The engineer who built the core attention kernel inside DeepSeek V4.1, one of the most powerful open weight AI models on Earth, is twenty-two years old. He turned down doctoral offers from CMU and Berkeley. He calculated that five years in academia could leave him stranded in a field that no longer existed by the time he emerged. He was right. It took twelve months.
Now he writes code that teaches machines to write code well enough to erase the value of his own expertise. He knows what he is doing. He keeps going because if he stops, someone else's machine gets there first, under terms he considers morally unconscionable.
This is a funeral held before the corpse stops breathing.
You don't even qualify for this crisis
The overwhelming majority of workers do not get to mourn the loss of craft. Work was never craft for them. It was a job, an instrument of survival dressed in office-casual. Watching a lifetime of mastery get flattened into training data for a neural network is a luxury form of despair. You need twenty years of expertise, and some love for the work itself, before you can lose it this way.
For most people, the bargain was simpler: the paycheck clears. So when an elite kernel engineer grieves the disappearance of the quiet afternoon, the hours spent on register allocation and warp scheduling, the private pleasure of shaving seven microseconds off a memory access, it can sound like aristocratic melancholy. Sometimes it is.
But the aristocrat gets hit first. If workers this specialized, in a field this technically demanding, are already writing their own eulogies, what exactly is the plan for everyone else?
"I Set a Target Before Bed. I Woke Up. It Was Done."
One infrastructure engineer working on the same open-source serving framework described the moment the floor dropped out. Using a frontier model to optimize DeepSeek V4.1's inference pipeline, he watched throughput rise from 30 tokens per second to more than 1,000 in just over a week. Ninety-five percent of the work was performed autonomously. The human contribution had collapsed to directional nudges. His description reads like a witness statement:
"From the moment I started using it, I felt the world was wrong."
He would set a performance target before going to sleep. By morning, it was hit. The gains held up under verification: no reward hacking, no hallucinated benchmarks. The model did the engineering while he slept.
Asked what advantage he still had over the machine, he did not cite knowledge, intuition or creativity. He said responsibility.
A human can still be held accountable when something goes wrong. In his account, the remaining human advantage is not superior knowledge or speed. It is simply that a person can be blamed.
That is a brutally thin moat.
The shit mountain cometh
If students use AI to complete every lab, assignment and systems-programming exercise from day one, they never develop engineering judgment: the instinct for when an abstraction is leaking, a design will buckle at scale, or elegant code is solving the wrong problem. A bad engineer without AI produces bad code slowly. Give the same engineer AI, and the output can become catastrophic at superhuman speed. The result is fragility at scale: invisible rot produced at industrial speed.
The same system attacks the training pipeline from both ends. It automates the master craftsman and deprives the next generation of the practice that would have made them one.
Who controls the compute controls the future
If artificial intelligence exceeds human cognition, power becomes decisive. It belongs to whoever owns the GPUs and controls access. That owner decides which version of machine intelligence everyone else is allowed to use.
Open model weights do not solve this. A 600-billion-parameter model without millions of dollars in infrastructure is a paperweight. "Open" may widen the oligopoly from two companies to twenty; for eight billion people, access remains somebody else's decision.
That logic explains both his choice of lab and the extremity of his language about one particular AI corporation. He frames its bid for dominance, in his own words, as civilizationally comparable to a fascist regime obtaining nuclear weapons before the Allies. For him, the contest is over whether frontier intelligence becomes something like a public utility or a private monopoly, with the gains shared broadly or concentrated into a permanent caste.
At twenty-two, he is already helping automate the skill that made him exceptional. Control of the intelligence he is helping build matters more than the fate of his job.
Who inherits the thinking?