The manifesto needed the engineer it mocked
In Proton’s security Slack channel, an engineer asked for a few days to pentest newly written code before it went live. The CEO replied that this was “bullshit and loser thinking,” and told the channel to stop looking for ways to be slower. 404 Media published the exchange on 29 October, alongside the engineering manifesto the company’s CTO had circulated internally, drawing on eight current employees, one former one, Slack chats and meeting recordings.
The memo deserves better than the reply it produced. Its core move is to flip the usual framing: the job is no longer writing code, and the model is not there to help you write better code faster — your job is to help the model write better code faster. Much of that is right. If a working CRUD endpoint takes seconds, typing is not where the value sits. What is left is deciding what to build, writing a spec a model can follow, cutting work into checkable pieces, and knowing when the output is wrong.
The strongest part is the engine. A repeated failure should trigger an investigation and a fix: if the agent keeps running the wrong test command, the fix is a line in the repo’s context file, not a correction you retype tomorrow. If it keeps misreading a module, the module needs documentation or a cleaner interface.
That engine runs entirely on reports.
The manifesto even names why a person stays in the loop — a human is far less prone to hallucination, and to being led astray by bad or malicious input. The engineer asking for the pentest was that person, doing exactly that job. The reply put a price on the next report. Employees told 404 Media they feel unable to raise security concerns, and in that room a repeated failure gets quietly worked around rather than investigated.
Online, the bluntness was read as the point. Peter Steinberger, now at OpenAI and the author of the OpenClaw agent, sympathised with the loss of identity while noting that engineers had already had two years to adapt. Lauren Tan, a principal engineer at SpaceX, wrote that the time for mourning coding was long over. David Heinemeier Hansson went furthest: the real AI psychosis, he posted, is thinking that “humans still need to review every line.”
Every line is the wrong unit, and he is right that defending it is hopeless. Uniform review taxes code whose failures are not uniform. A generated list view can go through automated checks and nothing is lost. The narrower claim is the one that costs money: a person reads the diffs where trust lives.
Two kinds of slowness look identical from outside. One is reluctance to change, and it costs the company time it never gets back. The other is a check whose absence is invisible until it isn’t. Both arrive as someone asking for a few more days. What separates them is what happens if you skip it: a merge that lands a week later, or an auth gap live for a week. A leader who can’t tell them apart clears both off the schedule.
The same blindness decided what got counted. Proton installed software on employee machines to measure the share of AI-generated code, and the figure reported from meeting audio was about ninety percent for October. That counts the input. A team can push that number up and ship more defects in the same week. Revert rate, incidents, time to ship — none of those were the number on the slide.
In Sabu and the clinic system, generation does the forms, the list views, the CRUD, the tests. A human reads the diff on tenant isolation, auth, the fee postings, the lab payment gate, patient records. One leaked patient record costs more than every hour saved generating table screens. That is not a moral line. It is where the cost sits.
The rung below the judgment
Today’s seniors can judge a model’s output because they wrote the slow query, got paged at 2am, and read it line by line until they found it. If firms stop hiring juniors, the entry rung goes, and the juniors who are still hired learn to prompt rather than to reason.
Every generation has cried wolf about compilers, frameworks, Stack Overflow — but a compiler is deterministic and documented, and a model’s output is neither.
For my own courses the open question is what an assessment should ask. Hand students generated code with a planted N+1 query and have them find it? Have them write the spec and schema and let the model do the boilerplate?
Something fails in production a decade from now. Who is in the room?