The Software Economy Is Eating Itself
Bon appétit!
OpenAI is not just a company, but a religious cult which believes AI can achieve miracles. Google is an established industry juggernaut seeking to extend its power and influence. Anthropic is a religious cult like OpenAI, but an apocalyptic one that believes in ‘AI safety’ and ‘alignment’, so the machine God it creates is merciful.
Companies are generally not faith-driven. They make choices based on maximizing shareholder value and pragmatism. Anthropic became the world’s leading AI company by most effectively communicating its ideals to companies in a common language they understood. That common language is vibe-coding. As it turns out, the most valuable business proposition of AI is not mathematics, propaganda slop, truth-seeking, drug discovery, investment advice, education, medical or legal applications, search, or romance. The ‘killer app’ is coding and coding underlies nearly all economically valuable endeavors in today’s world. Software ate the world; now it’s feeding off itself.
With faith-based market valuations and capital expenditures on AI infrastructure which according to Nvidia CEO Jensen Huang could get up $4 trillion annually within years, the AI companies have a lot to prove. In a rush to socialize the financial risks of AI through IPOs, Anthropic applies creative accounting methods, while OpenAI seeks long-term commitments from customers in exchange for token discounts.
What could go wrong? Nothing according to the legacy media. Major news outlets exercise restraint from critical reporting about ‘tokenomics’ (once a term associated with crypto), but farm engagement by cherry-picking the most shocking statements from tech CEOs about how their products will solve climate change, diseases and the need to employ humans.
Vibe-coding is presently the most impactful application of AI and illustrates better than anything the ‘human-to-AI intelligence paradox’. As AI technology continues to evolve and surpass the abilities of humans, overreliance and the whole attention-based economy which laid the foundation for general purpose models, erodes our thinking skills.
On the demand side, cutting-edge models like Claude Mythos can discover and exploit software vulnerabilities at scale; on the supply side, vibe-coding creates an endless amount of flawed and buggy code.
The 50 US-based internet companies that are part of Anthropic’s Project Glasswing with access to Claude Mythos Preview have allegedly found “more than ten thousand high- or critical-severity vulnerabilities across the most systemically important software in the world”. The bottleneck is designing patches for all of the identified software bugs, which takes two weeks on average for critical vulnerabilities. In general, we can expect that cutting-edge models can identify vulnerabilities, much faster than the companies can patch them.
AI’s potential in cyberwarfare adds a new dimension of threats to a digital economy where scams and scammers are already omnipresent and protecting users isn’t a major priority for the gatekeepers. Internet users have never been more exposed to financial frauds and data theft, not despite, but because of the rapid advancements in AI.
Two software engineers behind the agentic coding tool OpenClaw, Mario Zechner and Armin Ronacher, recently issued a stark warning against ‘vibe slop’ in the WSJ:
“Their core message: These systems are supposed to make senior engineers so productive that companies can lay off junior engineers, but in reality, many companies are trading near-term productivity for long-term woes. Not only does the pipeline of junior talent dry up, but residual effects include buggy software, service outages, security vulnerabilities and mounting technical debt.”
Yet, the gatekeepers of our digital infrastructure are revelling in vibe slop. Google estimates that 75% of the code it produces today is AI-generated. According to Microsoft CEO Satya Nadella, the share of AI code in new Microsoft projects is around 20-30%, but its CEO Kevin Scott claims that 95% of code could be AI generated within five years. Meta expects that 65% of its code will be AI-generated in the first half of 2026 and Snap is already operating with the same share.
For the industry leaders, keeping humans-out-of-the-loop is a goalpost and a source of pride with important signal value to investors. But turning programmers into agent orchestrators is the end goal. The end goal is recursive self-improvement (ref/Jack Clark), creating software that creates and maintains software by itself at a level of skill and abstraction no human coder can follow. The end goal is to do what Deep Blue did to Kasparov and what Move 37 did to the game of Go, reaching a ‘coding singularity’ where computers are no longer learning and instructed by human, but vice versa.
What does all of this mean for the future job prospects of coders ? I can’t tell since I’m not a coder, but Lars Faye weighs in with an interesting perspective on his blog.
If recursive self-improvement is achieved in coding, the same ‘ouroboros effect’ will be observed across most other industries, because the economy runs on code. Most while-collar work, including the jobs of lawyers, judges, doctors, accountants, financial advisors, architects, etc. would soon be disrupted by AI engineers, until those jobs are replaced by self-correcting code entirely.
Don’t think this is my projection! Microsoft AI CEO Mustafa Suleyman predicts it will happen within the next 12-18 months. Ironically, the shift to an AI-first society would also mean that consumer spending will plummet, since people no longer have any money to spend without a source of income for their skills and qualifications. The whole idea seems breathtakingly stupid. What an exciting time to be alive.
