AI Reality Check: Did the LLM Job Apocalypse Begin Last Week?

What It Covers

Cal Newport applies a computer science lens to three AI news stories from one week: Block's 40% workforce reduction attributed to AI, Anthropic's PhD-level intelligence claims tested against a Cornell freshman CS course, and survey data from 350+ professional programmers on real agentic coding tool usage.

Key Questions Answered

• •AI Washing in Layoffs: When a CEO attributes mass layoffs to AI without specifying which tools eliminated which roles, treat the claim skeptically. Block grew from 4,000 to 10,000 employees between 2019 and 2025, largely through crypto acquisitions that are now failing. Industry analysts and even pro-AI commentators like Ethan Mollick concluded the cuts reflect pandemic over-hiring, not automation.
• •Attributed Headlines as a Red Flag: Journalists increasingly place unverified claims in subheadlines followed by a comma and a source attribution, allowing a narrative to circulate without direct fact-checking. When reading AI job-loss coverage, check whether the article provides specific operational evidence or simply repeats an executive's self-serving framing to explain a stock price jump.
• •LLM Education Comparisons Are Misleading: Anthropic's PhD-level intelligence claim originated from performance on a single graduate math problem set. A Cornell TA tested ChatGPT, Claude, and Gemini across an entire freshman CS semester using real rubrics. Claude and Gemini earned C+, below the 2.5 GPA threshold required to declare a CS major at Cornell.
• •Agentic Coding Adoption Is Real but Uneven: Newport's survey of 350+ professional programmers found roughly 45% now generate the majority of their code using agentic tools like Claude Code or ChatGPT Codex. However, the productivity gains are partially offset by new overhead: prompt composition, output verification, re-prompting, and more rigorous code review when AI-generated code enters the pipeline.
• •Multi-Agent Hype Doesn't Match Professional Practice: Viral content promotes complex hierarchical agent systems with dozens of simultaneous AI workers. Actual professional programmers in Newport's survey rarely use this approach, finding it causes excessive context-switching and leads to accepting lower-quality outputs. Vibe coding — describing an app and returning days later to a finished product — remains confined to hobbyists and personal projects.