weekend ai reads for 2026-05-22

But their operations depend on something just as fundamental: access to shared, finite resources. Water to cool equipment. Electricity to power servers. Land to house the facilities themselves. And, perhaps most importantly, the willingness of communities to host infrastructure whose primary benefits often accrue elsewhere.

Seven in 10 Americans oppose constructing data centers for artificial intelligence in their local area, including nearly half, 48%, who are strongly opposed. Barely a quarter favor these projects, with 7% strongly in favor.

Friendship, however, has always depended on a certain irrational generosity. A willingness to waste time together magnificently. To listen to the same anxiety for the fifth time. To sit through silence. To remain available without agenda.

Across thousands of these small interactions, what you can actually build without an AI looking over your shoulder gets a little weaker every week.

AI-assisted writing tends to be less hedging, less passive, and more specific (for instance, it includes more numbers). But on the measures that capture whether a reader can actually parse and absorb the prose, AI writing is worse. AI-generated academic writing uses longer words, more complex sentence structures, more jargon, and more nominalizations.

Where is the vibecoded Photoshop. The vibecoded Excel. The vibecoded Maya. The vibecoded Blender. The vibecoded compiler that compiles itself. The vibecoded database, the vibecoded OS, the vibecoded anything-that-requires-architectural-judgment-to-hold-together. Huh?

I am not asking for slop. Slop exists, slop is easy, slop is not the question. I am asking for the coherent, complex, non-trivial things that vibecoding allegedly makes accessible to anyone who can prompt.

Silence. Every time. The category is empty.

It is now 2026 and prompting AI models is very different from when ChatGPT first came out in 2022. Using AI well is one of the most impactful skills you can develop. And people that are not yet at a cutting edge of AI usage often run into AI generating frustrating outputs.

The skills of producing work and judging it were deliberately distinct, but accomplishing the work itself used to teach the judgment. The first skill now belongs, in large part, to the machines. The second still belongs to us, though fewer are bothering to acquire or utilize it.

The strongest demand signals revolve around seven needs:

Use-case prioritisation
Reference architectures
Standards and interoperability
Model benchmarks
Operational governance frameworks
ROI and funding
Organisation and talent

Al is now a leading cause of U.S. layoffs, and employers who use neutral-sounding criteria like “Al fluency” to push out older workers--or who eliminate their roles only to hand the work to younger employees running Al tools--may be building an age discrimination case against themselves.

Use case prioritization is fundamentally about navigating two competing AI adoption strategies: going for transformational moonshots or quickly accumulating many smaller wins to compound.

Most successful portfolios I’ve seen combine both, but the prioritization mechanics differ depending on which path dominates. There's no universal step-by-step approach to this but I found a few practical principles that help me structure the chaos.

Nearly 40 percent of students admit to regularly using AI for coursework in ways their instructor may view as inappropriate or against class policy.

1. Gather less of it

2. Keep it in your hands

3. Translate to human

Coding in the open has been shown time and again to produce high quality and secure work. The looming threat of AI vulnerability scanners doesn't change that - security is a shared responsibility. Technical teams need to be well enough resourced to create secure systems; hiding code is as reliable as papering over structural cracks.

This means that on the optimistic side (50 watts, 40 tokens per second, and 10 years) the pro max is as cheap as openrouter. On the pessimistic side (100 watts and 3 years at 10 tokens per second) the pro max is 10x the cost. I think ~3x the cost per million tokens is likely the right number for local inference on the pro max from an accounting perspective.

Speed of inference is the biggest factor here though for most cases. Local inference is slower than cloud inference. Some of the gemma 4 providers on openrouter get up to 60-70 tokens per second, which is 3-7 times faster than what I'm seeing with the pro max (~10-20 tokens per second). For a human employee with a work laptop, their salary costs are going to be ~1000x the cost of the tokens they can generate locally. Throwing money at anthropic makes more sense in this context.

Our experiments show that annotators who frequently use LLMs for writing tasks excel at detecting AI-generated text, even without any specialized training or feedback. … Qualitative analysis of the experts’ free-form explanations shows that while they rely heavily on specific lexical clues (‘AI vocabulary’), they also pick up on more complex phenomena within the text (e.g., formality, originality, clarity) that are challenging to assess for automatic detectors.