A more analytical batch: arguments and evidence rather than vibes, on what AI is actually doing to economies and organizations, starting with why a compute tax backfires.
Brian Albrecht’s A Compute Tax Is a REALLY Dumb Idea is a clear application of public-finance theory to a trendy policy proposal. As people panic about AI, ideas like taxing compute (pushed by Andrew Yang, called “sensible” by some economists) have gone mainstream. Albrecht walks through why it’s about as bad as a tax can be: Diamond-Mirrlees says don’t tax intermediate goods, and compute is becoming the intermediate input across drug discovery, fraud detection, logistics, and more, so taxing it shrinks the pie before anyone gets a slice. It’s also a capital tax (Chamley-Judd says don’t), it may even warrant a subsidy on Pigovian grounds, it’s undefinable in statute, and it wouldn’t raise meaningful revenue anyway. If you want to help labor, cut taxes on labor rather than adding one to a form of capital.
Jake’s Talk Is Cheap is the most rigorous skeptical piece I’ve seen on LLM value, and it’s genuinely uncomfortable. Drawing on Faros.ai telemetry covering 4,000 teams over 2 years, he argues that at scale, LLM use may be destroying value in software. Developer-level productivity is up (~16% PR merge rate, not 10x), but system-level metrics are ugly: deployment frequency down 11%, code churn up dramatically, defect rates up, and lead times up 80-480%. Using queuing theory (Little’s Law), he shows the industry has pushed its bottleneck downstream to code review: “organizations are generating more code than the system can review and merge.” His frame from operations: productivity ≠ throughput, and throughput is the stuff customers pay for. Whether or not you accept every calculation, the data deserves engagement.
Avery Pennarun’s The Software Industry: Annealing, but Wrong is the smartest counterpoint to the “always ship small PRs” orthodoxy. He uses simulated annealing as an analogy: mature systems benefit from small, careful changes (low energy), but sometimes you’re stuck in a local optimum and need a big jump to escape. AI makes big, interconnected changes cheap to produce, and while most are bad ideas you can now discard cheaply, some are good, and not every big step decomposes cleanly into 500-line PRs. His real argument: in a world where writing code is easy and reviewing it is hard, invest more in AI-assisted review workflows than in AI-assisted development. From the CEO of Tailscale, with concrete examples from their Aperture product.
Barn2’s The AI Generation Gap looks at how different cohorts (by age, role, and tenure) are adopting (or resisting) AI tools, and what that means for teams and products. It’s a business-oriented complement to the madewithlove exposure argument: the gap isn’t just individual, it’s organizational and generational, and it shapes everything from hiring to product design. Useful framing if you’re managing a team spanning that divide.
The Atlantic’s No, Artificial Intelligence Is Not Conscious is a philosophical corrective to the growing tendency to attribute inner experience to language models. The argument distinguishes between behavior that resembles consciousness and the actual presence of subjective experience, and cautions against the category error of treating fluent output as evidence of a mind. A measured read in a discourse that badly needs measured reads, and a good counterweight to anthropomorphizing language throughout the rest of these lists.