LLMs were supposed to delete software engineering. The headcount kept rising.

The hype pattern here is a demo-to-payroll leap. A model writes a function from a comment. A CEO says 30% of a repo is AI-generated. A thread says "we had a good run." None of those automatically deletes a job category. But they do change who gets hired, how fast seniors move, and how scared the feed feels.

This file is about software engineering only. Not every desk job. Not extinction. The question is narrower and more testable: did coding LLMs take over software engineering jobs, and what does the record show so far?

What coding LLMs actually shipped

Before the CEO percentage tour, there was a product ladder.

GPT-1 (June 2018) was a language-understanding paper. OpenAI's 117M-parameter model beat NLP benchmarks. It did not ship as a coding product.

GPT-2 (February 2019) scared people about fake news and spam, not payrolls. OpenAI's release note listed misuse risks like impersonation and abusive content. By July 2019, developers wired GPT-2 into Deep TabNine, an autocomplete plugin. That is the first widely felt "AI writes code" moment for many engineers, framed as speed, not replacement.

GPT-3 (June 2020) went behind an API. WIRED captured Silicon Valley's gloomy predictions about programmer employment after code-generation demos went viral.

Codex and Copilot (2021) made coding LLMs a product. GitHub announced Copilot on June 29, 2021, powered by OpenAI Codex. The pitch was "AI pair programmer," not pink slip.

ChatGPT (November 2022) put natural-language coding in everyone's browser.

GPT-4 (March 2023) and the agent/editor wave after it (Cursor, Copilot Workspace, Claude Code) raised the ceiling again.

The montage compresses eight years into one mood: the robots learned to code, so the coders must be finished.

CEOs were already writing obituaries before Copilot shipped

The common memory is closer to the record than the neat "this started with ChatGPT" story. Tech leadership has been publicly targeting coders and entry-level programming since 2019, often before coding LLMs were a daily product.

The timelines slip. The target does not.

February 2019. Sam Altman, then running Y Combinator, told the New York Times New Work Summit that AI would "probably replace most of today's jobs" and that "entire classes of jobs will go away and not come back." He was not naming software engineers alone. But he was already selling a world where whole job categories vanish.

May 2019. Mark Cuban went straight at coders on Recode Decode: "Twenty years from now, if you are a coder, you might be out of a job." He argued a computer science degree would lose value as AI advanced and that humanities skills would matter more. That is not "months." It is still a billionaire on a major podcast telling students the coding major may not age well.

August 2019. Elon Musk told the World Artificial Intelligence Conference in Shanghai that AI would make many jobs "kind of pointless," then narrowed to programmers: young people should study engineering because programmers would be least vulnerable, "but, even then, eventually the AI will just write its own software." GPT-2 had shipped six months earlier. Copilot was two years away.

May 2020. Jack Dorsey, CEO of Twitter and Square, told Andrew Yang's podcast that AI "is even coming for programming" and that "a lot of entry-level programming jobs will just not be as relevant anymore" because machine learning's goal is to "write the software itself over time." He cited Brookings work ranking computer programmers among the occupations most exposed to AI. That is a sitting tech CEO, in 2020, aiming at the bottom of the engineering ladder.

March 2021. Altman's Moore's Law for Everything essay said "software that can think and learn will do more and more of the work that people now do." Again, not engineers-only. But the computer-work category was already on the slide.

June 2021. Altman tweeted that AI would make "the price of work that can happen in front of a computer" fall faster than physical-world work. Hacker News immediately read that as programmer exposure. Copilot preview launched three weeks later.

April 2022. In his DALL-E 2 launch post, Altman wrote that Copilot was "very far from being able to create a full program" but that AI was "increasingly going to make some jobs not very relevant." He named labor markets explicitly.

May 2023. Jensen Huang declared at Computex: "Everyone is a programmer now. You just have to say something to the computer." In 2024 he flipped the old advice: do not teach every kid to code; human language is the programming language now.

July 2023. Stability AI CEO Emad Mostaque said flatly: "There will be no programmers in five years."

The record is plain. The CEO replacement tour predates ChatGPT, predates the code-percentage earnings calls, and in Dorsey's case names software engineers directly in 2020. Fortune 500 bosses rarely say "all engineers gone in ninety days." The public version is a stack of "eventually," "entry level," "five years," and "maybe half the code," repeated on stages while headcount rises.

Not every executive sang the same song. GitHub CEO Thomas Dohmke said Copilot would write 80% of code "sooner than later" but also that "the developer is not going to be replaced." That is the industry split in one sentence: more AI-written lines, same human pilot, uncertain payroll.

