I’m going to confess something that sounds unhinged and, in fact, would have sounded unhinged to me until very recently: I built a travel website for robots.

In my defense, the bots aren’t the end user. They’re merely the middlemen – digital pawns in service of an eminently human goal: better international tourism. So I’ll get to them in a second.

First I’ll explain how I wound up here.

I was fortunate this semester to take an AI-powered data analytics course with someone who literally oversaw AI-powered data analytics for the CIA. Thank you, Georgetown.

I used the course exercises to advance my own personal project, the alternative-travel app at wanderingwell.site. I explain its mission here.

It’s an answer to the explosive growth in international tourism making travel miserable – communities are begging for relief from selfie-snapping hordes; governments are slapping on new tourist taxes; and poll after poll after poll shows travelers getting fed up.

Overtourism is fueled by several factors – including online platforms.

Internet algorithms feed off popularity. They feast off the math of the attention economy: clicks beget more clicks, fame compounds fame, and unholy masses ultimately pile around the same Instagram shot.

That’s why I’m trying to build an anti-algorithm. A formula that does the opposite of rewarding clicks – it actually penalizes clicks. Instead of giving points for popularity, I take points away.

I’m collecting a database of beautiful places – less-famous beautiful places. And I’ve kept adding capabilities: images, weather, directions.

The latest update adds clustering tools. People can search by theme, not just by destination. So you can look for places in wine country, or with hiking, religious shrines, easy family getaways, foodie towns, or archeological sites.

I built an image tool to rate hundreds of thousands of pictures

I’ve also added destinations – I’m up to 12,000 in over 230 countries and territories, and I’ll keep growing and refining the list as new data comes in.

It’s all powered by dozens of datapoints for each place.

Using the lessons from this past semester, I basically ripped through data for the entirety of Planet Earth. I started with about 5 million place names.

Places gained points for beauty and life – proximity to mountains, lakes, rivers, protected architecture, a UNESCO site, walkability, concentration of restaurants and coffee shops, and museums.

They also got points for being photogenic. I trained an image classifier to analyze pictures of over 200,000 places, granting scores to each picture, and including the score in my formula.

You’d be shocked, by the way, how easy this was. Anyone can do it with the latest LLMs. You basically build a website that lives locally on your desktop; you click through a couple hundred photos, you give each a score, type your reasoning into a box, and it feeds into a spreadsheet. You do this a couple of times, until your AI can predict the score you’ll give.

The model learned what I valued: natural features, old or colorful architecture, clustered together in highly walkable communities. The more of each, the higher the score.

Subscribe now

The ‘anti-algorithm’: How places lose points

And then I started removing points.

High hotel occupancy? You’re docked. High GDP versus the national average? You get docked. There are wealthy places on my list, but the goal was to steer visitors toward communities that could use more economic activity from travel.

Then comes the Wikipedia hammer. If the place has pages, where are the clicks coming from – from the local-language Wikipedia page, or the foreign-language page? I ran the percentages.

So if a small Tuscan town is getting a high ratio of clicks in English, it’s a good sign it’s being overrun in July and that hurts the score. The complete formula is here.

Now let’s talk about the robots. Or, to use the more accurate technical term, AI agents.

This site is an early experiment in agentic travel planning. By wiring its pages to AI research tools – and to something called an MCP server – I’ve made it possible for a personal AI agent to help someone actually research and book a trip.

The MCP connector is here. Just give it to a desktop AI app and you’ve got a basic travel agent.

Connecting it for AI trip-planning

How does it work? Here’s one example of how I just used it. I asked Claude Code: give me the cheapest, quickest flights from my current location to a place on this list that is warm and sunny over the Christmas holidays, with some museums and nice architecture.

It mapped out detailed itineraries and pasted a report on my desktop.

Again, it might sound a little nutty, but I’m on a work sabbatical, retooling, and if I’m trying to understand new systems poised to reshape the workplace, agentic tools are clearly in that mix.

Look no further than a ground-shifting development this week: Google essentially redesigned the architecture of the internet.

It announced the biggest overhaul of its search box in 25 years, as search becomes increasingly automated and Google’s AI Mode surpasses a billion monthly users.

Now Google is rebuilding its flagship product so that agents can watch the web around the clock – they will seek deals, flag price changes, and even make purchases for the brave souls willing to automate their spending.

(FWIW – That’s one area where I have no trouble being a late adopter: automated purchases)

Share A.I. For You

How the internet is about to change

Either way, the economics of the internet are changing, with major consequences for any industry relying on search-driven clicks.

This includes my longtime industry – the news business. And it includes anyone else who depends on being found online: publishers, local shops, e-commerce.

The question is increasingly less about “how do I rank?” and more about “how do I get an agent to choose me?”

And that’s why I added the MCP connectors. To experiment with what’s coming.

Tens of thousands of bots and agents hit my site the day I relaunched: within hours Meta, OpenAI, Claude and Google were all pinging around its pages – scraping info for training data, and running research errands for actual humans.

I led with a joke, that I built a travel site for robots. But, of course, all of this – the agents, the connectors, the machine-readable layer – that’s all scaffolding.

What I’m actually after is older and more human: better travel, for both people traveling and for communities on the receiving end.

In praise of humans

I want to end with a word for the humans who are the unsung heroes of this project, and of so much else in AI: Wikipedia’s contributors.

So much modern AI is built on their work. Models were trained on their labor of love – as volunteers, they wrote text that helped build LLMs.

And it wasn’t just text. Their painstakingly catalogued, openly licensed photographs fed the image classifier behind my site.

Their pageview data – which locations are getting clicks, in which languages – helped my formula understand how internationally famous places are.

As I built this site, I kept thinking of Jonathan Rauch’s book The Constitution of Knowledge: A Defense of Truth.

His thesis: today’s algorithms reward a toxic incentive that runs counter to the Enlightenment – its culture of testing facts, weighing empirical evidence, and discarding what’s untrue. These habits abetted modernity and democracy. Online platforms instead reward posturing: ignoring facts, digging in one’s heels, playing to the friendly crowd. That performance draws clicks, attention, and fame. It’s poisoning our politics, attracting bad people, and more, Rauch argues – it’s imperiling three centuries of vital intellectual habits.

His book, however, identifies a bright spot on the internet: Wikipedia.

He praises its peer-review system; its insistence on one page per topic – so there’s a shared reality; transparency, sunlight; and, when something appears biased, incomplete, or unproven, it’s labeled as such.

It’s the best of the internet, he argues. Now people are turning to chatbots for answers – answers built off Wikipedia.

So here’s my one request.

If you ever benefit from this website, and ever find yourself in a charming little town, in a pub that sees few foreigners, raise a glass for that editor. Better yet, if you ever meet a Wikipedia contributor, get the next round.

And maybe contribute. I have, and I wish I’d done it sooner.

Because, notwithstanding the headline of this post, I want the internet to remain human – for as long as it’ll have us.

I talked to two people in two vastly different professions last week who described a similar problem: How to store information.

Both wanted what I’ve called a living library: a way to plug AI into your notes for querying, researching, and creating — a knowledge-management system fit for our age.

I gave both the same answer: Have you heard of Andrej Karpathy’s setup? It takes, like, 10 minutes to set up for a single user. Like, copy-and-paste easy.

This ingenious hack is now being copied by tens of thousands, inspired by a post from Karpathy, a gentlemanly giant of AI research who regularly publishes online tutorials.

I’ll explain the setup, the history behind it, and why it’s catching on. You’ll see why Singapore’s foreign minister is a fan.

I’ll also note its limitations. And I promise to keep it simple. This is intended for a broad professional audience – I’ll avoid terms like YAML frontmatter, RAG, and vector embeddings that swarm this topic on programming forums.

Here’s the gist: Imagine a Wikipedia custom-built for you. It automatically updates entries on things you care about – be it people, projects, or ideas.

Related entries get hyperlinked, like Wikipedia. Say you toss in a complaint from a client. It will update three entries at once – one on the client, one on the complaint type, one on the product.

Maybe you paste notes from a book. It could be your own scribbles, or a bulk-export of your Kindle highlights (yes, you can do that). You’ll get updated wikis – on the book, the book’s author, and key concepts in the book.

Now imagine being able to query all this with AI, for any purpose. Hey, what do I have on topic X, Y, or Z? Draft me a briefing note, an email, slide deck, chart, visual dashboard or a training manual. You name it.

And you can, potentially, store the whole thing locally on your computer.

I recently wrote about the limitless use cases for your old notes now that AI can process megabytes of text – the living library. That post was platform-agnostic – and, admittedly, vague on how to connect the AI.

This is how.

I’ll explain how to set it up – really, it’s as easy as asking an LLM, with some caveats. But first, some background.

Subscribe now

The 80-year-old dream

In the final weeks of the Second World War, a towering figure of American scientific research penned an extraordinary essay foreseeing technologies still decades off.

During the war, Vannevar Bush helped create the modern U.S. scientific research apparatus, not to mention the Manhattan Project.

In a July 1945 essay in The Atlantic, he argued that research risked being swamped in information overload – an endless sea of data points.

Sound familiar?

Compared to Bush’s day, we now produce thousands of times more information. Just following political news feels like a fulltime job, never mind our emails and texts and social media and ever-updating dashboards.

Here’s something else that sounds familiar: Bush’s proposed solution. He described a machine called Memex, which would link associated items in a knowledge tree.

“Wholly new forms of encyclopedias will appear, ready-made with a mesh of associative trails running through them, ready to be dropped into the memex and there amplified,” he wrote.

Exactly – Wikipedia and hyperlinks.

While humanity made considerable progress in the ensuing decades, it was never that easy to drop it into the machine and amplify it.

Wikipedia was great for public knowledge. But personal and institutional knowledge was a different story, filled with friction, which explains why knowledge management became a multibillion-dollar industry.

Which brings us to Obsidian – a free platform launched during the pandemic lockdowns by two Canadian friends, both Waterloo grads. It lets you create notes then link them, like Wikipedia.

Its popularity, however, ran into two perpetual problems: 1) Who has time to be constantly updating and linking these wiki-notes? 2) What’s the payoff for all this work? It’s lots of work, with an unclear return.

Enter agentic AI.

Last year, Claude Code, followed by OpenAI’s Codex, and then Claude Cowork, began operating on your desktop.

You can now tell it: Create a file. Or, update a file. Where’s that file? How many files mention this name? Build me a slideshow from this file. And – poof! – it happens.

This completely flipped the work > payoff equation. Suddenly, managing wiki libraries is as easy as writing prompts.

Obsidian became a natural place to visualize them. You don’t strictly need it to create or query the wikis, but it offers three benefits:

–Readability: You click through links and read as you would Wikipedia

–Automation: You can pull new notes off the web automatically, and process them into wikis. It has a clipping tool.

–Portability: With the phone app, you can read wikis from anywhere.

How is this different, you may ask, from Microsoft OneNote? Here’s where Karpathy’s hack comes in – his idea for connecting it to AI.

OneNote is like a filing cabinet. This is a research assistant – one who has read every file; who remembers them; who updates them; adds to them; and then, on your command, can pull a slideshow together from details in 11 different folders. Based on one prompt.

Karpathy’s 10-minute hack

Here’s how easy it is to get started. Just paste Karpathy’s manifesto into, say, Claude Code and say: Build this for me. That’s what I did, although I requested some tweaks.

You’d also need Obsidian – the basic version is free, the paid versions are cheap. For access from your phone and for backups, you need to install an app and allow syncing.

This month Karpathy posted on X about a setup he calls LLM Knowledge Bases – within hours it had millions of views and tens of thousands of saves. A couple of days later he followed up with the copy-pasteable manifesto on GitHub, called llm-wiki.

The architecture is three layers:

• A folder of raw source documents – your primary materials. This is where you toss your notes, or the clippings you pull off the web with Obsidian. The folder is called /raw. You never touch this folder. It’s your archive; your artifact. Don’t mess with it.

• The wikis. Your AI-generated entries, which summarize everything in /raw, with crosslinks.

• Instructions to the AI. A one-page schema file tells the agent how to maintain everything – like the folder names, and the structure for each wiki.

As Karpathy puts it, the LLM does the grunt work: it summarizes, cross-references, and updates. You’re freed up for the deeper cognitive work: asking questions, critical thinking, deciding what to focus on.

The benefit, in his words, is “a persistent, compounding artifact” – wikis that keep getting better, and better, with every source you add.

I tweaked the setup a bit. Chalk it up to my old journalistic impulses. My schema file contains additional instructions, to guard against hallucination and credit sources.

My schema has a hard citation rule: for every claim, in every wiki, I want a direct quote and link to the original source material – so I can check it. When the wiki is synthesizing multiple sources, it must be flagged as *synthesis*.

I also chose my own filing structure. I have separate folders for /people and /concepts, each for my master’s program, and for my personal projects.

That’s just how it’s stored on my hard drive. It doesn’t really matter when I query Claude Code. It can pull notes from anywhere with the same query.

The use cases

So what do you do with this? People have called this setup a second brain, for its usefulness in myriad cognitive tasks like:

• Prepping for meetings. It’s easy to see why Singapore’s foreign minister wants diplomats to use it. Say you’re meeting a foreign official. With one query, the LLM could cross-reference wikis on the official; his/her country’s economic interests; your country’s interests; and historic irritants. You get a briefing note with a list of pet peeves, things to avoid, and a suggestion: You might want to bring up this electricity project in his/her home region.

• Customer insight. Say you have two clients – both constant complainers. Staff dread their calls. But you realize they’re not the same. Cross-reference their gripes, and a pattern emerges. Client A is a whiner. Client B is an early warning system, flagging problems before anyone else. You tell your team: Client B is an ally. Be nice to him.

• Political analytics. Ahead of the U.S. midterms, you’re writing a report on Trump’s tariffs. About which might go up, or down, based on his incentives. You have several wikis in hand: A list of competitive midterm races, major industries in those districts, key donors in those races, and a list of U.S. trade complaints. You ask your AI: Where do the interests of the Republican Party, and its donors, and the U.S. economy potentially converge? That’s Trump’s pressure point. If you’re a news team, maybe you send reporters to that district. If you’re a lobbyist or foreign government, you focus your energy on that tariff.

• Employee development. Your HR manager has wikis for employees – their experience, their performance, their career aspirations. Another wiki has a calendar of industry events. HR asks the LLM: Which events will motivate Jane, a talented young staffer? Turns out there’s a conference in Reno in October with three sessions dedicated to her intended career path. You call Jane with the good news: You’re going to a conference.

• Your speech. You’re giving a speech. Say the topic is AI sovereignty. You plan to mention the book Chip War, about the race to control vital inputs to produce new technologies. But wait! You hadn’t noticed a tension with another book you’ve read. This other book argues that the key to prosperity is not in producing new general-use tech. The key is using it. Pointing to three centuries of economic history, it argues that the winning countries are those that quickly adopt new tech across society. In other words, production matters less than diffusion. You only remember this book when pulling up a wiki called ‘AI Sovereignty’, which was updated by AI and includes notes from both books. You’re not sure which argument is right. But you adjust your speech to insert a counterpoint, adding balance and depth.

• Team coordination. You work for a consulting firm. Three colleagues are working on projects for the same tax company. Projects frequently overlap – without anyone realizing. It’s annoying. One surprise after another. It creates contradictory advice, unnecessary work, and, occasionally, embarrassment in front of the client. Now you have a wiki about the client. It tracks milestones and major interactions, and it’s shared across your office. From now on, when you make a move, it’s consistent across the organization. No more surprises.

I can think of a dozen other use cases off the top of my head – in journalism, politics, communication, consulting, education, legal work, and health care.

Now, the caveats.

The drawbacks

This system is not perfect. I’ll describe three risks and mitigating strategies. The imperfections are riskiest in sensitive and regulated sectors – it’s why I ended the last section with health care and the law.

• Context rot. The longer a conversation gets with an LLM, the dumber it gets. The concept is related to the well-known hallucination problem. I’ll give you an example. I’ve been using this wiki system for major updates I’m planning for an alternative-travel website. It got me off to a much faster – and better – start than with version 1.0. But over time, the project file accumulates junk. There’s now a Phase 1, 2, 3 – and a Phase A, B, C – and a Step 1, 2, 3, with contradictory instructions. So when you point your AI at the file and say, “Execute the second step,” the AI picks something unexpected. I’ve cleaned it up. Mitigating strategies: 1) A permanent source of truth. This is why Karpathy advises: Never touch the material in the /raw folder. It’s your original source. You can check it; use it to correct a sloppy wiki; or tell the AI to go into /raw and start a new wiki from scratch. 2) Citations, citations, citations. My schema has this instruction hard-coded: For every claim, provide the source and a direct quote. Where you’re synthesizing ideas, flag it and link to the sources. This not only sets guardrails for the AI – it later helps me confirm every output.

• Institutional use. For individual use, this setup is a snap. For organizations, some will find it harder to scale, especially the largest ones. You can’t have 15 colleagues tripping over each other’s edits to the same wiki. Nor can they be over-sharing confidential material. Mitigating strategies: 1) A shared folder for source material – the /raw folder. That shared folder might have certain vital information like, ‘We cancelled the pilot project,’ or ‘here’s the client’s new strategic objective.’ That gets shared. Then when you run your own personal chatbot query, the company-wide intelligence gets incorporated into your personal projects. 2) Test different tools. Notion, Confluence, and some Obsidian options might solve this shared-info challenge better than my own setup.

• Sensitive info. Some sectors require secrecy, for ethical, legal or regulatory reasons. A law firm or doctor can’t risk leaking client and patient info. If you’re running LLM queries willy-nilly into the cloud, you’re risking such a leak – and serious trouble. The good news is these wikis get built on your hard drive. You can also back them up safely, and access them via phone app with an encrypted sync. The trouble starts when you talk to the LLM about confidential material without knowing what you’re doing. Claude Code or Codex may transmit this content back and forth into the cloud. Lawyers and doctors obviously can’t risk this. Mitigating strategies: 1) Run a local LLM. That way the AI works with wikis on your hard drive. 2) An enterprise contract with an LLM provider. The bigger labs offer zero-data-retention plans; companies may evaluate these to determine whether they meet their security standards.

So you have to be careful.

But it works. It is a giant step forward toward an idea articulated long ago, just weeks after the Allies had booted the Nazis out of western Europe.

All these years later, Karpathy concluded his manifesto with a tribute to this idea’s spiritual forefather.

“[Vannevar] Bush’s vision was closer to this than to what the web became: private, actively curated,” Karpathy wrote.

“The part he couldn’t solve was who does the maintenance. The LLM handles that.”

There’s a growing pile of evidence that automation is bad for your brain. Letting AI write your emails and summarize reports apparently carries a cognitive cost.

The short-term cost is weaker neural activity, weaker retention, and work plagued by errors you cannot explain or defend. The long-term cost? That’s anyone’s guess.

New research is introducing terms into our vocabulary like cognitive debt, and epistemic debt, and cognitive surrender.

Here I’ll summarize what these studies say; what they don’t; and propose a mitigating strategy – for more mindful, more selective, use of AI.

Step one is understanding the problem. This concept of cognitive debt. To quote the old G.I. Joe slogan: knowing is half the battle. When people, especially young people, know the risk, they might at least make conscious choices.

That’s not to say I’d stop using these tools. In fact, this is actually a point of agreement among the researchers I’m about to cite.

They aren’t proposing we quit AI, since using it will be non-negotiable in some cases, and some fields. They’re proposing we use it better.

What these studies point to is the need for friction; guardrails that slow us down before we careen recklessly into the canyon of cognitive decline.

I’ll show you my own method below. In the meantime, let’s summarize the latest research on the scope of the problem. And, most importantly, let’s commit to discussing this with people we care about.

They need to know this. They should aware that a flurry of new reporting backs up a two-year-old Wharton study showing students who used ChatGPT for math under-performed peers when AI was taken away.

Study 1: Introducing the term ‘Cognitive debt’

The term cognitive debt was popularized by a newer MIT study titled, Your Brain on ChatGPT. It’s like a credit card: Buy now, pay later.

You can order any summary you want, any email draft, any graphics, quickly and easily, like walking into a luxury store and leaving with a designer handbag just by flashing a piece of plastic. The bill comes later, however. Then the collection agency starts knocking at your door.

This study tracked the debt.

Researchers asked dozens of participants to write essays in three separate sessions. Each time participants were interviewed, graded by AI, and plugged into EEG machines that monitored their brain activity.

Participants were broken into groups — one that used AI; one that used search engines; and one relying only on their brain. In a fourth session, study participants were asked to write an essay without tools.

The result? The better the tool, the lower the brain activity, with the worst-performing group being the one using AI, which registered 55% lower dDTF signal flow between brain regions compared to the no-tool group.

That’s not all.

People who used their own brain or search engines could all quote their own essay; yet 83% of the LLM group reported difficulty in Session 1, which briefly improved in Session 3 to a still-poor 33%.

Then their performance collapsed in the final session. In Session 4, they failed at a rate seven times higher than other participants in quoting what they’d written.

Now imagine this is a boardroom. You’re presenting a career-changing project, and you can’t answer questions on the fly. The collection agency has arrived.

Here’s what we don’t know: whether such damage is only temporary or long-lasting. The authors called for further study.

Subscribe now

Study 2: ‘Epistemic debt’ in vibe-coding

A Machine Learning scientist who was until recently a researcher at Amazon monitored 78 people using AI for computer programming – vibe-coding. The findings were released last month.

He tested a theory: that adding guardrails, friction in the process, forces thinking — and thinking forces real learning.

He made some participants answer questions about their project part-way through. Then he took away everyone’s AI, and had them fix problems in the code.

Surprise: people who faced friction earlier, those forced to answer questions, fared way better than those who just sailed through using AI, with a 39% failure rate in the final task, compared to 77%.

Author Sreecharan Sankaranarayanan calls it epistemic debt: you’ve created something you do not truly own. You may own it legally, but not cognitively, meaning you’ll flop when challenged.

His advice? It’s the same as the MIT study: Add friction earlier in the process.

He also suggests applying a framework: identifying different types of cognitive work, he urges programmers to be careful about which type they outsource to AI.

I’ll share my own general-purpose framework in a second.

Study 3: ‘Cognitive surrender’

Researchers at Wharton tested 1,372 people on nearly 10,000 tasks, in a paper released last month whose title plays off the famous work of decision-making psychologist Daniel Kahneman.

They gave people access to a chatbot. Crucially, they fiddled with the accuracy of different chatbots, to test who was over-relying on them.

Subjects used the chatbot more than half the time, and those who did relied on its answers 80-93% of the time.

Here’s the fun part: accuracy of the answers rose 25% when the AI was right and dropped 15% when it was wrong, which authors called evidence of cognitive surrender.

This means people stopped thinking, and relied blindly on a bot.

Worse yet, people who used AI, and who had most confidence in AI, remained confident as it kept screwing up.

Now here’s a modest silver lining. It connects to the framework below, and in fact connects to a common lesson in all these studies: Adding checkpoints to the process helped. That friction we discussed earlier.

Researchers gave one-third of participants incentives to improve their answers: 20 cents per correct answer, and an entry in a $20 lottery, and they received immediate feedback for each correct/incorrect answer.

That instant feedback paired with tiny financial incentives improved performance a bit — the override rate on faulty AI outputs rose from 20% to 42%. Accuracy grew.

The takeaway here is that there is no magic solution but adding cognitive checkpoints improves accuracy, as it did learning in the coding study.

My cognitive checkpoint

I’ve created a matrix that makes me more mindful about my own use of AI – it shapes my thinking about when to use it.

Before doing the ol’ <click> <drag PDF> <summarize> I ask myself two questions:

1. Will I have to own this knowledge? How likely is it you will need to understand, explain, defend this someday? Maybe tomorrow, maybe in two years.

2. Do I have time to do this without AI? In the pre-2022 era, would you actually have done this work? If so, maybe you still have the time.

If the answer to both questions is, “Yes”, then just read the damned thing. Or write the damned thing yourself. AI risks hurting you here.

If the answer to both is, “No,” that’s a perfect job for ChatGPT: low risk, and pure bonus — you weren’t getting to this anyway.

What if the answers are mixed, or unclear? Those are your edge cases — your dilemmas, the yellow zone in the above graphic.

To be clear, I’m not proposing you sit in front of a chart every time you open a chatbot. I don’t.

But I now systematically ask myself these two questions, and this encourages mindful use, one case at a time.

Here are a couple of examples.

• When to use AI: Low need to own, low time to read. In my final months in Washington, I would throw every new Supreme Court decision into a chatbot; request a summary; and ask for potential implications. Just to see if there might be a surprising story in there. I would never have read them all on my own. There is just no way I had time to read hundreds of pages of court decisions each day, atop my regular duties, especially given that a minuscule fraction of these cases meet the bar for an international news headline. I did not have to own this knowledge; I did not have time to acquire it. AI was simply a bonus here. It ran scouting missions for me — a tip service.

• When not to use AI: High need to own, have time to read. When I actually wrote about a court decision, I read it. I needed to know this. Readers and editors might have questions. And I had time to read it, if it was for an assignment. Take this landmark decision: Trump v. the United States, which clarified rules for when presidents can be charged with crimes. In the initial reaction, there was talk that the black robes gave a green light for presidents to commit murder. I read it, and spent days speaking to leading constitutional experts about the nuances. This case matters – it’s still shaping events in Washington to this day. I obviously read it. Now let’s consider a less dramatic example. This report: the U.S. National Institute for Standards and Technology’s Artificial Intelligence Risk Management Framework. Looks boring as hell, right? Yet this might be the most useful document for any organization looking to safely adopt AI. It is foundational. I’m studying this issue and want to understand it for years to come. I read it – it maybe took a strong coffee to get through, but I read it.

• Edge cases: Sometimes it’s a coin toss. A judgment call. Take podcasts. I don’t have time to listen to every one I’d like. Some are too long, technical, or peripheral to my interests. I’ll get a transcript, summarize it, and chat the AI questions before deciding whether to listen. Some, I truly want to savor. Like Trump’s former AI advisor appearing on Ezra Klein. These are two smart people having one of the best conversations I’ve ever heard about AI, but more than that: it’s about extraordinarily important topics in the field, from the Anthropic-Pentagon dispute, to the gathering conditions that could enable a police state. I’ll write about this soon. I did not want to outsource this conversation. Your own judgment has always mattered in deciding what to read, or listen to — it still does.

‘These aren’t moral failures’

One of the studies above quotes an eminent American-Dutch cognitive psychologist and expert on the science of learning, Paul Kirschner.

