Enter your search text in the box above

Select a result to preview

DL 435

The Cost Interpretation

Published: June 3, 2026 • 📧 Newsletter

Hi all, welcome back to Digitally Literate.

This week keeps pulling us toward a plain but uncomfortable idea. AI is making interpretation cheap. Search, browsers, and workplace tools are no longer just helping us find information. They’re increasingly deciding what the information is before we ever see the source. That collapses the distance between asking and accepting, and it changes the kind of judgment we need to practice.

This issue focused on the source layer thinning out, the hidden cost of convenience, and what happens when friction stops being an inconvenience and starts being one of the few things standing between us and a mistake.

As always, the broader archive and connected notes for this newsletter live at digitallyliterate.net.

🏗️ Microsoft's Second Act

Microsoft Build 2026 was this week (June 2) in San Francisco). The themes weren't standalone products, they identified one architecture. Microsoft rebuilding its entire stack around AI, with itself in the middle.

Microsoft had the best seat in the house for the AI boom. It funded OpenAI as a research lab, locked down exclusive rights, and wired its models into Windows, Office, Teams, and Azure before anyone else moved. Then it lost the market anyway. ChatGPT became the consumer default, Claude the coding default, Cursor the developer default. Copilot became the thing people tolerated, not chose.

The answer at Build was a full slate of homegrown models. MAI-Thinking-1, Microsoft's first reasoning model, leads a family of seven models. All trained from scratch with zero distillation on licensed data, paired with custom chips for cheap inference. This means that Microsoft built these models the slow, expensive way instead of copying rivals' models, using data it had the rights to, and running them on its own purpose-built hardware so they're cheaper to operate.

My biggest takeaway from Build 2026 is that Microsoft is betting the AI agent becomes the primary interface to computing. The irony is that Microsoft built its empire on the interface (Microsoft Windows) that agents would replace.

First up, was Project Solara, an Android-based platform for devices that run agents instead of apps. No app store, no browser, the agent as the interface. Microsoft also revealed Scout, an always-on agent in Teams whose internal rollout plan, per documents 404 Media obtained), names its first phase literally, "Make people addicted." One employee called the framing "very troubling."

Why this matters: We talked about this a bit last week, but the big AI labs are all making bets on how we'll interact with AI...but you're definitely going to interact. This isn't AI bolted onto Windows or Office. This is building the computing interface around the model. The software fading out, and an agent that's "always on" and built to serve you and optimizing for something. The question is who is optimizing whom?

🔬 Show Your Work

In late May, OpenAI announced that one of its models had disproved a notable 80-year-old conjecture, one of roughly 1,200 problems posed by the Hungarian mathematician Paul Erdős. The problem, posed in 1946, is simple to state but had resisted proof for decades. A mathematical proof is a logical, step-by-step argument that demonstrates a mathematical statement is true.

The planar unit distance problem, first posed by Paul Erdős, asked if you scatter n points on a flat plane, and what's the largest number of pairs that can sit exactly one unit apart? Mathematicians had long assumed a tidy grid was the best you could do. The model found something better. Independent experts called the result genuinely impressive, and a group of them checked the proof and wrote up a human-readable version.

This week, sixteen researchers published the Leiden Declaration on Artificial Intelligence and Mathematics, endorsed by the International Mathematical Union and open for signing worldwide. It doesn't dispute the proof. It disputes everything around it. The model is proprietary. The prompts, the training data, the methods, the compute...all undisclosed.

Strip away the formality and the declaration is making a demand every teacher will recognize: show your work. Not just the answer, show the steps, the sources, and the method. OpenAI showed the world a proof and a promo video. It didn't show its work.

Why this matters: Math has run for centuries as a rare open system. Proofs are published, methods shared, and credit is given. AI firms are drawn to it precisely because the answers are checkable, which makes proofs an endless source of training feedback. The risk is a field shaped around what a model can win at rather than what matters. Knowledge becomes something built to be impressive, not understood. That's why "show your work" was never about catching cheaters. It's how thinking gets passed on.

💭 One More Thing

Here's a story with no product launch and no valuation.

Last November, a UPS cargo plane crashed in Louisville. The National Transportation Safety Board (NTSB) did what it always does. It opened an investigation and published the docket, including thousands of pages, a video, and a transcript of the cockpit recording. U.S. federal law forbids the agency from releasing the actual cockpit audio, out of respect for the people who died in it. So it didn't. What it did include was a spectrogram, an image of the sound, not the sound itself. For decades, that distinction held. An image was safe. A recording was not.

Someone on the Internet noticed the image still held the data, and used an AI model to reconstruct an approximation of the audio. The audio contained the last thirty seconds as the two pilots tried to working the problem as the plane went down. As the clips started circulating online, the NTSB pulled its entire public docket system offline and, as of this week, still has 42 investigations sealed while it decides what to do.

Why this matters: In this newsletter we keep coming back to a pair of questions: could and should. Could is capability: is this technically possible, can the tool do it. Technology answers that one for you, faster and cheaper every year. Should is judgment: is this right, who does it harm, did anyone consent. No tool answers that. A person has to stop and ask it.

For most of history those two traveled together. When reconstructing that audio took real expertise and effort, the difficulty was a pause. Long enough that whoever had the skill to do it was also positioned to wonder whether they should. The friction wasn't just a technical barrier. It was doing moral work nobody had to name. The NTSB never needed a rule against resurrecting the pilots' voices, because the effort required was its own protection.

AI didn't just lower the cost. It added distance. One more layer between a person and the thing they're doing, making it easier to look away from what the act actually is. It removed a safeguard everyone forgot was load-bearing. Now could arrives in ten minutes and should becomes optional, something a person has to choose to ask when nothing in the tool will make them.

So much of what protected privacy, dignity, and consent in the analog world was accidental. It ran on the assumption that no one could be bothered. As that assumption dies, we have to make explicit the protections we used to get for free.

💭 Consider

But he did not understand the price. Mortals never do. They only see the prize, their heart's desire, their dream... But the price of getting what you want, is getting what you once wanted.

— Neil Gaiman

🌿 The Understory

The themes this week are all about hidden layers as AI turns information into knowledge. Each layer used to be one that we could quietly inspect before closing.

Each is a layer you used to be able to inspect, quietly closing. And underneath all three is the same dissolve. Friction, expertise, and obscurity...the things institutions quietly relied on to govern information...are not working. What was once technically public but practically out of reach is now searchable, interpretable, reconstructable. The protections we counted on were never really protections. They were just effort no one expected anyone to spend.

So what do you actually do with that?

First, treat could and should as separate questions, out loud, especially when a tool makes the first one effortless. The pause is the work now. Second, when something arrives as an answer, ask to see the layer underneath it: the source, the method, the steps. Make "show your work" a habit you model, not a rule you enforce. Third, name the protections you've been getting for free. The friction, the obscurity, the difficulty, and decide which ones you now have to build on purpose, before something forces the question for you.

None of this was put to a vote. The interface, the incentives, and the collapse of friction arrived as products, not as choices anyone offered you. The one choice still genuinely yours is whether to keep asking the question the tools won't. Not whether we can, but whether we should—and who gets to decide.

If these reflections help you, there are a few ways to support the work:

See you next Wednesday on the other side. As always, my email is hello@wiobyrne.com.

Previous: DL 434 • Next: DL 436 • Archive: 📧 Newsletter

🕸️ Connected Concepts: