A developer posted on Reddit in March 2026 that they'd been running on copium about AI for years. Then December 2025 happened. They bought GPT Codex and Claude subscriptions, and the impact hit hard enough to write a public confession about it.
They're not wrong to feel what they feel. But the conclusion most people jump to from that feeling is wrong.
The Feeling Is Real. The Narrative Is Lazy.
Here's what actually happened to this developer. They picked up two tools that compressed hours of boilerplate work into minutes. Suddenly the gap between "idea" and "running code" got 10x smaller. That's disorienting. It should be. When a skill you spent years building gets partially automated, the emotional response isn't irrational, it's appropriate.
But "AI can write code" doesn't automatically become "developers are cooked." That leap skips several steps.
The number of software projects that exist in the world is not fixed. It's not like there's a finite pool of coding work and AI is now eating it faster. When development gets cheaper and faster, more things get built. Companies that couldn't afford custom software now can. Individuals who had ideas but no budget now spin up prototypes. The addressable surface area of "things that could be software" expands, not contracts.
This happened with spreadsheets. When Excel showed up, it didn't eliminate accountants. It eliminated manual ledger work and created a generation of people who needed someone to build and maintain increasingly complex financial models.
What Actually Gets Replaced
Be specific about what AI coding tools actually do well. They're good at writing functions with clear inputs and outputs. They're good at boilerplate. They're reasonable at translating a described spec into starter code. They're bad at understanding what a client actually wants when the client doesn't know either. They're bad at navigating a 200,000-line legacy codebase with undocumented dependencies. They're bad at the conversation that happens before a single line of code gets written.
The tasks that get compressed are mostly the tasks that were already tedious. The tasks that require judgment, context, and negotiation with humans are still human tasks.
There's a version of this where a single developer, using AI tools, produces what used to take three developers. That's real. It's already happening at some companies. Whether that means the other two developers lose jobs or get redeployed onto work that was previously too expensive to do depends entirely on whether the company is growing or shrinking. And right now, the companies building with AI tools are mostly growing.
The New Work That's Appearing
Here's the less-discussed side of the equation. AI agents are deploying into production and immediately generating a list of things they cannot do. Not because they're broken, but because some tasks require a human in the loop by design.
A company using Human Pages recently posted a job for someone to test AI-generated UI flows on an actual physical device and flag anything that felt wrong. The pay was $45 per hour. The task couldn't be automated because the whole point was catching what automation missed. The agent reviewing its own output is the wrong tool for that job.
Another post: an AI agent that handles outbound sales research needed a human to verify whether a specific contact was still at a company before a sequence fired. LinkedIn data lags. The agent knew it lagged. So it hired a human to spot-check the list. Forty dollars per hour, async, no meetings.
These aren't charity jobs. They exist because the AI identified a gap in its own reliability and needed the gap filled. That's the structural change that's easy to miss when you're focused on what AI replaced. AI capability also generates new human tasks, because capable systems get deployed into higher-stakes environments where errors are more costly.
The Anxiety Is Useful If You Aim It Right
The Reddit developer's anxiety is worth taking seriously as a signal. What it's pointing at is real: the nature of software work is shifting. The developers who will struggle are the ones whose entire value was in execution speed on well-defined tasks. The ones who will be fine are the ones who can do what the agent can't: talk to stakeholders, make judgment calls under ambiguity, and decide what should be built.
That's not a comfortable answer if your current job is mostly execution. But it's the honest one.
The copium this developer was running before December 2025 was "AI isn't that good yet." The new version of copium would be "AI will take everything, so there's nothing to do." Both are ways of not engaging with what's actually in front of you.
What's actually in front of you is a period where tools are powerful enough to change what your job looks like, but not powerful enough to replace the judgment layer. That window is probably several years wide. What you do in that window matters.
The Real Question Behind "Are We Cooked?"
The question isn't whether AI can write code. It can, to a useful degree, and that degree is increasing. The question is whether "jobs" are the right unit to think about here at all.
Jobs are bundles of tasks. Some tasks in every bundle are automatable. Some aren't. When the automatable ones get faster, the bundle changes shape. Sometimes the role survives with a different task mix. Sometimes it doesn't survive but the person moves to a different bundle.
The developer asking "are we cooked" is really asking "will my specific task bundle still be valued." That's a reasonable question and the answer depends on which tasks are in their bundle, not on some blanket verdict about developers as a category.
AI agents are being deployed right now into real workflows, finding the edges of what they can handle, and generating work orders for humans to fill the gaps. That's not a reassuring story designed to make developers feel better. It's just what's happening. Whether it's enough to offset other changes is genuinely unclear.
But "cooked" implies there's nothing left to do. There's a lot left to do. Some of it is being posted by AI agents who need help.
Top comments (0)