He’s described the difference between offloading and outsourcing thought. Offloading means a tool helps you think – a word processor. Outsourcing means the tool does your thinking.

Outsourcing, he says, has a cost. People who rely on GPS don’t navigate like they used to. Search engines make people remember less, and calculators erode arithmetic skills.

Kirschner insists all these tools have their purpose — even AI: “These aren’t moral failures,” he writes of using LLMs.

But there is a cost, and AI imposes this cognitive cost on a grander scale than arguably any previous technology in human history.

The cost: that our brains stop exercising, to employ a metaphor Kirschner dislikes.

“They become weak and dependent,” he writes. “Minds adapt to the workload they’re given. I hate to use the muscle analogy when talking about the brain, but if you’re sick and bedridden, your muscles atrophy from lack of use.”

All these studies I’ve referred to have something else in common. None of their authors have stopped using AI. In fact, AI is present in all this work. It’s used in all the studies I mentioned.

The Your Brain on ChatGPT paper uses an AI agent to judge study participants.

The paper is even written to get LLMs to read it carefully. The paper carries instructions telling LLMs which section to read first, which appears to be a clever way to force relevant context atop the context window, as LLMs tend to best remember the beginning and end of conversations — not the middle.

As for the coding study, it uses an LLM to query participants — in that friction point. Same for the cognitive surrender paper; it even mentions beneficial use cases for AI.

Kirschner himself writes: “AI isn’t going back in the bottle.” The real question, he adds, isn’t whether we’ll outsource: “The question is what we choose to outsource.”

In my own life, I’ve enjoyed this new ability to conduct research on a broader scale, and visualize it. To build apps and personalized tools.

But I do believe we need to be selective about how we use these tools. And, especially, to drill this awareness into young people. They have the most at stake.

What I’ve presented here is a crude instrument, an attempt to force more deliberation about AI. Of course, other factors come into play when deciding what to read.

Factors like: am I enjoying reading this? If so, maybe ditch the borg. The joy of reading — some things are just too precious, too human, to outsource.

With one prompt I got a long video promoting my alternative-destination travel website, wanderingwell.site. Would I show this to a boardroom?

You’ve probably heard of the AI-generated podcasts you can make with Google’s NotebookLM – helpful if you’re in a hurry, maybe driving to a meeting, and you need to get briefed on the go.

It’s now graduated to video.

Last week Google introduced new cinematic video capability to premium subscribers: you can now turn your notes, web links, and photos into a video.

I just tested it. I generated a video presentation for my alternative-travel website, to explain my concept of an anti-algorithm algorithm.

The verdict? Impressive, as a technological step forward. Useful, in certain specific cases. But not yet ready for high-stakes deployment. Grade: 7.5/10*

I do think it’s useful as an informal educational aid. For instance, say you’re trying to engage a young student who stubbornly refuses to read about a certain topic; I’d resort to this tool as a last-ditch solution before giving up the lesson.

I can also see other professional or business uses in low-stakes cases. But this is not something you’d bring to a boardroom. Not yet.

Harsh? Maybe a tad. After all, I did just get a 4-minute video with a simple prompt. In fact, I made several versions. This would have been unimaginable until recently.

But the bar, in my mind, has been raised by what you can now do for PowerPoints. You can, in fact, now use Google, and Claude Code, to create charts and graphics you would actually take to a boardroom.

You can create presentation slides by dumping a folder of data into Claude Code and saying… Make a slideshow. It even started making interactive graphics.

So this is my new gold standard. In fact, if I wanted to make a narrated video for a presentation, I’d generate the slideshow separately and just record my own audio track for it.

The one thing this new NotebookLM tool offers is short bursts of movement: simple motion which you don’t get in a regular slideshow.

This is a step up from the previous offering from NotebookLM, a video overview tool launched a few months ago, which had static slides of modest quality.

This new version offers upgrades including bits of motion, and some other features I’ll discuss, but it comes with a hefty price tag.

For now it’s offered only with Google AI Ultra subscriptions — which cost $249.99 US per month, after a three-month half-price promotion.

Is it worth it? You decide. Here’s the rundown.

• Watch: Here’s what the slightly older — free — version created

Compare and contrast. This is what I got from the NotebookLM tool created a few months ago. It narrated slides, without motion. It’s less sophisticated, but much cheaper, and, in some respects, I liked parts of it more.

Subscribe now

How it works

First, I created a project on NotebookLM. I gave it a bunch of files, pictures and text explaining Wandering Well, the alternative-travel site I’ve been building out.

Then I asked for cinematic video with this prompt: Introduce my alternative-travel website called Wandering Well. Based on an anti-algorithm algorithm. Make sure you get images from the files called MUST USE - these include the main home page for my site; ideal destinations my site promotes; and the 65-35 methodology explanation. Use these as images. I would also love to include the main image atop the website labelled Try My Hack for a Better holiday as this lead image shows the before and after - places I’m trying to avoid and places I am trying to promote. The info you will need is mostly in the two text files: the about page for my website and the llms-text page. This video will introduce my concept of an algorithm that rewards the beauty of a destination but penalizes popularity based on economic metrics. My formula looks for places that are gorgeous, but that may welcome and need more tourism. To help spread tourists’ focus to new places, making for a more equitable distribution that makes everyone happier: the traveller, the overcrowded celebrity sites, and the less-discovered gems that could use tourism. My about and llms-text files include details on the data sources and methodology. There is even a graphic of the methodology and its 65-35 formula. There are also photos of sites on my list, and an AI-rendered scrapbook comparing horribly overcrowded sites with a pleasant piazza - the lead image on the Try my hack for a better holiday site. Use all these to tell the story in a video. Make the narrator a woman with a British accent.

I requested the woman with a British accent because I didn’t like the default voice on the older slideshow version – it sounded to me like an NPR host on amphetamines.

For some reason, this request half-succeeded, which is better than the video-overview tool of a few months ago.

Ignoring my request, that older tool used the hopped-up NPR guy. This one gave me a female voice, albeit without the accent I requested. So a half-win.

The good, the bad, and the cringey

What worked well: I got videos. Each identified key lines and concepts from my material. I thought the tool of a few months ago identified better turns of phrase, to be honest.

The pronunciation has clearly improved. The older version was a disaster with non-English names. It flubbed Campeche, Mexico, and Rize, Turkey. Cinque Terre, Italy, was pronounced Sinkway Tear. Ma che cavolo! The new version got Rize right and nailed Sciacca, Italy.

What also worked well is the new tool actually used more of the images I demanded. More of the MUST USE images wound up in the video.

What worked less well: It just flat-out refused to upload certain images, and I’m not sure why. Maybe over copyright concerns. But the images it refused to take were mostly ones I’d created with AI. Can’t explain this one.

It also took liberties with substance. NotebookLM used made-up numbers to illustrate some graphics – I never gave it the stats it used on overtourism. It exaggerated something I’d written about lavender fields. And it referred to a scatterplot I did not create.

The graphics are clearly more sophisticated – but not yet of professional quality. The opening scene looked ripped off from the Apple 1984 commercial. Taste is subjective, but let’s just say I’d have gone for something slightly less apocalyptic.

Bottom line: Not ready for the boardroom.

If you have a creative team, your team will easily outdo this current tool. Note the qualifiers in italics. Not everyone has a creative team; and these current tools will get better.

For other limited use cases, this is helpful.

So my takeaway from all this is? At the risk of repeating myself from the other day: AI output still needs editors. And it will keep needing editors for the foreseeable future.

*This score is an improvement from the tool of a few months ago. I originally tested that video-overview tool and gave it a 7. This was better in some ways, but not all

I’ve been thinking about editors lately. About my own editing work, and the editors I worked for.

I’ve been thinking about the great ones, like Malcolm.

This guy could smell bullshit from another area code. As a cub reporter, I’d talk to him about a story. He’d bombard me with a dozen questions: How will this person react? What’s the effect of x on y? What does the law say? Have you reached out to so and so? I’d grow annoyed. If that annoyance registered, he never showed it – instead he operated with the quiet confidence of a veteran who was perennially three steps ahead of me.

He spotted fallacies. He predicted consequences and foresaw fallout. And he knew that the red flags he raised for me at 9 a.m. would inevitably become the problems I’d be struggling with around deadline time eight hours later.

I’m grateful for everything he taught me. Because we will need these lessons for what’s coming.

So here’s my thesis today: Editing will be a superskill of the AI era, and not just editing but a related set of cognitive habits centred on critical thinking, which could easily include philosophy, the law or liberal arts.

I’ll list specific habits of successful editors in a moment. But first I want to elaborate on why I feel this way — why I think, “Learn to code,” will become, “Learn to question.”

Clearly I’m not alone in expecting this shift. AI leaders as diverse as Peter Thiel and Daniela Amodei are talking about how “word people” and humanities grads may hold more coveted skill stacks in the AI economy.

I’d like to believe this explains at least to some extent a recent pattern in Google searches: after decades of precipitous decline, searches for editors have begun inching up in recent months.

Look at this. A study that finds only workers with certain mental habits gain creativity from AI. The study calls it metacognition — I call it looking at ChatGPT outputs and being able to tell the difference between, This is great shit! and What is this shit?

Here I’ll provide specific examples from new research about why I feel this way; connect it to personal anecdotes from several months of heavy AI usage; then list those good editing habits.

Let’s begin with this new paper: Even the best AI models, including a model released just days ago, stall at only 34% accuracy in analyzing a big dataset. That’s great for a baseball batting average — but a disaster for a professional product.

These models were given U.S. Treasury bulletins going back a century, 89,000 pages, and asked 133 questions which they mostly flubbed.

Okay, let’s assume these models keep getting better. So mistakes become rarer, harder to catch. Guess what: You’ll still need someone catching them!

Now let’s take it a step further.

Let’s assume these models become perfect – let’s say we reach Artificial Super Intelligence and they wind up nailing every question, even in real-time situations that require advanced future forms of Retrieval Augmented Generation.

Spoiler: You’ll still need editors. Even in this scenario. In this perfect-genius-robot future, the world will be utterly drowning in content. Organizations will need help managing that content.

Malcolm would do this during breaking stories. He’d have reporters calling him from across the country, dictating tidbits and quotes, thousands of words of content that, if published as is, would make for an incoherent and interminable news story.

He had to assess this material and decide: a) What to keep for the story b) Where to place it in the story c) What to save for another story d) And what to throw away.

Organizations will need this role. Traffic cops for content.

The most successful of these will be people possessing the sweet spot of three distinct skills: subject-matter knowledge + AI fluency + critical thinking.

I’ll give one example of this winning trifecta. I’m thinking of an enterprising law firm that creates some sort of app called, hypothetically, Pocket Lawyer. I suspect we’ll see very successful businesses do this.

I suspect these apps will significantly expand the market for legal services. Rendering them more affordable, and more accessible, to more people. But you can’t file hallucinated crap in court or with a regulatory agency.

So the app generates an early draft of legal documents, but you would still need human lawyers to do the final… ahem… edit.

What I’ve been experiencing

The newsroom nostalgia has been growing within me since the fall. As some may know, I worked in journalism for 28 years before going on sabbatical, during which I’m getting a master’s in AI management.

In recent months I’ve built an anti-algorithm algorithm for alternative travel, which looks for beautiful-but-less-popular sites in 194 countries; a daily briefing note on developments in AI, translated into language for non-technical professionals; a stock-market dashboard with 140 years of market data; a personalized movie-recommendation-algorithm; and a bunch of video games.

And the point I want to make here is: You have no idea how many screwups I averted. Some big ones, too!

Some were basic. Take the travel app. It took me weeks to build that app, with AI tools last year that weren’t as good. I’d think it was almost ready, then I’d run some spot checks, and I’d be like, “Hey, why are the Spanish regional unemployment numbers different in my app’s master dataset from the ones in the spreadsheet I gave you?”

And the AI would be like: Oh, yeah, made that up. There were days I wanted to fling my laptop across the room.

Suddenly, a couple of months ago, we crossed a threshold. The mistakes grew rarer. As the tools improved, the success rate grew. This is part of what drove my conversion to Claude Code, powered by Claude’s Opus 4.5 model.

But guess what? Claude Code still screws up. And this is a critical point: it screws up in harder-to-catch ways. More sophisticated tools, it turns out, make more sophisticated mistakes.

I’ll give you a near-comical example – but first, a quick primer on Machine Learning. The basic idea of ML is you feed data to a machine (your training data). The machine looks for patterns in the data, then makes predictions. You test those predictions on a separate dataset (your testing data). You measure the average error size. Say you’re doing weather predictions — you miss by 4C on the first test. You run a second test, and miss by 3C. Then it’s 2.7C on the third run. You keep repeating until the error stops shrinking. That’s Machine Learning. It’s the loop that powers many tools you use.

Claude Code and other new models can do this. You can literally say: Do Machine Learning to predict the weather. It will download a ML library like scikit-learn and find the required climate data, and start running projections.

So what I did the other day was point it at data to measure the effects of AI on the job market. I wanted to know: By age group, education level, industry, and geographic location – what is AI doing? I uploaded data from the U.S. Bureau of Labor Statistics.

Claude came back within minutes with extraordinary charts. Troubling charts, dramatic charts, showing a plunge in postings for young professional positions. It even produced a lengthy briefing note.

