The tech industry's loudest voices have spent the past two years declaring that coding is solved — that LLMs can replace software engineers, that English will replace programming languages, and that the remaining human role is "taste." Mehdi Daoudi, a systems engineer with two engineering degrees and years of building AI-powered products, calls this what it is: a brain-dead narrative pushed by people selling tokens. His core argument is structural, not emotional. Code creation is cheap; code maintenance, reliability, security, and scalability are where the real costs live. These non-functional requirements — the stuff that keeps healthcare systems from killing patients and financial platforms from losing money — are precisely where LLMs remain weakest. LLMs are stochastic and probabilistic engines wrapped in deterministic harnesses. They struggle with logic, degrade with context window size, and cannot be held accountable when they fail. You cannot fine an LLM. You cannot imprison it. You cannot even meaningfully punish it. Daoudi identifies exactly three categories of software where you can skip reading the AI-generated code: personal projects, proofs of concept, and deliberate cyberattacks. The first two tolerate high failure rates. The third weaponizes them. Everything else — healthcare, finance, aviation, defense, manufacturing — demands accountability that no AI system can provide. The people most confident in LLM-generated code, he observes, are the ones least equipped to evaluate it: a textbook Dunning-Kruger dynamic where not reading the output correlates with higher confidence in it. The technical argument is precise. LLMs succeed at code generation primarily because engineers have built feedback loops that pipe compiler and runtime errors back into the model until surface-level bugs disappear. But this is error suppression, not understanding. The same engine that cannot count the R's in "Raspberry" is being trusted with production logic. Capabilities are following an S-curve with diminishing returns: more expensive models are not proportionally more productive. The claim that "English replaces code" ignores that programming languages exist specifically because natural language is too vague and conflicting for deterministic systems. Daoudi reserves his sharpest criticism for the downstream effects consumers are already experiencing. Google and GitHub services are degrading with bugs that accountability and quality gates would catch. The "move fast" ethos, measured in vanity metrics like lines of code and PR counts rather than service levels and customer satisfaction, is producing motion without progress. Leadership pressure to inject AI into every workflow is burning developer skill sets while enriching AI vendors. The piece also diagnoses what he calls "AI overdose" at the individual level: zero tolerance for disagreement, outsourcing cognitive challenge to chatbots, abandoning long-form reading, spending more time with AI than humans. His prescription is blunt — if your output quality equals or falls below AI output, upskill. Don't sacrifice long-term relevance for short-term velocity. The "taste" argument, he notes, is "the lie retired chefs tell themselves" — everyone has taste, and taste alone doesn't pay. This is not an anti-AI argument. Daoudi acknowledges genuine utility in proof-of-concept work, personal automation, language transformation tasks (translation, summarization, format conversion), and the ecosystem of harnesses, agents, and architectures built to compensate for LLM shortcomings. His target is the specific, extractive narrative that overstates capabilities to sell compute while the people who depend on reliable software absorb the cost of degraded quality.
