The Code Delusion: Why the Entire Software Industry Is Lying to Itself

// 01. THE PLACEBO ERA IS OVER. THE GUILLOTINE ERA JUST STARTED.

For two years, the software industry ran a $7 billion experiment on itself. Forty-one percent of all code written in 2025 was generated by AI. Microsoft announced 30%. Google said 25%. Twenty million developers signed up for GitHub Copilot. Ninety percent of Fortune 100 companies deployed it. The industry declared victory before anyone checked whether the software actually got better.

It did not. Google’s DORA report found that despite 90% AI adoption, organizational delivery metrics stayed completely flat. Individual output was up: 21% more tasks, 98% more pull requests merged. But the software reaching production was no better, no faster, no more reliable. Pull request sizes inflated 154%. Code review time grew 91%. Bug rates increased 9%. Security vulnerabilities in AI-generated code were 2.74 times higher than human-written code. A peer-reviewed METR study ran 246 real tasks with experienced open-source developers: they predicted AI would make them 24% faster, believed afterward they were 20% faster, and were measured at 19% slower. A 39-point gap between what developers felt and what was true. The Copilot era was a collective hallucination. The industry spent billions on tools that made experienced developers objectively worse at their jobs while convincing them they were better.

That was the placebo era. Then, on February 3, 2026, Anthropic launched Cowork. Forty-eight hours later, $285 billion was wiped from SaaS stocks in a single trading session. Thomson Reuters dropped 18%, its worst day on record. LegalZoom fell 20%. The S&P 500 Software & Services Index collapsed 20% year to date. Two days later, Claude Opus 4.6 shipped. The market understood in a single week what most engineers have still not processed: the tools that arrived in February are not a better version of the tools that failed. They are a different species of technology. And they are not here to help developers code faster. They are here to make the act of coding, as a human profession, structurally unnecessary.

The industry spent two years celebrating tools that inflated output, degraded quality, and made developers measurably slower. Then in a single week, $285 billion in market value evaporated because the real tools arrived. The placebo era is over. What comes next is not an upgrade. It is a restructuring.

// 02. THE TECHNICAL CASE FOR WHY THIS NEVER STOPS

Most people discussing AI coding are debating whether the current tools are good enough. That is the wrong frame. The right frame is the trajectory, and the trajectory is governed by mathematics that have held without exception for six years. Neural scaling laws predict model performance as a power function of compute, data, and parameters. They have been validated at every scale. No ceiling has been hit. Every time a frontier lab doubles its compute budget, performance improves on a smooth, predictable curve. And now there is a second independent scaling axis: test-time compute. Models that think longer on harder problems follow their own power law. Two independent exponentials, compounding simultaneously.

Context windows went from 200,000 tokens to one million in twelve months. At 200K, a model holds a few files. At one million, it holds an entire large codebase in a single reasoning session: every dependency, every edge case, every architectural decision buried in a file nobody has touched in two years. The model sees the whole system at once. At ten million tokens, which is not a question of if but when, it holds a full monorepo with its complete history. At a hundred million, it holds every version of every file ever written, every design document, every issue tracker conversation. The model does not just understand what the code does. It understands why every decision was made.

But the most consequential technical development is the recursive loop. Claude Code now writes Claude Code. Boris Cherny, its creator at Anthropic, confirmed that 100% of his code contributions since November 2025 have been written by the tool he built. A team of four people built Claude Cowork, a full multi-user product, in ten days. When asked how much code Claude wrote, the answer was: all of it. Anthropic’s CPO said publicly that Claude is being written by Claude. This is not metaphorical. The models are improving the infrastructure that trains the next generation of models. Better models generate better synthetic training data. Better training data produces better models. The loop is running. And each cycle compresses the time to the next capability jump.

METR, the research organization that proved Copilot-era tools made developers slower, also tracks a different metric: the length of software engineering tasks AI can complete at 50% reliability. That number has been doubling every four to seven months. Claude Opus 4.5 could complete tasks that take humans five hours. Extrapolate at the current rate and you reach sixteen-hour tasks by early 2027, five-day tasks by mid-2028, and month-long projects by 2029. These are not predictions from companies selling AI. This is observational data from the organization most famous for proving AI tools do not work. Even the skeptics’ own data says the trajectory is vertical.

// 03. THE EVIDENCE IS NOT THEORETICAL. IT IS ALREADY SHIPPING.

Jack Dorsey cut Block from 10,000 employees to under 6,000 on February 26, 2026. His shareholder letter was explicit: “Intelligence tools have changed what it means to build and run a company. A significantly smaller team, using the tools we’re building, can do more and do it better.” His formula: “100 people plus AI equals 1,000 people.” When asked what triggered the decision, he said: “Something happened in December of last year where the models just got an order of magnitude more capable, and it’s really shown a path forward in terms of us being able to apply it to nearly every single thing that we do.” Block’s stock surged 24% the next day. The market did not punish a 40% workforce reduction. It rewarded it.

