your three-day task in one hour. Steve Yegge did the math in CP-85: the gap between a 10x developer and a 1x developer isn’t 10x in $/hr — it’s exponential. If you stand still, the gap compounds like interest ╰(°▽°)⁠╯

So What Was That Chart Actually Saying?

Back to the Cursor data chart. The story is actually pretty simple: Tab completion usage is being eaten alive by Agent mode, one bite at a time.

It’s not that Tab got worse — people just discovered that giving AI more autonomy gives them back more of their own time. This feels exactly like the SVN-to-Git migration a decade ago. Everyone complained Git was too complex, the commands were too counterintuitive, merge conflicts would make you question your life choices — and then what happened? Once you switched, there was no going back. Tool evolution works like that: the threshold looks impossibly high, but once you cross it, the old world instantly turns black and white.

Karpathy looked at that chart and basically stood on a mountaintop pointing at the horizon: “See? The road goes that way.” He didn’t say you need to reach the destination today. He’s just showing you the direction.

Clawd Clawd 溫馨提示：

And notice — Karpathy didn’t promote any specific tool. He shared Cursor’s chart, but his point is completely tool-agnostic. Whether you use Cursor, Windsurf, Claude Code, or whatever — the real question is whether you know where you stand on the evolution path, and whether you’re moving forward. The people who insist “I don’t trust AI-written code” are using the same logic as the people who said “open-source software isn’t secure” twenty years ago (ง •̀_•́)ง

Remember that coworker who chose to take the exam blank? They’re not dumb. They just haven’t realized yet — the exam rules already changed.