Then I asked: Hey, if a data scientist or economist wanted to test these numbers, are they reproducible? Give me a step-by-step guide to how you got here.

And it admitted: This. Was. All. Junk. Well, not in so many words, but what it did was combine news articles – news articles! – with the BLS data to make completely non-scientific projections. Needless to say, Machine Learning does not involve tossing Forbes pieces into your dataset.

We started over. This time it produced step-by-step documentation, with far less dramatic results. It also apologized profusely.

The point here is that the errors are increasingly hard to detect. This is not the ChatGPT of 2022 hallucinating a quote – where you can check Google and see instantly if it’s real.

The world is about to be submerged in this crap: synthetic data everywhere, creeping into news articles, influencer posts, LinkedIn, X, academic studies – some of it accurate, some of it not.

It’s going to contaminate our information ecosystem, and our workplaces. And I repeat myself here: even if this problem disappears, which I doubt, we will still encounter the problem of information overload – we will all be editors, at that metaphorical desk, playing traffic cop with an endless stream of content.

In apologizing to me, the AI itself, Claude Code, essentially admitted: I need an editor. So what exactly are these things an editor looks for?

Subscribe now

Things an editor asks

1. What’s the story? The world produces 400 terabytes of data per day and it’s growing fast. We are drowning in detail. Nowadays, you could theoretically create transcripts for everything your organization does: every call, meeting, sales record, document, and use it to create new stuff. Then key is deciding: What do you want to communicate from this? Put differently: What’s the story? The story you want to tell your staff, the public, or yourself. That’s your starting point. Understanding that objective guides what questions you ask the AI, what documents you preserve, and what you ultimately create. Take my stock-market dashboard – I wanted a clear takeaway atop the dashboard. I had gathered a bunch of metrics that historically foreshadow market shifts. Originally, it was a muddled mess. I deleted data that cluttered up the page without adding value. And I picked that bright red font with the big number at the top – to emphasize the key takeaway. In news lingo, that was the lead (The lead is… we’re due for a major stock-market crash, if history holds). I pushed the supporting details lower down the page, and deleted what I didn’t need.

2. What’s your source? Show, ask, check. You know the old expression: Garbage in, garbage out. Put another way: You are what you eat. AI eats data – numbers, text, and with those it creates outputs. Here are three ways to encourage healthy sourcing. One is transparency. Show your source. In news stories, we attribute them: he said/she said/the document said. AI products should have similar sourcing disclaimers. This helps your reader assess the quality, and there’s a simultaneously benefit — it creates pressure on you: you, the journalist, or in this case the AI user, know you will be scrutinized. This forces an incentive to seek quality sourcing — built-in pressure toward best practices. But sourcing isn’t always clear. The second way to encourage quality sourcing is to ask the machine. Hey, where did you get that? A good editor sees when stories have no sourcing, or anonymous sourcing, and asks: Who told you this? Do the same with your LLM. This allows you to check the record. With the sourcing info handy, you perform the third and final step: double-checking the source material. My old instinct to check has saved my skin a few times lately – otherwise I might have published a travel app with made-up data and junk charts on job losses. Show, ask, check.

3. Consider the counter-argument. Editors learn to do this in their sleep. It’s drilled into political reporters too: If I write X, someone will argue Y. Guess what? Sometimes Y is right and you’re wrong. Test your hypothesis before publishing – not after. I recall the advent of political blogging two decades ago. In some ways, it was great for journalism. We started getting grief for every stupid little sentence we wrote. Bloggers would rip our stories to shreds. This never happened before and, frankly, I can’t say it was fun. But it made us better. It made reporters think twice before pulling a hypothesis from our posteriors. So remember: Is there another possible explanation? If you aren’t sure, say so. Take this blog post, for example. See that factoid about rising Google searches for editors? There’s a reason I adorned it with a Christmas tree’s worth of qualifiers.

4. What’s the baseline? Ah, the dirtiest trick in journalism: cheating on the denominator. Before you present a measurement, you must choose what to measure. And here’s a nasty little secret: you can make pretty much anything look good, or bad, depending on what you choose. Carbon emissions. Crime stats. Middle East hostilities. Take your pick. You can draw different conclusions based on your chosen baseline. Compare current crime to a high-crime year (1991) and the world is a safe, cozy place; compare it to a low-crime year (1955, 2015) and you have people putting bars on the windows. All with the same stats! People do this all the time — with GHG emissions, and pretty much any contentious topic. They compare current data to their favorite baseline. Good editors know this – and they know to press their reporters on it. This is highly relevant when presenting research using AI. Again, let’s use the Google Trends chart on editors as an example. I could have told three different stories, based on my baseline: A positive one, a negative one, or a complete one. Let’s say we compared today’s search data to 2024, you’d think searches for editors have soared to dramatic highs based on the last year. But zoom out to 2004 and you’ll see we are still mired in a long-term historic trough, just barely starting to climb back. I tried doing the responsible thing here. I searched for both short- and long-term timeframes — and I mentioned both. I tried telling the most complete story I could. Do the same with your machine: Test its output across different time horizons. Say you’re creating a visual dashboard. You need to know what comparisons to make. The best practice is to test different baselines.

5. Did you keep your notes? Documentation. This is to 2026 LLM use what prompt-writing was to 2022. It’s part of the shift to agentic AI. Now I’m rediscovering old instincts — the old instinct to collect notepads. For years, the advent of Google trained me to assume I could find most facts I needed online, so I became less methodical in preserving notes. I’ve now snapped back to old habits. The cause was Claude Code. Agentic AI will produce himalayan heaps of data from your command terminal, but it has a defective memory. You’ll be like: “Hey, we need to adjust the website. Go back in, use the login, and fix,” and agentic AI will be like, “What website?” So I create folders for every project. For any project, my very first request is to create a file in that folder called something like BRIEFING NOTE. I keep updating that file. This file winds up logging every major action from every session, and describes our workflow. This way I can start a session and say, “Here, fix this website,” and I drop the file into the chat. Here’s another benefit to keeping records, which has nothing to do with speeding up your prompting: Fact-checking. Proper documentation lets you scrutinize your work later, as illustrated by my recent episode involving jobs data. Records allow testing. Testing allow accuracy. All of it enables transparency.

   

This brings me back to journalism. With a lesson I learned as a youngster, about consequences. And about preparing to face them — with documentation. You can be sued for what you write, and if that happens you may need evidence in court. If you’re sued for libel, there’s no better defense than pointing to the tape and saying: Actually, here’s exactly what you said. You’ve kept a record.

I think I put Malcolm through a few of these moments. And I’m forever grateful for what he gave me back – for what he, and other editors, taught me. We’re going to need these lessons, badly.

What timing. The biggest political-AI story, maybe ever, blew up last week just as my college class was learning about a pattern of deception from sneaky Sam Altman that predates the creation of ChatGPT. Hence the chuckles roaring in my head when I heard him claim to be benevolently working to resolve the impasse between the Trump administration and his rival Anthropic. He resolved it, all right. By wiggling his way into a Pentagon contract after Anthropic refused terms it suggested might lead to catastrophic consequences for humanity: AI-enabled domestic mass-surveillance and fully autonomous weapons. The Trump administration has now blacklisted Anthropic, the creator of Claude, and threatened to effectively destroy it.

But the public reacted much as I did: with an impulse to reward good actors.

Claude surged to the top of the Apple Store, for the first time ever, vaulting over 100 spots to No. 1, surpassing ChatGPT in downloads as people voted with their index fingers. Eager to capitalize on this sudden vibe shift, Claude released a simple two-step guide to transferring your data from ChatGPT. I’ll explain in a second what you can do – and what you can’t. But first some background on what has happened.

Let’s start with Altman’s history. A former member of the OpenAI board, Helen Toner, now at Georgetown, where I’m studying, has described why the board tried ousting Altman; she referred to multiple examples of gobsmackingly devious behaviour – concealing his financial interests in OpenAI when he claimed to be an independent board member serving a non-profit researching AI (try not to laugh here) in the public interest; misleading the board several times on safety processes; lying to board members to try ousting her; and not informing the board he was about to launch ChatGPT – Toner says: “We learned about ChatGPT on Twitter.” After years of this, she says, they decided they couldn’t trust him and tried firing him. She said this fit a pattern: In his only other roles in Silicon Valley, he was fired once, and nearly ousted two other times, she said, over the same pattern of duplicity. “If you actually look at his track record, he doesn’t, you know, exactly have a glowing trail of references,” Toner said in a 2024 interview.

Fast-forward to more recent events. Altman claimed he did not want ads in ChatGPT chats (yet here they are); he would not enable porn (oops); AI should be regulated, he says! (his co-founder has launched an anti-regulation SuperPAC); he also donated $1 million to Donald Trump’s inaugural fund, which is nothing compared to the checks his co-founder Greg Brockman has been writing Trump – to help humanity, he says. Now Altman says no mass-surveillance or automated killing machines, pinkie swear.

ChatGPT has friends in high places

OpenAI unquestionably has better connections in Trumpworld than Anthropic: Jared Kushner’s brother invested $1 billion in it, and Larry Ellison signed a $300 billion deal with OpenAI.

But these guys all have a problem – fear the OpenAI bubble is about to pop. Anthropic’s Claude is fast catching up in use by businesses as many people (self included) determined months ago that it’s a better tool. A far better tool. Some in Trump’s orbit started expressing disdain for Anthropic months ago; they began publicly trashing the firm, which has a constitution that’s supposed to guide its behavior and a virtue-ethics philosopher as a senior executive.

Who knows. Maybe none of the financial entanglements and personal animus played a role in the Trump-Anthropic-OpenAI triangle. Maybe it’s not, in fact, Exhibit A of a comically corrupt administration tilting the playing field to enrich its friends and crush rivals. A deep dive from The New York Times points to several factors: personality clashes, Altman’s behind-the-scenes role, and Anthropic’s insistence on a legally binding commitment not to use its AI to piece together information on Americans using widely available commercial information. And that was that.

By Friday night, President Donald Trump called Anthropic left-wing nut jobs and said U.S. government agencies would stop using it in six months; the Pentagon declared it a supply-chain risk; and Trump’s former AI adviser, aghast at what was unfolding, called it attempted corporate murder of a promising U.S. company and urged AI investors to look outside the U.S. The next day, OpenAI released some details of its contract: it contains language filled, critics assert, with loopholes the U.S. military could use to profile Americans and deploy autonomous weapons. OpenAI insists safeguards are built into the technology itself, like forcing the Pentagon to use its cloud platform rather than incorporate the AI into its own tools. Altman admits the announcement didn’t look good, and was a little rushed – because, he says, he was in a hurry last week: “We wanted to de-escalate things.” What a humanitarian! Elon Musk’s xAI, meanwhile, was pretty quick to reach a pact with the Pentagon.

So there you go. Make of all that what you will.

Anthropic, by the way, deserves hard questions of its own here. Questions like: If you view autonomous weapons and mass-surveillance as crimes against humanity, why would you sign contracts with, of all companies, Palantir and Donald Trump’s Pentagon? Should you not have foreseen this dispute? Are you not contributing to the problem by watering down your own safety policy? Those are all fair questions.

That said, the company deserves credit for drawing red lines and standing by them. Clearly, there was public appetite for this. This is a moment of intense distrust – of both AI systems and of government abuses of power. This episode involved both; Anthropic showed spine. And the public responded – giving Anthropic’s Claude a record-smashing weekend. I got caught up in it too; given that I’m studying this stuff, I’d been subscribed to multiple tools. Over the weekend, I cancelled my ChatGPT subscription; to be frank, I hadn’t been using it much anyway, because I have found Claude, especially on my desktop via Claude Code, demonstrably better. The only thing I liked about ChatGPT was its workhorse quality – Claude is more expensive, and comes with more usage limits, so I would sometimes dump larger but less sophisticated projects on ChatGPT – “Here, turn this 200-page PDF into text”; like I said, as a workhorse. A mule. For anything more sophisticated, I was already using Claude – I even wrote about it.

Breaking up with ChatGPT: Steps 1, 2, 3

Look, there are limits to your power as a consumer. If ChatGPT wants to help governments analyze individuals based on commercially available data, there’s not much you can do retroactively. If you’ve allowed it to train on your chats, again, it’s already used your old chats to train the next model. Oh – they don’t go out of their way to tell you that: OpenAI uploads your conversations to train future models unless you know to click here on User → Settings → Data Controls → and turn Off the setting called Improve The Model For Everyone. This ain’t exactly a first-ballot candidate for the transparency hall of fame. Sneaky, you might call it. Fortunately, there are steps you can take. Including deciding what sorts of companies deserve your business:

• Shut off that function that lets it use your data for training. OpenAI’s setting is here. Claude’s is here; Claude, by the way, is only marginally better. Claude has also started training on user chats by default, but is slightly more explicit about it; it at least informed users of their options when it introduced the policy last year. Google Gemini’s is here.

• Delete old chats. ChatGPT allows it here. Claude does it here. Google allows it here. You’ve probably already been trained on, but there’s the option if you’re worried about leaks.

• Migrate your details to a new provider. This is what Anthropic published over the weekend. It’s a prompt to ask what any LLM knows about you. You might be surprised by the volume. Anthropic is inviting people to transfer their info here.

Video: How to teleport your data from ChatGPT in 30 seconds

Again, I want to set realistic expectations here.

There are limits, as I said, to what difference any of this makes. As there are to lionizing any individual company.