Table 3: The replacement clock (and the payroll that ignored it)

This is the proof table. Every row is a dated claim about engineers or programmers being taken over, with the stated deadline and the BLS headcount in the same year or the year after. Headcount uses BLS occupation 15-1256 (developers + QA testers combined) for 2019–2020, then 15-1252 (software developers only) from 2021 onward. The series is not perfectly apples-to-apples, but the direction is.

Date	Who said it	Stated deadline	What they said	Source	U.S. developer headcount nearby
Jan 2019	Essayist on Medium	No date, structural decline	"We've reached peak code. From this point forward, coders will be in decline."	Medium	1,406,870 (BLS 2019)
May 2019	Mark Cuban	~20 years	"Twenty years from now, if you are a coder, you might be out of a job."	CNBC	Same year
Aug 2019	Elon Musk	Eventually	Programmers safest for now, but "eventually the AI will just write its own software."	CNBC	Same year
May 2020	Jack Dorsey	Soon / over time	AI will "soon write its own software"; "entry-level programming jobs will just not be as relevant anymore."	CNBC	1,476,800 (BLS 2020)
Jul 2020	Trade press on GPT-3 demos	Imminent wipeout (mood)	WIRED on viral code demos: "prompted gloomy predictions about the employment prospects of programmers."	WIRED	Same year
Oct 2020	The Next Web	2030 / "few months" rhetorical	"Developers might be obsolete by 2030" and, in the same piece, "in a decade, in a few months even, you'll probably be doing things you can't imagine."	TNW	Same year
Jun 2022	Stephan Schmidt (CTO essay)	~10 years	"Software Engineering is dead in 10 years"; "the endgame is near and it's the end of software engineering."	AmazingCTO	1,534,790 (BLS 2022)
Oct 2022	Thomas Dohmke (GitHub CEO)	~5 years	Copilot could write 80% of code within five years; "some programming jobs redundant."	The Decoder	Same year
Jan 2023	Matt Welsh (ex-Google, Fixie)	Next few years	AI-generated code "being the norm within the next few years"; "fire all the devs" framing.	Medium	1,656,880 (BLS 2023)
Dec 2022	Tech press on ChatGPT	2–3 years	Consultant told TechTarget ChatGPT would become "a proficient coder within the next decade" and that "advancements are happening such that we'll get there in two to three years."	TechTarget	Headcount still rising
Jul 2023	Emad Mostaque (Stability AI)	5 years	"There will be no programmers in five years."	Decrypt	Same year
Mar 2025	Dario Amodei (Anthropic)	3–6 months	"In three to six months, where AI is writing 90% of the code." In 12 months, "essentially all of the code."	Business Insider	1,693,800 (BLS 2024)
Jan 2026	Dario Amodei (Davos)	6–12 months	"We might be six to twelve months away from when the model is doing most, maybe all of what software engineers do end to end."	Entrepreneur	No 2025 annual OEWS yet; junior hiring signals down
Jun 2026	Elon Musk (xAI)	End of 2026	"You don't even bother doing coding" by year's end; AI writes the binary directly.	Machine Design	Clock still running

Read the deadline column against the headcount column. The clocks reset. The occupation grew.

When Amodei said in March 2025 that 90% of code would be AI-written in three to six months, September 2025 arrived and LessWrong and Skeptics Stack Exchange were already scoring the miss. Six months later, at Davos 2026, he moved the window to six to twelve months for models doing "most, maybe all" of software engineering end to end. That is not a one-off mistake. It is a recurring format: short horizon, maximal claim, payroll unchanged.

The clocks are not one slogan. Leadership keeps selling a short horizon, then replacing it when it expires: three to six months, six to twelve, "soon," "within a few years," "no programmers in five years." The Next Web in 2020 put "in a decade, in a few months even" in the same essay that told developers they would not become obsolete. The vibe and the caveat live in the same paragraph.

Table 1: Releases, CEO rhetoric, and the labor market at the time

This table pairs the coding-relevant model moment with what top executives said about engineering jobs and what U.S. labor data looked like nearby. CEO quotes are about intent, fear, or bragging. They are not payroll reports.

