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

Did AI just cost forty percent of jobs at Block, can today’s models clear a freshman CS course, and what are coders really doing with agentic tools? I’m Cal Newport, and you’re listening to AI Reality Check.

Quick programming note: I’m spinning AI coverage into short Thursday episodes as an experiment, and you can share them easily from YouTube.

Q: Did Block really lay off almost half its staff because of AI?

A: Jack Dorsey’s note said the company is shrinking from just over ten thousand to just under six thousand, credited to new intelligence tools and flatter teams, and yes, I noticed the no‑caps vibe in a message that serious. Media ran with the AI angle and the stock jumped, but that framing deserves scrutiny.

A: Block ballooned headcount during the pandemic, bought into crypto and blockchain bets that cooled, and has missed earnings targets recently, which looks a lot like industry right‑sizing we’ve seen at other tech giants.

A: The claim lacks specifics, agentic tools are too new for a sudden company‑wide fifty percent efficiency leap, and even AI optimists like Ethan Mollick called it a smart story to tell investors, while a FinTech analyst covering Block flatly said don’t buy the excuse.

A: Bottom line, AI will change jobs, but calling this wave right now is AI‑washing; track real impacts carefully and skip the apocalypse vibes.

Q: Can top LLMs pass an advanced first‑semester CS class?

A: I ran every graded task from Cornell’s CS 2112 through ChatGPT, Claude, and Gemini, and scored them with the same rubric we used on students.

A: The models handled some early work and the recall‑heavy final fairly well, but they also hallucinated, ignored constraints, and even printed a goofy “hello world world” when the output was just supposed to be hello. Final grades: Claude and Gemini landed at C plus, ChatGPT at B plus, below the class median.

A: The takeaway isn’t a gotcha, it’s that “army of PhDs” or “country of geniuses” talk is misguided; these are specialized tools that perform best with guidance, checks, and task‑specific tuning, not human education labels.

Q: How are professional programmers actually using agentic AI right now?

A: From hundreds of practitioner reports I’m reviewing, many crossed the usefulness threshold after the fall Claude and ChatGPT updates, with some generating most code via an agent while they design, plan, iterate in chat, execute, and verify in Git.

A: Others are more cautious: AI crushes boilerplate, scaffolding, and quick docs lookups, but prompt crafting, re‑prompting, manual refactors, and heavier code reviews add friction that can erase gains.

A: Expect best practices to settle in the middle with clearer standards for where to auto‑generate and less chatter in planning and review, while the flashy many‑agent hierarchies stay mostly a hobbyist party trick.

Take AI seriously, not every headline; I’ll be back Monday with the main show and likely another Thursday reality check next week. See you then.