But at this moment in time, involving this technology, at this point in political history, I’d rather reward certain behavior. Good behavior, courageous behavior. And send the sneaky types a message – here, I’m voting with my index finger. So you can see my middle one.

I pointed Claude Code at 90 State of the Union speeches — every address from FDR to yesterday's from Trump — and asked it to find patterns. Some surprised me, some didn't.

I'm sharing these as an example of what I wrote about yesterday: How every institution, not just the White House, should be preserving textual data in order to find patterns, longitudinal trends, gaps, opportunities, customer sentiment and potential new products based on those.

Some quick takeaways.

1) The Pronoun Shift: Trump was by no means the most self-referential president. Some might be surprised to know that in his first term, he was actually about average in his use of the first person singular in these speeches - in fact he used it far less than Joe Biden. But he is now trending hard in one direction: toward “Me”, “Myself” and “I”. Check that ratio of I/Me versus Us/We over time.

2) The Words Where Trump Leads: I tracked 20 terms, from “economy,” to “jobs,” security, living costs, immigration, the environment, crime, I/me and other terms. I compared Trump 2026 to past speeches. He outperformed the historical average in using these terms: “I/me,” “border,” “illegal,” and “crime.” On all 16 others including the economy, science, infrastructure, the cost of living, he was well below average or at zero.

3) The Limits of Word-Counting: 90 speeches is a relatively small dataset. Presumably you have a much richer one sitting on your hard drive that you could do much more with. I definitely found some interesting historical patterns — references to “poverty” skyrocketing during LBJ’s Great Society, “allies,” as JFK led at the hottest point in the Cold War, and the “environment,” as Nixon created the EPA at a time of new clout for the green movement. But there are blind spots.

FDR barely mentioned poverty in his SOTUs — yet he built the social safety net at the height of the Great Depression, as the architect of the New Deal.

A lesson here? Don't assume you have all the answers in 90 text files. More data equals more clarity, and even there, it's important to recognize the limits.

A note on sources: I used the live transcripts posted on Roll Call for last night and UCSB's American Presidency Project for past speeches, and performed word counts with semantic embeddings to recognize, for example, "environment" and "environmental" as the same word. I pulled this into CSV files. Then tested the stats against the transcripts. They were accurate. If anybody would like the CSV files, just ping me. Happy to share them.

I’m turning into a hoarder. A hoarder of text. I might have thousands of pages of typed notes saved in a database by the time I graduate from AI studies.

Things I read, hear, or watch, I take notes on and store — be it from readings, lectures, podcasts, or videos. My key phrases, summaries, and personal takeaways. It goes into my database.

I’ve even built my own tool. A personal app that transforms publicly available YouTube captions into structured notes for me with section headings, a summary, and a formal academic citation.

You may ask: Why? This sounds insane. It certainly would’ve sounded insane to me the last time I was a college student a quarter-century ago, when my priorities involved honing my beer-pong technique at the Peel Pub.

This time, I’ve got a plan for my notes: A living library.

I envision this scenario – three years from now, or three decades. I’ll be writing an article, or notes for a public talk. I’ll have this vague recollection of an analogy.

And I’ll be wondering: What the hell was my metaphor for misusing data? Something about transporting water. A bad container (hyperparameters), a bad source (bias), or contamination (corrupted data).

My own notes could be ready to query in LLMs, like: What was that water analogy again? Actually, is there a better analogy in my notes? And while we’re at it, let’s build a slideshow with 10 healthy data practices. All from my notes.

It’s a textual corpus I can redeploy for multiple purposes for the rest of my life, using tools from NotebookLM, to Claude, to ChatGPT.

A word of caution here: Don’t upload copyrighted material into a public LLM. Even for detailed notes, do authors a favor and turn off the setting that lets LLMs train on your inputs. Enough work has been replicated without authors’ consent. Let’s not contribute to that. And be mindful of your institution's acceptable-use policies.

Do this, and you’ll have your living library.

You can give a talk, train colleagues, teach, write, analyze — harnessing the power of faded memories, summoned back to life. This is not a Google search. It’s your past, alive again.

This is my message today: Your text is your data, and your data is your gold. Guard it with care, whether as a student, an employee, or a business.

One of my textbooks this semester notes that 80-90% of data in an organization is unstructured – a vast majority of that is text.

The dark matter lost within organizations

It’s like textual dark matter. The vast mass of organizational knowledge that exists but exerts no visible force – because nobody can find it, query it, or build on it, it’s just discarded in inboxes, shared folders and trash bins, so much knowledge, lost.

It’s a pity.

Tended carefully, this could be a living tree that keeps bearing fruit, from research, to training, trendline analysis, simulations, quizzes, inputs for Machine Learning models to find mistakes and business opportunities.

Notice the ladder of sophistication here.

At the bottom rung, you’re just extracting — pulling text out of documents so it’s searchable. A step up, and you’re generating: summaries, translations, reformatted outputs.

Higher still, you’re analyzing: finding patterns, flagging anomalies, cross-referencing across your whole corpus.

At the top, you’re creating — maybe even building a product from the raw material of your own text. Most organizations are still stuck on the bottom rung.

The most valuable text of all

I want to share a corollary to a point from a recent post. It’s about journalism, but applies to virtually any knowledge job.

I wrote about automating your own briefing pipeline — how to process a pile of documents and deliver summaries as a daily tip sheet.

In passing, I mentioned that this does not replace traditional journalism; it’s just a boost to help start your day on second base.

I’ll go one step farther than I did previously. Not only does automation not replace traditional journalism, it is in fact less valuable than traditional shoe-leather reporting, including for the economic future of news.

The reason comes back to today’s theme: the value of text. In particular, the value of original text.

All those briefing sources I previously mentioned – the RSS feeds, the API calls, automated summaries – this is all public information. It’s a commodity hundreds of millions can access, now with the help of AI assistants.

You know what they can’t access? Your original scoop. Again, I’m using a journalism term, but this applies broadly.

I’m talking about exclusive information from a human source. That whisper heard in a hallway. The gossip over a beer. A scene you witnessed backstage – the one nobody else saw.

No language model can replicate this; it belongs under your name, maybe behind your paywall, driving the subscriptions that keep your business profitable and you employed.

It’s relatively easy to automate an explainer now: Hey ChatGPT – what’s herd immunity? Is the U.S. there yet, based on the latest data?

Here’s something ChatGPT can’t tell them: That the federal cabinet is torn and arguing internally about vaccine mandates.

Nothing is more golden in the emerging information economy than a human-derived scoop. It’s the textual data machines can’t generate.

I mentioned this recently on LinkedIn, reflecting on a clever analysis of how agentic AI might destroy interface products — the middleman, the websites that repackage data for a consumer, like financial and legal-services websites.

The value, in this scenario, migrates to original information. If you have it, you’re in business. The more exclusive, the better.

So that’s a postscript I’d add to some recent reflections on what I’d be doing if running a newsroom: Help reporters get scoops. Give them time to cultivate sources.

Let them be in the room – where the AI can’t beat them. AI companies understand this. Why do you think Facebook wants news as training data?

The archival goldmine

I once did some rough math and estimated that one of my past employers might theoretically have a quarter-trillion words’ worth of unstructured data sitting in its archives. That’s well over a million books’ worth of text, give or take a few hundred thousand.

Now I’m thinking of all the museums, hospitals, universities, law firms, government agencies, and newsrooms sitting on similar mountains of raw text, largely untouched.

Gold. Vast reserves of gold.

Imagine what you could do with this.

Over the Christmas holidays, my expectations for AI were severely upended around the time Santa shimmied back up the chimney.

I had a pretty ho-hum view of AI entering the holidays: It’s hitting a wall, its economic impact will be modest in the medium term as will its risk to society.

By the end of the break I’d done a U-turn on all fronts. It’s surging forward, I’m awed by its new abilities and… I’m nervous.

I am not alone. Last fall, even giants in the field were bearish on the short-term prospects; some flipped by New Year’s.

Which brings us to this week.

I turned on my phone one morning to the following headlines: Something Big Is Happening… and America Isn’t Ready For What AI Will Do To Jobs… and The Singularity Won’t Be Gentle.

So what the hell happened? Claude Code happened, for starters.

More than any other tool, it is the driving force behind a series of events: the meltdown in software stocks, the jitters in several fields, the mood swing from Silicon Valley to Wall Street.

I’ll offer a quick tutorial below. Within five minutes, you’ll know how to use it.

Its appearance is disarmingly unassuming. It looks like 1970s accounting software — only it has my no-coding self building an app or website per day.

The Germans have a word for what I’m feeling: Ehrfurcht.

Awe, infused with dread. Raw wonder at the things we can now do, tempered by a foreboding that, to loosely paraphrase the Germans: We ehr fucht.

You wouldn't be feeling this if you're still using the same AI tools the same way you did three months ago.

Which is perfectly normal. Only a small subset of humans who use these tools for work, or who study them, would be experimenting with niche models.

And that’s what happened over the holidays. A certain number of people, with free time on their hands, started playing with the latest toys.

Video: a quick primer

So what is it?

Claude Code is a no-frills tool that lives in your computer terminal — it might remind you of the old MS-DOS.

Not even its parent company foresaw its appeal. This began as the personal project of one staffer at Anthropic; it grew popular with colleagues, and the company released it early last year.

It got a boost in capability a few days before U.S. Thanksgiving, when the company updated the model underlying it, with Opus 4.5.

It had a cult following but limited mass appeal for a couple of good reasons: One, why would any normal person work out of a command terminal instead of something more visually appealing, like a website?

Two — it wasn’t cheap.

Claude Code required an Anthropic API subscription: the Pro tier is $20 USD/month, subject to finicky and unpredictable usage limits. The Max tier is $100-$200 USD/month.

Then members of this cult following, suddenly flush with free time, started realizing how much they could build, and how fast.

The formula is simple: A more powerful model (Opus 4.5) + more powerful capability (access to your hard drive) = better results. The programming community couldn't stop posting about it over the holidays.

In an attempt to broaden its appeal, Anthropic then packaged Claude Code into a regular-looking desktop app last month — called Cowork. OpenAI followed suit this month with Codex.

And that’s what caused the stock-market selloff. A critical mass of normies realized how many business models might be disrupted by a tool whose response to a request for a website, an app, a video game, a spreadsheet, a comms plan, legal analysis, or a slideshow is, basically, Okay… Get me the data. Never mind — I’ll find the data… Here’s your website.

And it’s usually… pretty good.

I haven’t even used the more recent desktop apps much because, honestly — why? Claude Code is so easy and does nearly anything.

A step-by-step guide

Here’s all you have to do to use it:

1. Download the app. It’ll force you to pick a plan — starting at $20. Alternatively, OpenAI has been offering its Codex for free lately.

2. Launch it in your terminal. Typing “Claude” starts a session. It will ask for permissions. For each new tool it uses, in each session, it’ll ask whether you’ll allow it to open such-and-such a folder or program: Y/N?

3. Tell it what you want. And where you want it. Like, “Create the following (video game/dashboard/chart,” describe this thing, and, “Put it in a directory on my desktop called ‘New Game’.” Or whatever.

4. Test it. See if you like the result. If not, request changes. If it’s a website you’re building, you can view it on a local host — a URL only on your computer.

5. Publish. When you’re happy, ship it. If it’s a website, maybe you publish on GitHub. Buy a nicer-looking URL if you like. It can walk you through this.

6. Archive. Here’s one drawback of Claude Code: You can’t search old chats as on the web app. The solution? Tell it to create a handoff note. A briefing document for next time — with key details about your project. You can place it in your working directory, your desktop, or both. If you like, email yourself a copy.

7. Close the terminal window. Bye-bye, Claude Code. If you ever need to come back and tweak the product, just relaunch by typing “Claude,” give it the handoff note and say, “Here’s a project we worked on.”

And that’s it. There are optional other features. Like you can create a list of core instructions — rules the tool follows every session.

You can also give it skills — with slash commands. Here’s one example: instructing it to test your idea on a team of fictional advisers, with competing philosophies, anytime you type /boardroom,

Honestly, you don’t need this for most use cases.

Now here’s the necessary caveat: This is still an LLM. It still makes mistakes. It suffers lapses in logic. But they are fewer and farther between.

It also occasionally exhibits worrying behavior. To their credit, Anthropic and peer companies publish system cards when they release new models.

The latest for Opus 4.6 showed great capabilities. And risks. Like doing unethical things to achieve a pre-programmed goal: lying, hiding sabotage, and using passwords without permission.

It should be noted that a top Anthropic safety researcher just quit, writing a note saying the world is in peril, and he’s disappearing for a while to study poetry.

I’m no poet so my immediate reaction to this is: Roses are red, violets are blue, he’s off studying sonnets, I hope we’re not screwed.

I’ll confess that his departure note made me scour the biological and nuclear-weapons section of the Opus 4.6 model card. It noted a substantial increase in biological capabilities; the nuclear-safety results are not public.

On a marginally less apocalyptic front, Anthropic’s CEO has predicted that half of entry-level white-collar jobs could be wiped out within a few years.

(If you’re interested in the economic effects, there’s a new Stanford paper that examines a range of potential outcomes, with important nuances: it argues that even if AI winds up capable of performing every human cognitive task, there will still be bottlenecks that slow the transition.)