Release / moment	Date	What leadership said about software engineering jobs	U.S. labor signal nearby	Mood on the ground
GPT-1 paper	Jun 2018	No major CEO job-loss tour. Research benchmark, not a coding product.	No coding-LLM-specific labor shock. National unemployment ~3.9% (BLS).	ML Twitter excited; most devs untouched.
Altman at NYT summit	Feb 2019	AI will "probably replace most of today's jobs"; "entire classes of jobs will go away and not come back." Not coder-specific, but sets the frame.	Unemployment ~3.8%.	Valley optimism plus job-loss talk in the same keynote.
GPT-2 staged release	Feb 2019	OpenAI warned about fake news and abuse, not engineer layoffs.	Unemployment ~3.7%.	"Too dangerous to release" meant misinformation, not HR.
Cuban on Recode	May 2019	"Twenty years from now, if you are a coder, you might be out of a job." CS degrees will lose value.	Unemployment ~3.6%.	First major billionaire-on-podcast shot at coding as a career bet.
Musk in Shanghai	Aug 2019	Programmers least vulnerable for now, but "eventually the AI will just write its own software."	Unemployment ~3.7%.	General AI-job panic; coding named as a temporary safe harbor.
Deep TabNine + GPT-2	Jul 2019	No CEO layoff rhetoric tied to the product. Trade press: faster autocomplete.	Unemployment ~3.7%.	Devs impressed; BBC later would frame Copilot-era tools as "could put them out of a job" while quoting coders who wanted help with boring syntax.
Dorsey on Yang Speaks	May 2020	AI "is even coming for programming." "A lot of entry-level programming jobs will just not be as relevant anymore." Cited Brookings ranking programmers third-most AI-exposed.	Unemployment 13.0% (COVID peak month).	Sitting tech CEO targets junior devs before Copilot exists.
GPT-3 API	Jun 2020	No formal CEO "fire the devs" campaign. Valley press ran programmer-gloom pieces after demos.	Unemployment 8.1% (COVID). Hard to attribute anything to GPT-3.	Awe and niche dread in tech Twitter.
Altman computer-work tweet	Jun 2021	Predicted AI would cut the price of computer-based work faster than physical work. HN read it as programmer deflation.	Same month Copilot preview follows.	Forum panic ahead of the product launch.
GitHub Copilot preview	Jun 2021	GitHub pitched "AI pair programmer." Dohmke later said 80% of code from Copilot eventually, but devs not replaced.	Software developers: 1,364,180 employed (BLS May 2021). Unemployment falling to 5.4% in 2021.	"Beginning of the end for programming jobs" on forums; daily use mostly autocomplete.
Altman on DALL-E 2	Apr 2022	Copilot "very far" from a full program, but AI "increasingly going to make some jobs not very relevant."	Headcount still climbing.	OpenAI boss names labor impact while shipping coding tools.
ChatGPT	Nov 2022	Still pre-earnings-code-percentage era.	Software developers: 1,534,790 (BLS May 2022).	Every dev tries it once; "we had a good run" memes.
Huang at Computex	May 2023	"Everyone is a programmer now." Speak to the computer; programming barrier "incredibly low."	Software developers: 1,656,880 (BLS May 2023).	Democratization frame: fewer specialists needed.
Mostaque podcast	Jul 2023	"There will be no programmers in five years." Claimed 41% of GitHub code already AI-generated.	Same year: OEWS count still rising.	Shortest CEO clock on record.
Altman Senate testimony	May 2023	GPT-4 will "entirely automate away some jobs" while creating others; government should help on displacement.	Same BLS year as Huang/Mostaque.	Institutional panic begins.
Altman on tasks vs jobs	Jun 2023	AI is "good at tasks, bad at jobs" for now. If every developer becomes 3x faster, the world still needs more software.	Same year: headcount still rising on OEWS.	Relief and skepticism in the same breath.
Pichai Q3 2024 earnings	Oct 2024	"More than a quarter of all new code at Google is generated by AI," then reviewed by engineers. Helps them "do more and move faster."	Software developers: 1,693,800 (BLS 2024 baseline).	Code-% headlines recycled as job-loss proof.
Nadella at LlamaCon	Apr 2025	"Maybe 20%, 30%" of code in some Microsoft repos is AI-written, rising steadily. Same week Zuckerberg said maybe half of Meta's development could be AI-driven in a year.	Microsoft later said it still plans to hire engineers, with "more leverage" per head.	CEOs compete in code-percentage bragging.
Altman on Stratechery	2025	Each engineer will "do much, much more for a while." Then "maybe we do need less software engineers." Said AI coding is past 50% in many companies; "agentic coding" is next.	Stanford/ADP study: 16% relative hit to 22–25-year-olds in AI-exposed occupations including software.	Fear gets more specific: fewer juniors, same seniors.