Tobi Lütke of Shopify issued an internal memo that said: “Before asking for more headcount and resources, teams must demonstrate why they cannot get what they want done using AI.” Shopify went from 11,600 employees to 8,100 while revenue grew 21% annually. Marc Benioff of Salesforce stopped hiring engineers entirely in 2025, citing a 30% productivity increase from AI agents, then told a room full of CEOs: “We are the last generation to manage only humans.” Dario Amodei of Anthropic said at Davos in January that we may be six to twelve months from AI performing end-to-end what software engineers do. Mark Zuckerberg told Dwarkesh Patel that every engineer at Meta will become a tech lead managing their own army of agents.

Andrej Karpathy, who coined the term “vibe coding” a year ago, has already moved past it. He described giving an AI agent a single paragraph of instructions: set up vLLM on his DGX Spark, benchmark a vision model, build a video analysis dashboard, wire up systemd services, write a markdown report. The agent ran for thirty minutes unsupervised. It hit errors, researched solutions online, debugged, tested, configured services, and came back with a finished system. His conclusion: “You’re not typing computer code into an editor like the way things were since computers were invented. That era is over.” His new framework is “agentic engineering”: orchestrating long-running AI systems with tools, memory, and instructions that manage multiple parallel coding instances. Not using AI to code. Using AI to manage AI that codes.

Stanford researchers analyzed ADP payroll data for millions of workers and found that employment for software developers aged 22 to 25 has declined nearly 20% since late 2022. Junior developer hiring at Big Tech collapsed from 32% of new hires in 2019 to 7% in 2026. But employment for developers over 30 grew 6 to 12% in the same period. The two cohorts moved together for years, then diverged sharply the moment AI coding tools arrived. AI did not eliminate engineering. It eliminated transcription. And Anthropic stress-tested Agent Teams with sixteen agents building a C compiler from scratch. Rakuten ran Opus 4.6 against a 12.5-million-line codebase and achieved 99.9% accuracy in seven autonomous hours. A senior Google engineer told the San Francisco Standard: “I’m basically a proxy to Claude Code. My manager tells me what to do, and I tell Claude to do it.” His predominant feeling was grief. The skill he spent years developing was just commoditized to the general public.

Dorsey did not cut 4,000 people because Block was struggling. He cut them because 100 people plus AI equals 1,000 people, and the math got undeniable in December. That is not a layoff. That is a restructuring of what a company is. And he said publicly that most companies are late, and within a year the majority will reach the same conclusion.

// 04. THE VERDICT

I run 65 engineers at BearPlex. Not one of them was ever hired to type code. Every single one was hired to think in systems: to decompose ambiguous problems, define constraints and failure modes, architect the structures that run without human intervention, and make the decisions that machines cannot make. When the Copilot era produced nothing but inflated pull requests and a 39-point perception gap, we did not participate. When Opus 4.6 arrived with a million-token context window, Agent Teams, and the ability to sustain coherent autonomous work for hours, we did not panic. We deployed it. Because we had always built the organization around architecture, not transcription.

The uncomfortable truth that the industry will not say plainly: most of what passed for software engineering was never engineering. It was transcription with an engineering title. Taking a Jira ticket and turning it into a pull request. Taking a design document and turning it into functions. Taking someone else’s architecture and implementing it keystroke by keystroke. That job is gone. Not threatened. Not evolving. Gone. Fifty-five thousand roles were cut explicitly because of AI in 2025 alone, twelve times the number from two years prior. Fifty-five percent of companies that rushed to replace humans with AI now regret it, according to Forrester, because they replaced the wrong humans. They cut architects and kept typists, or could not tell the difference. The companies that understood the distinction accelerated. The ones that did not are now rehiring at a loss.

The discourse is “will AI replace software engineers?” That is the wrong question. The right question is: did you ever have software engineers, or did you have typists with engineering titles? Because the data answers it. The old tools made typists slower while they believed they were faster. The new tools made architects an order of magnitude more productive while making typists structurally unnecessary. Scaling laws guarantee the models keep improving. The recursive loop guarantees the rate of improvement itself accelerates. Cost curves guarantee access expands to everyone. The trajectory is not ambiguous. It is vertical. And it does not have a ceiling that anyone can identify.

Levels builds $3 million in revenue with zero employees. Karpathy converts weekend projects into thirty-minute agent tasks. Dorsey eliminates 4,000 roles and his stock surges 24%. Sixteen AI agents build a C compiler from scratch. A 12.5-million-line codebase is modified with 99.9% accuracy in seven hours. These are not future projections. These are last month’s case studies. The companies panicking are the ones that never built systems. They accumulated code. The companies thriving are the ones that always understood that code is an output, not an asset, and that the only durable asset in software is the architecture that determines what gets built, how it gets verified, and why it exists at all. Hype is a temporary tactic. Architecture is a permanent advantage. And every month the models get better, that advantage becomes harder to fake and more expensive to lack.