ai

My family first got online with Prodigy dial-up when I was 8. We connected to it via a local phone number, and so I thought that meant the offices of our local phone operator down the street was where “internet” lived. As I got older I began to understand it was far more complex than that.

One of my next-door neighbors growing up was a tall man named Craig. He spoke in a loud, but friendly, North Carolina accent that likely wouldn’t make you think he was the kind of guy who gets really into computers. With his characteristic strawberry-blond horseshoe mustache, he looked, talked and even dressed like a long-lost cousin of Dale Earnhardt.

But in the same way my dad just knew how to work on any car, Craig knew computers. Every so often he’d help my family out, upgrading our first desktop from a 2GB hard drive to 8GB and later helping us make the jump from Windows 95 to Windows 98.

I’ve been thinking about Craig recently because he was the first person I ever knew who worked in a data center. When I think about data centers, I still picture them exactly the way he described them: large, cold rooms filled with rows of whirring computers packed into racks and watched over day and night by people like him.

I can’t remember where Craig worked, and I may not have known at the time. For all I know, what he oversaw was a private network for a single company and not even part of the public internet. But the image sticks with me because it was likely the first time I understood, even vaguely, that rooms full of computers somewhere were doing something important enough that somebody had to watch over them day and night.

How Claude marks AI-generated content:

When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response.

Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.

Today Anthropic rolled out a new support page aimed at users in the European Union as part of signing the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. The new watermarking will apply to Claude models launched in the EU on or after August 2, 2026. Existing models are covered by a transition period but Anthropic says they’re working to add marking support to them too.

While I understand the concept of how this can work in theory, how it would actually work in practice is mind boggling. Even more so that this will apply even to API outputs from providers like AWS, Microsoft and Google.

How many characters will you have to use for something to be detectably Claude-marked? If I asked Claude to reply only with a single character, how can we be sure it’s marked? The answer is that we don’t know yet, and perhaps Anthropic has really just shipped an CYA page, especially as it pertains to marking older models.

Everyone seems to have an opinion on the effectiveness of “AI detectors” these days, but it will be interesting to see how well Anthropic’s works. If it’s any good, I expect other large economic powers will come knocking.

But I don’t see how the marking system they describe can be any good. Unless the Claude mark is complete gobbledygook (and they said it won’t be), there’s plausibly already text somewhere that would “match” this mark that wasn’t created by Claude. What’s the point of a mark if it’s more likely to cause confusion than no mark at all?

Imagine an author submits their original work to a publisher and they check it against an official Claude mark detector. Even a work with no use of AI whatsoever could may still not pass 100% because there’s nothing preventing someone from choosing the same words Claude might. And if doing so you choose to believe the author, why even check it in the first place?

So far it sounds like a system that solves for a very specific regulatory requirement, but no real problems. I’d anything, it might actually create more problems.

The case for open models just keeps getting better and better.

Jynn Nelson, writing for The “Inside Rust” Blog:

At the time of writing, there are 1,281 open PRs to rust-lang/rust. This represents a staggering amount of time invested by both authors and reviewers. We have long had the problem that there are more people who want to write code than people willing to review it. With the advent of LLMs, this problem only gets worse.

If you’re technically minded, the policy they agreed to adopt is an interesting read, but the overview has its own summary:

It’s fine to use LLMs to answer questions, analyze, distill, refine, check, suggest, review. But not to create.

Jon Udell:

I dislike the phrase “human in the loop” because it cedes authority to the machines. Let’s flip the narrative. It’s our loop, we work the same way we always have, now we recruit agents to join the team. An agent-assisted process need not be a black box that takes in prompts and emits features.

Jon made me realize how easily I accepted the term “human in the loop” in the first place. Of course we’re in the loop, we made the damn loops! Just saying “human in the loop” implies there might one day be a loop without humans at all, which some people are trying to sell but not many people want.

I tend to agree with Jon that “agent in the loop” is a more appropriate disclosure, but if AI was ever capable of doing what big tech said it could, we might end up saying “agent(s) in the loop.” Not because we rightfully acknowledge and wish to disavow machine-centric language, but rather that we would want to acknowledge when an AI agent was keystone reviewer of a loop. As of right now, humans are still better than AI at a large variety of important review tasks, and until that changes “human in the loop” is the better name because human operators are the most valuable part.

I’ve been a fan of Hank Green, and his brother John, since their “Vlogbrothers” series in YouTube’s early days. As well-meaning and well-spoken nerds, they rapidly assembled a community of like-minded terminally online followers. One channel quickly turned into a network, laying the groundwork for educational franchises like SciShow and Crash Course. But with a big audience and community, they’ve also made good. In 2012, the brothers launched a foundation aligned with their goal to “decrease world suck”, and since then it has granted $17 million to more than 100 charities. Earlier this year, Complexly, the company the Green brothers started to produce educational content for their channels and partners like PBS, became a non-profit after both brothers donated their equity in the company.

Which is why the recent fallout over Hank admitting to using ChatGPT for research feels like such a seismic internet event. Hank Green isn’t a faceless media conglomerate chasing a payout, he’s a beloved creator at the center of the internet’s fiercest debate about automation, scale, and burnout. Hank first addressed the issues in a reply to a tweet, but the original tweet to which Hank replied to was deleted, and later posted a longer Reddit apology, admitting he had fallen into an unhealthy habit of relying heavily on generated notes to keep up with his brutal publishing schedule. He announced a pause on several of his personal projects, conceding that “the level of dopamine I’ve been getting from interacting with LLMs… is not healthy for me or good for the world.”

If you interacted with my last LinkedIn post, I owe you an apology. It was 100% AI slop.

This post isn’t, but here’s why I did it.

In December 2025, Anthropic invited Claude users to sit down with an AI interviewer and share how they use AI, what they hope it could become and what concerns them about where it’s headed. Over 80,000 people across 159 countries participated, and Anthropic recently published their findings.

The findings are about what you’d expect. People want AI to do their boring work and they’re worried it’ll take their interesting work. The things they love most about it are the same things that scare them. Anthropic’s term for this is “light and shade,” which is a polite way of saying nobody knows if this is going to be good or bad.

This more or less describes how I feel too, and that’s pretty relevant because I was one of the 80,508 people interviewed in this study.

For a few years now, I’ve been chipping away at a historical research project about Will Rogers, the early 20th century humorist and one of the most widely read newspaper columnists in America. I find his story and his role in American cultural history fascinating, and as a fellow Oklahoman I feel somewhat obliged to help tell it. He died 91 years ago in 1935, so the vast majority of his life’s work has already rolled into the public domain, and the rest will soon. My project’s bottleneck has never been a lack of material, but getting access to it in the digital vaults where it is imprisoned. There are many paid-only, private databases that hold incredible, sprawling troves of public domain content but are barely indexed and have interfaces that seem almost designed to punish efficiency. Using them feels like the company is daring you to reverse engineer a better option.

The most profound insights about technology often come from direct experience rather than theoretical analysis. Last October when I gave my first-ever conference talk on agentic AI I emphasized process over code, specialized roles over general capability, and sequential collaboration over full autonomy. I was right about these architectural principles, but for entirely wrong reasons. The real limitations turned out to be more fundamental: accountability, security and an imperceptible line between capabilities and constraints.

I’ve spent a lot of time this year pondering AI’s impact on labor, especially in professional fields like mine.

While it’s a bit unsettling, history is clear: automation leads to obsolescence.