Read the last column against the labor-signal column. The scary quotes got louder after the tools went mainstream. The national developer headcount line mostly went up.

Table 2: The software engineering labor ledger since the replacement clock started

This table runs from 2019, when the public "coders are finished" tour was already rolling, through the coding-LLM era. BLS 15-1256 (developers + QA) for 2019–2020; 15-1252 (developers only) from 2021.

Year	U.S. software developers employed	Change vs prior year	Unemployment (all workers)	Replacement headline that year	What the tape shows
2019	1,406,870 (BLS 15-1256)	baseline	3.7%	Cuban: coders may be out of a job in 20 years. Musk: AI will eventually write its own software.	Headcount high; panic mostly forward-looking.
2020	1,476,800 (BLS 15-1256)	+5.0%	8.1%	Dorsey: entry-level programming jobs "not as relevant." WIRED: GPT-3 demos spark programmer-gloom predictions.	Count rose in COVID year on this series.
2021	1,364,180 (BLS)	−8.2%*	5.4%	Copilot preview; Altman tweet on computer-work prices falling.	*SOC break to dev-only series; not a clean YoY compare.
2022	1,534,790 (BLS)	+12.5%	3.6%	ChatGPT at year-end; AmazingCTO: software engineering "dead in 10 years."	Strong rebound on dev-only count.
2023	1,656,880 (BLS)	+8.0%	3.6%	Mostaque: no programmers in five years. Huang: everyone is a programmer now.	Headcount still climbing.
2024	1,693,800 (BLS EP)	+2.2%	4.0%	Pichai: quarter of new Google code AI-generated. Code-% tour goes mainstream.	CEO percentages treated as layoff forecasts; tape still up.
2025 (early signal)	No clean annual OEWS yet	Tech layoffs large; AI cited in some cuts	4.3% (BLS)	Amodei: 90% of code in 3–6 months. Then 6–12 months for full SWE automation at Davos 2026.	IZA: junior vacancies down 14–15%. Stanford: entry-level employment down ~16% in exposed roles. Macro count still not collapsing.

Two caveats belong in the margin. BLS OEWS is employer-reported headcount, not "developers on Twitter." It lags the news cycle by months. The 2019–2020 and 2021+ series use slightly different SOC buckets. And 2022–2024 still includes the post-COVID tech hiring hangover, so you cannot attribute every row to AI alone.

Still, the direction is clear: from Cuban and Dorsey through Amodei's months clocks, the national developer occupation did not shrink on the tapes we have.

Graph: U.S. software developer employment vs the replacement montage

From the 2019 combined-series baseline to the 2024 developer count, employment is up about 20% while the replacement headlines never stopped. That is the whole paradox in one line.

Graph: How developers feel about AI taking their job

National headcount is one tape. Sentiment is another. They can diverge for years.

The fear line wobbles, but the majority still says no. In the 2025 survey, 63.6% said AI is not a threat, 15% said yes, 21.3% were unsure. That is down from 68% "not a threat" in 2024, so anxiety ticked up, but it did not flip the profession into majority panic.

At the same time, 84% use or plan to use AI tools, while only 29% trust AI accuracy and 46% distrust it. Developers are adopting tools they do not fully trust while telling surveyors they are not doomed. That is the social version of the productivity paradox.

The paradox of AI coding

Here is the weird part everyone feels but few headlines name cleanly.

Paradox 1: More AI-written code, more engineers (so far). CEOs cite rising AI code share in the same period BLS developer employment rose. Those metrics are compatible. More AI-generated lines per engineer is not the same as fewer engineers. It can mean more experiments, more features, more review work, more slop to delete.

Paradox 2: Developers feel faster while RCTs show slowdown. In METR's 2025 randomized trial, experienced open-source developers on familiar repos took 19% longer with AI allowed, mostly using Cursor + Claude. Before the study they expected a 24% speedup. Afterward they still believed AI had sped them up by 20%. The tools felt helpful while the clock disagreed.

Paradox 3: Macro headcount up, junior ladder down. Brynjolfsson et al. find a 16% relative employment decline for 22–25-year-olds in AI-exposed occupations including software developers, with experienced workers stable. IZA finds junior software vacancies down 14–15% relative to senior after ChatGPT, with higher experience requirements within the same job titles. The job category grows. The first rung gets narrower.

Paradox 4: "Almost right" code eats the savings. Stack Overflow reports 66% of developers frustrated by solutions that are almost right but not quite, and 45% say debugging AI-generated code is more time-consuming. The demo shows the happy path. Maintenance pays the invoice.

