The New Problems AI Is Creating (And How People Are Solving Them)

Full analysis

NLW spent an episode with no news in it, and it's better for that. He's cataloguing the second-order problems AI creates once you actually deploy it: slop, token bills that behave like payroll, deskilling, and a broken expertise pipeline. For an engineering lead shipping AI into production, this is the operations manual nobody wrote yet. The question underneath it all: does the "AI is labor, not software" reframe change how you budget, staff, and route models? Type 2 decision, mostly reversible. You can re-architect your model routing and your token accounting quarter by quarter. The deskilling piece is closer to Type 1, because you can't rehire a junior bench you spent two years dismantling.

The Skeptic. Half of these "new problems" are old problems with a fresh coat of AI paint. Deskilling? We've been outsourcing junior work to offshore teams and Stack Overflow for fifteen years and the seniors still showed up. The BCG number is executives predicting a threat in three to five years, not measuring one. That's a survey of vibes. And "60% of executives expect distributed de-skilling" is exactly the kind of stat that sells a consulting engagement to fix it. For a PM: most of this is real but slow, and the people ringing the alarm are the people selling the fire extinguisher. The one thing I'd take seriously is the token bill, because that's a number on an invoice, not a forecast.

The Compute Pragmatist. The token-economics point is the only part of this episode with hard edges. Companies burning an annual AI budget in months is worth paying attention to, and the fix is boring and correct: tiered model routing. Stop pointing the frontier model at every task. Most of your workflow is classification, extraction, and summarization that a mid-tier model does at a tenth the cost. Route by problem type, cache aggressively, and put a cost-per-workflow number next to every agent you run. For a PM: your inference bill now scales with usage like a phone plan, not a Netflix subscription. If you don't meter it, it meters you. Greg Shove's line, don't token-max and don't token-minimize, is right and unsatisfying.

The Researcher. The "Tragedy of the Cognitive Commons" framing shared by Zara Zhang is the one genuinely new idea here, and it holds up. AI eats junior roles, which were the training ground, while simultaneously requiring expert judgment to catch its errors. Each firm cutting entry-level work is acting rationally; the profession collectively loses the pipeline that produces the experts who supervise the AI. That's a real coordination failure, not a vibe. For a PM: the person who checks the model's homework in ten years has to learn the job somewhere, and you're deleting the place they'd learn it. Sarah Friar's four-question scorecard, did it complete work that mattered and what did it cost including human review, is the most transferable artifact in the episode. Steal it.

The Builder. Varun Anand's Clay writing policy is the thing I'd ship Tuesday. Stand behind every sentence, writing is thinking, don't waste the reader's time. That's a governance doc that took a LinkedIn post to write and got eight thousand reactions because everyone is drowning in AI-padded PRDs and design docs. Adopt it, adapt it, move on. The harder build is Friar's finance reframe: if AI spend is opex that scales with utilization, then someone owns a utilization review cycle, and right now nobody does. Wire cost-per-workflow into your observability the same way you already track p99 latency. For a PM: you have a dashboard for how fast your agent responds. You need one for what each response cost.

The tensions. The Skeptic says deskilling is a forecast, the Researcher says the mechanism is airtight, and both are right: the mechanism is real, the timeline is guesswork, and by the time the invoice arrives you can't undo it cheaply. Second, the Compute Pragmatist wants tiered routing to cut cost, but every dollar saved by routing cheap models at junior tasks is a dollar that used to fund a junior human doing that task and learning from it. Cost discipline and expertise pipeline pull in opposite directions, and nobody in the episode reconciles them.

What this actually hinges on: whether token cost per workflow keeps falling faster than your usage grows. If inference gets cheaper 3x a year, the budget panic solves itself and tiered routing is a nice-to-have. If usage outruns price drops, the "AI is labor" reframe becomes the dominant procurement fact of 2027 and finance restructures around it. That's the one to instrument now. Run the boring test: put a cost tag on every agent workflow for one quarter and see whether your bill tracks headcount math or SaaS math. The answer tells you which budget line AI belongs in.

Prediction: By the time the major labs post their next round of enterprise pricing updates in Q1 2027, per-token frontier-model API prices for the flagship general-purpose models (GPT, Claude, Gemini tiers) will be at least 40% lower than their August 2026 levels for equivalent capability.

Confidence: Medium. Price-per-token has fallen every cycle, and competition plus cheaper inference keep compounding.

Why: The episode's core panic is enterprises exhausting AI budgets in months because token cost behaves like payroll, and NLW frames inference economics as the binding constraint on adoption. But the historical pattern cuts the other way: frontier model pricing for a fixed capability level has dropped multiples annually across OpenAI, Anthropic, and Google as labs compete and inference stacks get cheaper. Three labs fighting for the same enterprise seats plus falling hardware cost per token makes continued steep price cuts the likely path. The opposite, prices holding flat, would require the labs to stop competing on price, which none of them has done yet.

Revisit by 2027-03-31: We're right if flagship general-purpose model API pricing (per million tokens, equivalent capability) is down 40%+ from August 2026 at any of OpenAI, Anthropic, or Google. We're wrong if the best available price for that capability tier is flat or down less than 40%.

That price trajectory is exactly why I'd hold capital allocation steady, as Greg Shove argues, rather than re-architecting everything around today's token bill. The bill you're panicking about is a depreciating problem. The junior bench you're cutting to pay it is not.