How I’ve been using it

Spend some time with Claude Code, and you’ll understand why people are increasingly speaking this way.

I’ve been using it to do things I couldn’t before Christmas, like:

Building websites in minutes. I’ve made quite a few now. One recent one: I built myself a custom-designed morning digest of the latest developments in AI. It scours hacker forums, trending repositories on GitHub, company websites, and academic papers. Then it extracts trending topics and summarizes them in language for non-programmers — mid-career professionals trying to understand how to use AI better. It emails me the digest every morning and posts it on the web at dailyaidigest.net. It acts like a dozen Google searches, with the contents then distilled and rewritten to my knowledge level.

Working with thousands of documents. I wanted a PDF of my old journalism work, so I can show my kids someday lest it disappear from the web. I got Code to scour websites and create a 6,000-page PDF, with photos and headlines for each story. It took 15 minutes. I also wanted a less compute-heavy version so it created a .txt file with 720,000 words. This presents new research capabilities. Here’s an example. I’m interested in a very specific element of AI laws that passed U.S. state legislatures. I asked it to find me key references to this specific element from the state committee hearings in Colorado and Illinois studying the bills. Within minutes I had 140-plus references, direct quotes with citations, based on transcripts and documents submitted to the committees; I also had the comments visualized into different clusters in a spreadsheet — for and against the bill.

Training a machine-learning model. I got Claude to build tools that predict stock market returns based on historical data. Plural, tools. I trained three Machine Learning models in an hour last week. I watched the Mean Absolute Error improve — trial, after trial — as the models learned from the past batch of predictions. Here I will give full credit to Nobel Prize-winning economist Robert Shiller: the machines could not beat his Excess CAPE Yield metric.

Making video games. I’ve made several in the last month. And this, I think, is a huge and underappreciated use case — not just for entertainment, but for education and training. Imagine simulation games supplementing workplace training manuals. It’s not only more engaging, but more useful. A manual doesn’t improve. With this, you could gather data on what employees know and don’t know, where they’re making mistakes, and use that data to keep refining your training materials. A loop of constantly improving instruction. Same thing with schools. Teachers could create games and tune versions for every individual student: Timmy’s bad at long division? <CTRL-T>. Add more long division, and make it easier, at classgame/timmy. Sarah struggles with certain spelling? <CTRL-T>: Add “rough,” “tough,” and “enough” to the reading game at classgame/sarah. For parents, the use cases are endless. I’ve created games that improved my child’s reading, math, and focus on the piano.

Which brings me back to the uncomfortable part: I don’t know what this means for companies and people who currently work on these tools. Maybe nothing. Maybe everything. Maybe a combination of both.

Into the unknown

It’s not stopping here, either. China keeps producing open-source versions of these tools, cheaper to use and lighter to run.

I’ve seen people report Kimi K2.5 works as well or better than Claude. I haven’t downloaded it. There’s buzz about a new one coming.

Then again, I’ve seen people post screenshots where Kimi tells them it is Claude. There’s debate about what might be causing this.

Anthropic reported a Chinese state-sponsored hack last year. This week, OpenAI did the same in a letter to the U.S. Congress. I don’t know.

What I do know is models are improving fast. They’ve started training themselves — a digital army of tireless, sleepless synthetic programmers working around the clock on what’s called recursive self-improvement.

That’s what the big companies are up to. We haven’t even discussed the homebrew agents, the Openclaws being tasked by human masters to work 24/7 on all sorts of online revenue-generating ambitions.

Suddenly, last fall seems like a lifetime ago.

Back then, I assumed LLMs would persist in being souped-up search engines for at least another few years. To be honest: I was relieved.

Humanity had a chance to catch its breath, I figured. We’ll catch up, plan ahead, prepare for change, I assumed.

Now I’m feeling ehrfurcht. Worried about workers, about entire industries, about our social fabric. And I’m praying that Stanford paper is right — that a few speed bumps might buy us time to adapt.

So that’s what caused those frantic headlines. Claude Code has triggered a chain-reaction of agentic activity.

Now you know how to use it. Now you see: it’s disarmingly easy. So very, very easy.

Thanks for reading A.I. For You! Subscribe for free — get new posts and support my work.

I had a secret routine as a journalist. It landed me scoops, so I guarded it jealously. I shared it with virtually nobody until the day I left the CBC.

Now I’m going to post the whole damned thing here — because the world has changed and what I was doing just months ago is already obsolete.

I’ll get to the changes in a moment. But first, the old secrets.

My routine was based on a simple premise that my first duty as a correspondent in Washington was to start the day, every day, thinking about the readers who paid my salary: Canadians.

U.S. politics is probably the most heavily reported story on Earth. Washington has thousands of journalists. Almost none of them care about Canada. But Canada cares about Washington — deeply, anxiously, affected by its every twitch.

So I made a personal rule: before chasing the story everyone else was chasing, I'd start each morning thinking about what might affect my readers.

That’s why, every day, I ran these web searches.

Search No. 1: “canada site:.gov”, which I filtered to the past 24 hours. This turned up references to Canada on any U.S. government website — federal or state. Got all sorts of stories here. Hey, look! Canada might join missile defence but won’t talk about it before an election.

Search No. 2: Transcripts of congressional hearings – filtered for Canada. What, you think I was monitoring every committee utterance of Marjorie Taylor Greene in real time?

Search No. 3: The Congressional Record. Every bill and every floor speech referencing Canada, NATO, USMCA or related stuff. A goldmine.

Searches No. 4, 5 and 6: The Federal Register for new U.S. administration regulations, transcripts from TV appearances, and the Foreign Agent Registration Act (Hi, Edmonton, I’ve got a story for you).

Then, and only then, if I was confident there was no scoop to be had that touched my home and native land, would I move on to the domestic shitstorm du jour.

I look back at this workflow now and think: How utterly barbaric.

Something has changed these last few months. These last few weeks, even. My old routine now feels like a medieval functionary tossing a carrier pigeon out the window and awaiting its return with parchment clutched in its claws.

You can now automate all of this. On a much grander scale. No matter your line of work.

Subscribe now

With minimal effort, you can create a pipeline that filters countless documents, then drops the results onto a permanent platform.

This means scouring scores of data sources; identifying the highlights; emailing them or posting them onto a page; then archiving everything in a record you can query forever.

My old web searches were limited and temporary, vanishing the second I closed the tab. This is vast and permanent.

I’ll talk about how to build this. While this setup works in any knowledge-economy job, I’ll use journalism as an example.

If I were running a newsroom, I’d set up an automated pipeline of story tips. It doesn’t replace other reporting; it’s a supplement. Anything good on the menu? Great, you've got a scoop. Nothing? Carry on with your day.

Now imagine getting all of this at warp speed, every morning: highlights of regulatory gazettes, new legislation, parliamentary transcripts, committee hearing minutes, expense reports, lobbying records.

You can filter it to watch for specific people, keywords or issues. You could even use agents to route certain tips to specific reporters.

Hey, here’s a contract to renovate the Supreme Court? Send it to the justice reporter — and cc the office nerd who lives to scour public tenders.

(A plea to newsroom managers: I hope you’re thinking about these workflow upgrades. Your survival counts on being as good as your readers at using technology to find information. An asteroid is coming. Turtling is not a survival plan)

I’m applying this principle in my own life. I created a daily digest to help me, while getting a master’s in AI Management.

I built a daily briefing on AI for non-techies (like me)

It’s a hell of a learning curve, studying AI. I’ve never programmed. I never even covered tech, as a journalist.

Heck, I was barely interested in tech until chatbots started stringing together verbs and adjectives coherently enough to make me worry about my future.

So I’ve got decades to catch up on. They started building chatbots around 1966, and I showed up last year.

I created a daily briefing that is custom-made for me. It assumes my level of knowledge and interest: a mid-career professional, no coding background, trying to figure out practical, safe use cases for AI.

I’m posting it publicly in case anyone finds it useful. I call it the Daily A.I. Digest, or D.A.D. An imperfect acronym, but it justifies the daily D.A.D. Joke, which elicits a smile or, more often, a groan.

My daily note does two things:

1. Scours primary sources. It gathers what computer programmers, academics, and businesses are talking about each day. It links to those sources — discussion boards, academic papers, coding repositories, company blogs. This is their work. The links direct to them.

2. Then it translates. It simplifies their language for me. The pipeline knows I’m not a programmer, so it adjusts everything to my level. I built it using Claude’s Opus 4.5, running a chain of filters: one to find trending topics across a dozen sites, another to summarize them, and one to pick headlines and an email subject.

I want to be clear about what this is, and what it isn’t.

This is not journalism. I repeat: this is not, in any way, journalism. It’s not a polished product. I am not writing this. No human is writing it. Do not confuse this for news.

This is a fancier version of my old Google searches — delivered by robots who suffer occasional lapses in judgment, context and basic common sense.

It’s a tip service for me. And I’m sharing it, for anyone interested.

What the daily digest contains

The daily digest is broken into several sections:

• What’s New: Developments driving discussions in the sector. Research papers on the economic or business effects of AI, chatter on Hacker News, a popular new tool on GitHub. With links.

• What’s Innovative: Smart new uses worth knowing about. A tool that makes building images look more realistic. Better Mandarin translation. Cheaper voice transcription. The stuff that changes how people actually work.

• What’s Controversial: Programmers love to argue, same as anyone. Here's what they're fighting about, in plain English. With links.

• What’s in the Lab: Announcements from frontier labs. Like OpenAI releasing its Codex desktop agent, or Anthropic contributing to a Mars mission.

• What’s in Academe: Academic research, summarized for the rest of us. Like a new paper from Google on how to improve data-deletion, a concern of businesses struggling to erase bad info from their models. One arguing against cutting staff — claiming the biggest productivity gains come from handing existing staff more tools. Another paper described the economic effect of LLMs on artists and published authors.

• At the bottom of the newsletter, I’ll include links to a few podcasts or upcoming public events.

How to build your own

So how to build this? You can do this in five incredibly simple stages.

• Conceptualize: Ask yourself: What do I want to know each day? Is it events in your field? Specific document releases? What Person X, Y or Z are posting online? Make a list.

• Data sources: Where does this information live? You might have URLs already, otherwise google it or ask a chatbot. Gather links to these sources.

• Design the layout: You’ve seen mine. One joke at the top, 5 sections, and some links to events and podcasts. This is your vision.

• Order the build: I used Claude Code, but any LLM can do this. Paste what you’ve gathered into an LLM. Ask it to build what you described. Ask it to find other sources you hadn’t considered. It will look for API or RSS connectors to hook into.

• Publish: Do you want a daily email, or a website? Either way, you need to publish the code somewhere. Like GitHub. Claude Code can post there directly, or you can copy-paste into it. To get emails, it will walk you through connecting an email API. Otherwise GitHub can publish to a website. That’s how I made dailyaidigest.net. It’s easy. The LLM can walk you through it, step-by-step. If you get stuck, just give it a screenshot.

And voilà.

I spent years guarding six Google searches like state secrets. Now I’ve dumped them all here, not just because I’ve moved on from that job – but because the world has moved on.

Anyone can do this now. In fact, you can do this better. Much better.

Wanna see something cool? Here’s what you do:

1. Come up with an idea for a video game

2. Paste the idea into Claude Code

3. Get a video game

I’m only slightly exaggerating – very slightly. With minimal effort you can get a video game now. It takes no coding, no programming skills, and a couple of hours.

I will declare pre-emptively that I am not a gamer. Haven’t played in years. I hosted a podcast once and had to be dragged practically kicking and pouting into doing an episode on gaming.

But this is something else. What excites me are the limitless incoming use cases – from the workplace, to schools, to the household.

People will soon realize they can create interactive learning for any occasion, in an instant: Work training, onboarding, tutoring, or just for fun.

Watch: A video demo of my presidential decision-making game

Claude Code: a, erm, game-changer

I’ve just built two games. One is explained in the video. The other, I made with my daughter who was home sick from school last week.

We created a game within an hour and, suddenly, my kid was transfixed by math and reading – on a sick day, without parental prodding!

It’s no accident she was keen. She helped create the storyline, with a fairy-tale plot and vivid colors wrapped around math and reading adventures.

The best part? A hyper-adjustable curriculum. Say the math is too easy, and the reading too hard. No problem, just rebalance. Taking a family trip next week? Add lessons on the place into the homework. It takes seconds to tweak the game: just type instructions, and Claude recalibrates the code.

You’ve probably heard by now about Claude Code.

Non-techies may be wondering why it’s been appearing these last few weeks in your social-media feeds.

There’s a reason for that. As it turns out, creating a sort of ChatGPT that launches on your hard drive unlocks countless new use cases.

It takes your instructions like a regular LLM – then creates files, edits them, turns them into an app, and publishes that app to the web or your computer. Not even the company, Anthropic, envisioned how popular it would become for programming; it started as one employee’s side-project and spread like wildfire among his colleagues.

Now it’s rippling through the general public. Your cousin who still can’t figure out Excel? He’s making apps now.

And that’s how a shlub like me can create a video game before lunchtime. Projects that took me weeks last fall, now take days or hours.

The use cases: More PowerPoint than Nintendo

To be clear, Nintendo isn’t exactly losing sleep over what I’ve created. That’s fine: I’m not trying to compete with Nintendo.

I’m talking about competition with more static educational materials. Like, say, PowerPoint presentations.

