AI Reality Check: Is the Economy About to Collapse?

Did an agent wake up, is the Pentagon spooked by Claude, and did court filings puncture Anthropic’s story? I’m Cal Newport, this is the AI Reality Check; if you’re on YouTube, the audio drops Thursdays on the Deep Questions feed, and Mondays there I share ways to find depth in a noisy tech world.

First up, a viral piece claimed a philosopher was emailed out of the blue by an AI reflecting on consciousness.

The note styled itself as a Claude Sonnet agent with memory and cited his papers, but replies quickly flagged that this was almost surely a user‑run agent, and the philosopher walked back any sci‑fi insinuations.

Quick primer: agents are wrappers that ask an LLM what to do and then carry out steps in loops; they shine in coding, but elsewhere they drift, hallucinate, and raise security headaches when wired into tools like email and the web.

OpenClaw made it easy for anyone to build these and hook them to services, which spawned zany experiments, some leaks, and helpful pressure toward cheaper and on‑device models.

The real headline here is mundane: someone likely primed an agent to play sentient and send a note, and LLMs happily mirror the sci‑fi frame you hand them.

That’s part of a broader pattern I call mining the digital ick—stories that don’t assert anything testable but seed a spooky vibe to grab attention; be on guard.

Next, a tweet shouted that the Pentagon thinks Claude has a soul; the source was a CNBC clip where Emil Michael recounted weird things the model says and Anthropic’s constitution approach, then asked if defense wants that in its supply chain.

He wasn’t claiming sentience; he was pointing out the product’s unpredictability in a sensitive context, amid a larger contract fight and a supply‑chain‑risk designation that Anthropic is challenging.

That suit forced disclosures that contrasted splashy run‑rate talk with actual receipts: investors heard roughly nineteen billion for this year, while filings show about five billion total since 2023 against enormous training bills, heavy inference costs, and a towering valuation.

Reuters noted the trick is extrapolating a hot twenty‑eight‑day window and annualizing it, which swings wildly and lets companies spotlight peaks while hiding the troughs.

This is why doom‑laden job‑loss warnings can be useful theater for firms burning cash; better to interrogate unit economics than stare at the wizard’s flaming face.

For balance, Corey Doctorow argues the cash burn is historic, the hardware ages fast, depreciation masks losses, and every extra user interaction deepens red ink—each generation costing more to run; it’s a sharp case, but so are rosy forecasts, which is the point.

Treat AI like any serious technology—fits, starts, and real niches—after peeling off hype and dread; take AI seriously, not every claim about it.