That is the AI coding paradox in one sentence: the montage sells replacement; the micro-data sells friction; the macro tapes still show growth; the ladder may be where the damage lands first.

What research actually shows

Strip the CEO tour and the memes. The research file is smaller but sharper.

Employment at occupation level: BLS counts more U.S. software developers in 2024 than in 2021. The occupation's own outlook still projects 16% growth from 2024 to 2034, faster than the average for all jobs.

Entry-level compression: The strongest hiring signal is not mass unemployment but fewer junior slots. Stanford's administrative payroll data and IZA's vacancy data point the same direction after ChatGPT: seniority-biased change, not a cliff for everyone with a GitHub account.

Productivity: METR's experienced-dev slowdown is one setting, not universal law. Other studies find gains on greenfield tasks, boilerplate, or learning speed. The honest read is task-dependent, not "always 10x" or "always fake."

Adoption vs trust: Stack Overflow shows near-universal experimentation and falling trust. Developers treat AI like a fast intern who sometimes hallucinates a library.

Layoffs: Tech layoffs in 2025 were real and loud. AI was cited in some memos, but so were overhiring, rates, and restructuring. A code-percentage keynote is not a severance spreadsheet.

How people feel (and why that matters)

Feelings are data here because adoption runs on them.

Engineers in 2021 mostly heard Copilot as power tools. The BBC could headline job loss while quoting developers who wanted less typing.

Engineers in 2023 got the Senate version: some jobs automated away, many created, nobody sure which bucket is yours.

Engineers in 2025 live in a split screen. CEOs quote AI code fractions. Surveys show most devs still reject the "my job is doomed" label. Entry-level applicants see fewer junior postings. Seniors use Cursor daily and distrust the output.

Stack Overflow's team calls this cognitive dissonance straight: you use the tool because it helps, because your employer expects it, and because falling behind feels risky, even when you suspect it threatens the rung below you.

That emotional mix matters for policy and career choices even when the national headcount line is still rising. People do not live inside BLS tables. They live inside hiring loops.

The replacement clock kept moving. The payroll did not.

Fear of coder obsolescence has been public and continuous since 2019. By the mid-2020s the deadlines had shrunk to months. The national developer count did not follow them down.

Table 3 is the receipt stack. The pattern is not one bad prediction. It is a conveyor belt:

1. 2019–2020: Billionaires and CEOs name coders, entry-level programming, or "AI writes its own software" (Cuban, Musk, Dorsey). WIRED documents GPT-3 demos driving "gloomy predictions about the employment prospects of programmers." BLS count: up.

2. 2021–2023: Copilot, ChatGPT, "no programmers in five years" (Mostaque), "software engineering dead in ten years" (AmazingCTO), "AI-generated code the norm within the next few years" (Welsh). BLS developer count: up again.

3. 2025–2026: The clock gets shorter. Amodei: 90% of code in 3–6 months, essentially all code in 12. Six months later, Davos 2026: 6–12 months for models doing most or all of software engineering end to end. Musk: coding obsolete by end of 2026. National OEWS: still no wipeout on the annual tapes.

The public version is a near-term wipeout, whatever window the quote actually used. Trade press in 2022 already had experts saying ChatGPT would be a "proficient coder" in two to three years. TNW in 2020 mixed "obsolete by 2030" with "in a few months even" in the same piece. The feed compresses every horizon into right now.

That does not mean nothing is changing. Junior hiring and vacancies are the real bruise. But the cinematic version, engineers deleted on a quarterly clock while headcount cliffs, is not what the BLS series shows. The montage keeps saying months. The payroll line keeps rising.

What would change this file

We would update on evidence, not vibes.

Sustained decline in BLS software developer employment for multiple years while AI coding adoption rises.

Unemployment rising among developers specifically, not just "tech layoffs" headlines.

RCTs showing durable speedups on real production work, not only greenfield demos, reversing METR's experienced-dev result at scale.

Junior compression spreading from hiring rates into falling median experience, rising unemployment for new grads, or collapsing CS placement rates with AI as the documented driver.

Agentic systems closing end-to-end tickets with production liability accepted at scale, not demo repos.

Until then, the file is this: coding LLMs are real, daily, and emotionally huge. They have not yet deleted software engineering as an occupation on the national counts we have. The damage is more likely to show up first as fewer entry-level slots, weirder productivity math, and CEO code-percentage theater than as an instant wipeout the thumbnails promised.