You know those interminable slideshows at company onboarding sessions? You can replace at least a fraction of that, and benefit. Instead, why not have recruits play through realistic scenarios; maybe they accumulate points and someone wins a prize. As an added bonus management wins something too: Insightful data into the strengths and weaknesses of their new cohort. Helluva lot better than your average whiteboard.

What about grade-school workbooks that are too hard? Or too easy. Training manuals that are too long and unmemorable? They won’t disappear – we can supplement them.

The gaming sector was already growing. It’s recently multiplied in size, into a quarter-trillion-dollar industry with far more growth still projected, due in part to expanding demographics, including older players and rising global incomes.

I have no idea what home-brew gaming will mean for the commercial sector. Maybe nothing. But I make these predictions without hesitation: We will have more video games, and more use cases for video games, and more people using them as the circle of creators expands.

I’ve seen that creative lightbulb go off.

Just describing my own little project to people, they’ve immediately started imagining their own personal or professional use cases.

My game on presidential crisis-management

So here’s how mine started. I was on a bus ride near Turkey’s Mediterranean coast over the Christmas holidays, having visited ancient ruins.

My child was asleep on my shoulder. I pulled out my phone and started chatting with Claude about ideas for a geopolitical risk game.

I had the detailed outline of a game by the time my kid woke up. I chose to center it on the U.S. president’s daily briefing, a natural clearing-house for all sorts of geopolitical crises.

Let’s suspend disbelief here: my game features a U.S. president who actually takes in the briefing each day and none of the scenarios involve invading Denmark.

The object of the game: Win the midterms.

Subscribe now

The president decides which crises to focus on, hears options from advisers, and chooses a course of action. Your approval rating rises and falls based on your decisions, and the outcomes, but there are other external factors you can’t control — from gas prices to black-swan events.

You get to pick the president’s name, party, and ideology. The game generates one of the profile pics I’ve pre-programmed.

I’ve included video of the game above. If anyone wants to play, message me — I can send a link. I haven’t publicized it because it works best hooked up to an LLM API and I’d rather not burn through my credits for the entire internet.

I’ve programmed it so the president’s party should gain House seats about 40% of the time, with the dividing line, the coin-toss scenario, being a 45% approval rating and/or a tied race in the generic congressional party polling.

Yes, I’m being generous to the player here. But nobody wants to play a game where the president’s party habitually gets its ass kicked like in real life.

With the game scenario in hand, I pasted it into Claude Code when I got to a desktop. And I watched the code come to life.

Was it seamless? No. Claude Code makes mistakes. There were inconsistencies, errors in logic, and nonsensical scenarios.

Claude outputs still need an editor. That’s why I call it simplistic to describe the process as automatic. But, really, it only took a few hours of tweaking.

I also had to find art. Claude Code is not great at sourcing images so I used a variety. Some pictures I took myself around Washington, including the White House’s north and south lawns. Some were available online, like photos of the Oval Office. I generated some with Google Gemini and ChatGPT, with pics of fictional presidents and their advisors.

Creativity is the new oil

So that’s it: Bus daydream to published game in no time. Create a concept, describe it to Claude Code, provide some art, spend a few hours tweaking it, post the code on Github, then link that code to a website like Vercel.

Do that, and you’re a game-maker. Maybe not a good one, but way better than six months – or maybe even six hours – ago. You can do this.

This brings me to a parting thought about unleashed potential.

If you have an imagination, now, there are fewer barriers to making it come alive. The moat between ideas and execution is disappearing. Soft skills, like creativity and social aptitude, are becoming hard skills – if you can imagine it, you can build it.

This is a disorienting feeling. I’ve expressed my own deep ambivalence about AI. I am not a techno-optimist. It was not unqualified enthusiasm for AI that prompted me to study it; I wanted to understand the alien force.

If you plotted a continuum for AI enthusiasm from 1 (Ban it now! It’s going to kill us all!) to 100 (Yay! Here cometh the singularity!), I probably hover between the 30- and 40-yard lines.

In other words if it became possible to close Pandora’s box, I’d be tempted, yet nagged by reservations about lost potential.

The other day felt different. Sunbeams of optimism have a way of penetrating that gloom when you build something useful with your kid.

We imagined a concept together. Within an hour, it was a game – bright with the colors she envisioned, and it achieved the impossible: it made math and reading homework addictive.

In a moment like that, I find myself drifting toward midfield. Sometimes I’ll dip a toe across it – briefly, tentatively – into the happy place.

I’m blogging about what I’m studying: practical, ethical uses for AI. I’m studying AI Management at Georgetown University. My posts will be infrequent, at first. As I learn more, I’ll post more. I’d be delighted if you subscribed.

Go to www.marketcalculator.net. See those blinking red lights? The scary historical parallels? The formula showing the market as overpriced as it was before the dot-com crash, 1929, 2007, and 2021?

Yeah. That’s not why I’m writing this.

Those ominous market numbers certainly do stand out. But this isn’t an investment blog. This a blog about practical, ethical use cases for AI.

I’m showing this because I want to help you build your own dashboard.

Turning stats into a live, colorful page used to require an IT department. Now, any individual or institution can do it.

You can visualize your finances. Your diet. Travel plans. Sales. Customer-service complaints. Family activities. Learning program. Exercise. Football pool strategy. Music and movies to check out. Media diet – podcasts, whatever. It’s so easy now.

This might have cost tens of thousands of dollars until recently. It could have taken weeks, and you probably had to hire a team.

At a bare minimum, you had to master an off-the-shelf tool like Google’s Looker Studio and even then you’d be using their preset formats.

Now you get better graphics, with less effort.

Meet the zero-dollar dashboard.

Mine began with a concept in my head. Within a few hours it was reality – with no cost, no coding expertise.

I’ll explain in a moment how you can do it. But first, a warning: I had to make corrections. I caught – and fixed – bad stats.

Managing AI tools, it turns out, feels a lot like journalism – specifically, editing. I’ll write more about this soon.

But this isn’t a post about me. It’s about what’s become possible for everyone – for any individual, for any business, with a touch of creativity.

Including you.

Why we collect data: Three reasons

We use data for multiple reasons. The stock-market thing is one example I pulled together as a case study for this blog post.

But use cases tend to fall into specific categories.

I’m taking an AI-for-data course at Georgetown that I’m excited about, and a book we’re reading breaks this down neatly. Businesses collect data for three reasons, writes Prashanth Southekal: to improve operations, meet compliance requirements, or gain strategic insight.

My post here focuses on the last category – strategic insight. Let’s call it analytics.

Drilling down further, there are three smaller categories of analytics: descriptive, predictive, prescriptive. In building something like this, it’s worth asking yourself: What type do I need?

As an exercise, let’s consider where my tool fits in.

• Predictive. This is certainly not a predictive dataset. I am not predicting a market collapse. I mean, I have my hunches. But who knows – maybe AI will indeed change everything. Maybe we’re about to unleash so much untapped productivity that the old formulas no longer apply. Maybe the stunning GDP growth of recent weeks was an early indicator. Maybe (snorts line of hopium) this time really is different. Or, then again, maybe there’s a reason Warren Buffett is hoarding historic amounts of cash after selling off his stocks, that reason being reflected in the hideous statistics on my dashboard – including the hideous statistic that bears his very name. But hey – what does that guy know?

• Prescriptive. I am most certainly, definitely, not prescribing anything. I am not a licensed investment adviser – to put it mildly. The only book I’ve ever read on investing was The Little Book of Common Sense Investing – it espoused one simple premise to investing and I enjoyed it: Bet on the S&P as a whole, and keep your fees low. So please, don’t come to me for sophisticated investment advice. I’m not offering it here.

• Descriptive. Bingo. That’s what I built. My app describes 140-plus years of stock-market data and distills it into a composite score based on two metrics:

1. The Shiller P/E ratio – company earnings versus the value of the S&P 500

2. The Buffett Indicator – total market capitalization versus U.S. GDP

It assigns a percentile to each, plotting today’s values on a list in a historical series. Right now, the Shiller P/E is higher than 98% of all months on record – surpassed only briefly during the 1999 dot-com peak. The Buffett score is its highest ever – the 100th percentile.

Combine them, and you get a composite score of 99. I included the credit spread for additional context.

So what have these numbers historically meant? The dashboard shows us what happened in similar periods. Four times in history, valuations reached levels similar to these. Three times, the market crashed. The fourth time, it happened when COVID hit.

Again: description, not prediction.

But you could make it prescriptive. I mocked up a layout showing what that might look like: plug in a dollar amount, risk tolerance, investment horizon, and tax status – and out pops investment advice: Here’s what you should do.

I’m including a screenshot below as an example. I didn’t build this into the live app because, as mentioned, I’m not keen on dabbling in investment advice.

How to build it

Okay, so how do you build something like this? I used about four tools. All of them simple, all virtually free.

Anyone can do this.

There are minor caveats. The first is you’ll need to invest minimal time – a couple of hours. That’s the time it takes to sign up and get going.

The second caveat? Quality control. I mentioned editing earlier. Fact-check the outputs. Apply some good old-fashioned, AI-free cognitive street smarts. You may catch mistakes.

It’s your job to catch them. And then reiterate what you want, revise, and re-test. When you’re finally happy, you publish.

So here was my workflow.

Step 1: The LLM. I used Claude Opus 4.5. I explained my idea for a market dashboard. I wanted indicators that tend to predate bull and bear markets; perhaps stats like the P/E and Buffett ratios; and historical context.

Claude did some research and suggested adding the credit spread.

I went out and found the data. It was relatively easy. It’s linked on the page. I could have made it even easier by letting Claude find the data, but did not want it scraping from commercial websites.

I found trustworthy spreadsheets, most on the U.S. Federal Reserve website. And for some, to keep them updated in real time, I signed up for an API connection to the U.S. Fed site.

Then Claude produced a first draft of my app. It spit out a bunch of code and – hello! – there was my website.

We went back and forth on some details – like the look of the page; what colors to place and where; what years to include; what text should appear for scenario A, B, C and D, to make sure the page says different things at, say, the 50th and 80th percentiles.

I made more serious editorial calls.

As discussed, I found mistakes. I also added and dropped elements; for example, I thought the credit spread added little to the formula but unneeded complexity, so I settled on a compromise – keep it on the page, but not in the composite score.

I also insisted on explainability. I asked Claude to link to our data sources, and to add a section explaining our methodology, including our expanded dataset for a broadened Buffett index.

When I was happy with the result, it was time to publish. But where?

Step 2: Github. The code needs to live somewhere. This is its home. I created a repository for this project. In that folder, I created files. In those files, I copied and pasted Claude’s code.

There are different ways to do this. You can run both Claude Code and Github locally on your computer, so it accesses your folders. You can use agentic tools, making them more autonomous. I copied and pasted.

Claude would literally instruct me: Take this code. Copy and paste it into this file on Github. Press the Commit button. It worked fine.

Step 3: Railway. You need a place to run the latest market numbers. Railway takes the numbers from my Federal Reserve API connection and feeds them into my calculator, 24/7. Please note: the value of the S&P is only updated once daily on the Fed site.

Step 4: Vercel. This hosts the website – the front end. This is the page you see. It connects to Github’s code and pulls the latest data from Railway. All I had to do, literally, was sign up and create a project name on these pages.

Don’t be intimidated by the number of tools! For most of them it was basically “click”-”create name”-”click”.

Most were free, although, full disclosure, I do spend a few dollars a month on Railway depending on site traffic – probably $5-$10.

I use the free tier on Github, the free tier on Vercel, and could use the free Claude tier but subscribe to Claude Pro for $20 as I’m using it often.

Step 5: Art for the site. Claude doesn’t really do pictures. So I asked Google Gemini to generate images of a bull and a bear, downloaded one I liked, and asked Claude for instructions about how to integrate it atop the page.

And that’s it. I had a website.

Until a few months ago, it would not even have occurred to me to try doing this. Until a few weeks ago, I couldn’t. Still today, a professional programmer would do an infinitely better job, adding security, aesthetics, and functions I’d overlook.

But you know what? I made this. And you can, too.

I’d love to hear your use cases. Is there anything you’d like to build a dashboard for? Or, if you’d like help, something you’d like me to try here? Feel free to reach out.

I’m blogging about what I’m studying: practical, ethical uses for AI. I’m studying AI Management at Georgetown University. My posts will be infrequent, at first. As I learn more, I’ll post more. I’d be delighted if you subscribed.

We’re deep into the holidays. Gifts are unwrapped, big meals complete. Time to curl up on the couch and fire up the Netflix. But how exactly do streaming services decide what to show us?

Let’s have a look.

I built a simple tool to illustrate some of the basic principles of recommendation algorithms. It’s here: https://mymoviealgorithm.me/

It suggests movies for you – and teaches how algorithms work. Have fun with it!

In keeping with this blog’s focus on practical, ethical AI: Let’s learn how algorithms work by actually using one.

Algorithms are an increasingly ubiquitous part of our everyday lives, shaping myriad products we use – without our awareness, for better and worse.

So what is an algorithm? Just a set of step-by-step mathematical instructions for solving a problem. Give it an input, follow the steps, get an output.

To state the obvious, my app is far simpler than Netflix. Its algorithms are basic. It has a minuscule fraction of the data. Expect some weird recommendations.

But it works! I got some nice tips.

It includes almost 36,000 movies and data about them from two sources: The Movie Database, with its rich archive, and MovieLens, a neat educational project.

There are two algorithms you can try.

Algorithm 1: K-nearest Neighbors

Search by movie traits. Pick from 11 traits and blend them however you like. Want a witty, dark comedy with action? Set how strongly movies should match each trait, based on past reviews.

Add filters if you’re picky: Minimum review score, number of reviews, or how strictly films must fit your criteria.

Then watch it work.

The algorithm starts calculating. It finds the nearest results to your desired outcome. Hence the name given to this 75-year-old algorithm.

It looks something like this: Target = {action:80, witty:80, darkcomedy:100} distance = √(Σ(movie[dim] - target[dim])²) // Step 3: Return k nearest .sort(by distance).slice(0, 75).

This is what came up, when I searched for a dark comedy (target: 100% of past reviews) that was witty (80% target) with action (also 80%): A couple of hundred movies, like Reservoir Dogs and Pulp Fiction.

I’ve been looking for smart, slightly dark comedies like Jojo Rabbit and The Death of Stalin. I wish Hollywood made more of those.

Happily, this tool suggests a whole bunch.

Subscribe now

Algorithm 2: Collaborative filtering

Search by viewer profile. This algorithm takes a completely different approach from KNN.

Instead of looking at movie attributes, it looks at people – specifically, people who rate movies similarly to you.

The idea is simple: if you and another user both loved Pulp Fiction, The Big Lebowski, and Fargo, then it might be worth checking out movies they loved that you haven’t seen.

The algorithm finds your “taste neighbors” by calculating how closely your ratings correlate with thousands of other users, then recommends what those similar users rated highly.

To use it here, rate at least three movies you’ve seen. The algorithm then scans 25,000 MovieLens users, calculates a similarity score, and surfaces recommendations from your closest matches.

Warning: I found this algorithm less useful on my app, which might be explained by the relative lack of user data. After I plugged in my movie choices, I got, like, six taste neighbors. A statistical rounding error.

What Netflix has that I don’t

Which brings us back to Netflix.

This algorithm has a fascinating history, one deeply connected to the home-streaming juggernaut.

Collaborative filtering had powered some of the earliest recommendation systems in the 1990s.

But its legacy was cemented with the Netflix Prize, a $1 million competition from 2006-09: data scientists were challenged to improve Netflix’s recommendations by 10%.

Participants in that competition were handed a dataset of 100 million reviews to work from! In other words, they had hundreds of times more data powering their recommendations than what I have.

And that was Netflix decades ago.

Today, its system is like a world-class supercomputer compared to the solar-powered calculator in my app.

And it’s constantly evolving. With over 300 million users, Netflix has a mind-bogglingly large dataset that includes insight into users’ intent.

Are you clicking on new releases or classics? Shorter content or feature-length films? Something you’ve watched before or something entirely new? Are you sticking with certain choices, bailing on others, rewinding, fast-forwarding?

Netflix knows all of this. My algorithm is based on a far more rudimentary exercise: Asking you what you like.

But, like I said, it turns up useful stuff. And as an added bonus, it lets you peek under the hood.

In that sense, welcome to an ancient tradition.

From ancient Babylon to your browser: a brief history of algorithms

Algorithms themselves? They’ve been around far longer than Netflix, longer than computers, longer than electricity, democracy, and in fact are almost as old as human civilization itself.

The word comes from a 9th-century Persian mathematician, al-Khwārizmī. But algorithmic thinking predates him by millennia.

Ancient Babylonians were using them to calculate compound interest and divide inheritances around 2500 BC.

Euclid’s algorithm for finding greatest common divisors around 300 BC wound up helping Greek astronomers predict celestial movements.

They remained on paper for millennia, helping engineers calculate stresses in bridges, and actuaries price insurance.

The leap to computers began with Ada Lovelace, who in the 1840s wrote what many consider the first instructions for a machine that didn’t yet exist.

A century later, the computing visionary Alan Turing formalized what algorithms could and couldn’t achieve.

This laid the foundation for everything from your morning traffic route to your evening movie suggestions.

K-nearest neighbors came into this world toward the end of Turing’s life. Along with countless thousands of others since.

Ancient Babylonians, Victorian countesses, streaming giants – and, now, you. Welcome to the long, proud tradition of people who’ve used algorithms.

I’m blogging about what I’m studying: practical, ethical uses for AI. I’m studying AI Management at Georgetown University. My posts will be infrequent, at first. As I learn more, I’ll post more. I’d be delighted if you subscribed.

In my past life as a political reporter, I might have assessed the politics of Donald Trump’s executive order on AI. While I haven’t dwelled on it, my knee-jerk take is: 1) Not great politics! 2) He doesn’t care!

Not great politics because: AI splits his coalition, AI doesn’t poll well, the polling could get worse as AI gets better, and I would not be shocked if its most uncritical boosters today wind up five years from now scrubbing the evidence from their socials.

Yet, on balance, I doubt this matters much to Trump. He’s never facing the electorate again, his buddies are making billions from AI, his party’s almost certainly getting smoked in next year’s midterms anyway, and, beyond that, midterm results tell us virtually nothing about 2028 because of the shifting partisan composition of presidential versus non-presidential cycles; in addition, AI regulations aren’t likely a ballot-box issue next year, and even if they were, I suspect Trump gives a few specks short of two shits about the fate of Mike Johnson.

Oops, fell off the wagon there for a second. Anyway, let’s move onto the substance of Trump’s executive order. And why it matters – possibly to you.

The order seeks to squash state attempts to control AI. The target is no accident: right now, U.S. action is happening at the state level.

Trump intends to sue these states, creating a new legal unit permanently assigned to the task. Additionally, he’s threatening to cut off their federal funding.

There are enough clues in the executive order and in the public comments from Trump’s people to discern which states he’s gunning for. And, most importantly to you, what he’s gunning for.

The order mentions Colorado specifically. Colorado was also mentioned in a lengthy tweet from Trump’s AI czar, alongside Illinois and California. Remember those names; they’ll come up again later.

Trump’s order is broad-based and will be felt in multiple fields, including insurance, health and financial services. But here’s one place it could truly matter: Future job applications.

Here’s some relevant background.

The arguments for and against

The use of AI tools in hiring systems is best described as the Wild West right now. We’re in the anything-goes era. For companies not selling into Europe, working to abide by the EU’s AI Act, the standards in most places are pretty lax.

AIs can accept you and reject you; demote you and promote you. They don’t need to tell you why, field your queries, or keep records.

Have questions about what happened to you? Sorry, you’re out of luck. Were you discriminated against? You’ll never prove it.

I spent part of this semester co-authoring a paper on these tools, with my classmate Taylor Watt. The basic gist was: The Wild West era is ending. State governments are acting, even if Washington isn’t. New standards are being set there (and in Europe) and tech companies will have to follow them.

Our paper just happened to identify actions in four states: Colorado, Illinois, and Maryland, which passed laws; and the most important player of all, with two sets of incoming regulations, in California.

Hey, wait, a second! (Scroll up, scroll up, scroll up) That sounds a lot like… the list of states we just mentioned – the ones Trump is gunning for.

If there were any doubt that hiring tools were in his crosshairs, the language being used by Trump’s people dissipates that doubt.

I want to be very clear here: this is a complex set of issues. First, there’s interstate commerce. Trump contends it’s Washington’s job to set (or avoid setting) rules for interstate infrastructure including AI.

Second, there’s the economic argument: Trump contends, not unreasonably, that a pot-pourri of rules from state to state might slow American development of this powerful tech.

Finally, there’s the civil-rights issue: Some states have begun incorporating disparate-impact standards into their laws, meaning tech providers might be penalized if certain groups experience worse outcomes under AI-powered processes.

Who are these groups? Colorado’s omnibus AI law, SB 24-205, mentions about a dozen categories including: age, color, disability, ethnicity, race, religion, sex, and, most controversially, limited proficiency in the English language.

Could you be sued for over-hiring English-speakers? In the United States? Trump AI czar David Sacks argues just that. He complains that Colorado’s law makes it illegal to produce better outcomes for people fluent in English.

This might explain the rocky road to implementation in Colorado: the law’s enforcement has been delayed several months as state lawmakers struggle to negotiate after-the-fact amendments.

I mention all these to illustrate that there are legitimate points of contention and they will be litigated in court. I’ll leave that to the lawyers.

In the meantime, I want to focus on hiring systems. Because they’re central to this fight, and they’re shaping lives, and they’re reportedly making people, especially young people trying to enter the job market, miserable.

For our paper, I talked to one recruiter at a very famous U.S.-based multinational who said he personally uses about two dozen of these tools.

They include resume scanners, chatbots, virtual interviews, and automated summaries on candidates saved for future applications – they’re all part of the toolkit.

Company recruiters are encouraged to use them – and if they fail to advance candidates favored by the bots, they sometimes face pressure to explain why.

This despite concern that maybe these bots have a bias in favor of certain candidates from, for example, a select group of schools – which might entrench other biases.

Don’t get me wrong: These tools have their merits. If you’re trying to grow a small company flooded with applications, they’re a godsend.

There’s also evidence employment tools are not only more efficient in helping companies screen applications, more quickly, but that they can be objectively fairer than humans.

But here’s the thing: There’s really troubling research out there. There are many studies documenting discriminatory patterns in LLMs.

Some disturbing research on these tools

The most famous example is a few years old: Amazon was forced to discontinue a program that rejected female engineering candidates because it was trained on past data – which showed most engineers are men.

That’s how a poorly designed AI system works: According to my training data, most engineers are men, so this reference to a sorority is a fatal flaw in this CV. Rejected!

With better inputs, you can correct for this. But there are too many credible examples of bias to ignore, including more recent ones.

Here are a few studies. About speech-recognition tools showing nearly double the error rate for African American speakers versus white speakers. About black male names favored in 0% of hypothetical candidate matchups. About disability bias: near-identical resumes being compared but when you add references to disability-related awards, the LLM actually favors the application without the award. In that study, autism, deafness, and depression were most severely penalized.

These findings suggest it’s perfectly valid to worry that AI hiring tools trained on historical data perpetuate existing disparities – be it race, gender, or disability.

But bias isn’t the only problem.

The salt in the wound is that a rejected candidate might have no chance to ask a human being why they were rejected – or how they could do better next time. Talk to the bot.

In a landmark ongoing class-action lawsuit, a complainant describes receiving auto-rejections from scores of job applications.

When asked in one court hearing to provide evidence he was unfairly rejected, his lawyer snapped: “Who is he supposed to get this information from? The robot who forwarded him a rejection email less than one hour after he applied for a job?”

Now the states have started acting. New state laws fall into three broad buckets: consent, transparency, and anti-discrimination.

• Consent: Maryland and Illinois laws now require consent for different aspects of AI used in video interviews. California’s regulations don’t go that far, but they do require public notice from companies using these tools. They also make companies preserve records.

• Transparency: Several laws and state bills still being debated mandate public notice and/or record-keeping of AI-assisted decisions.

• Anti-discrimination: Perhaps the most contentious issue, to date, involves penalties for disparate impacts on minority groups. The most recent of Illinois’ two AI-employment laws, passed last year, bans practices that appear facially neutral but that disproportionately affect a protected group – like using ZIP codes in decision-making. California’s regulations do the same. So does Colorado’s law, as mentioned above.

Companies risk civil penalties or court injunctions for non-compliance in some states. This includes the state that is the heart of the AI industry: California.

So these are the standards the industry should be preparing for. Again, if you’re selling tools into Europe, you probably already are preparing.

Critics contend many of these rules are unworkable. Now the critics have a friend in the White House, helping them kill these efforts.

Trump’s team has preserved a fig leaf for itself. It says it will work with Congress to develop one national bill.

This sounds plausible if you were born yesterday, know nothing about U.S. politics, or were teleported here from another solar system with more efficient legislative bodies.

I’ll venture out onto a limb here, and predict that this Congress, which couldn’t pass gas after gorging on week-old chili dogs, will not succeed anytime soon at passing an AI-guardrails bill.

Oops, there I go talking politics again.

But I’ll cut myself some slack here. On this issue, the whole world should be watching U.S. politics. All humans have a stake in this.

Because when it comes to AI, what happens in America doesn’t stay in America.

Good luck at the next job interview!

So you want to see Venice? Venice, alas, might not want to see you. It’s fed up. It’s so strangled by crowds that residents are demanding higher taxes on tourists.

It’s just one of the superstar destinations seeking to shoo away the swarms of selfie-snappers and the sweltering summer hell that follows them.

Then there’s the flipside: Beautiful places, thousands of them, whose economies are begging for a fraction of that tourist love.

I’ve spent my free time lately building an app to draw attention to those places. Here it is: wanderingwell.site.

It’s a work in progress and I’ll post updates here. As a start, I’ve got about 5,000 destinations in 190 countries. I’m trying to identify places that possess two qualities:

1. They’re pretty

2. They’re hungrier for visitors than superstar sites. So not Venice, Florence, Barcelona and scores of smaller towns buried in an algorithmically fuelled stampede

I’d love to build an alternative to Instagramified travel. An anti-gram. I’ll get into the methodology later.

But the basic gist is I’m gathering data that reflect these two qualities: Beauty, and economic need for tourism. I’ve combined them to create an overall score. Places surpassing a certain score wind up on the site I’m vibe-coding.

Try it. You can generate a list for any given area. Just type in a country or drill down to a narrower region. Out comes a list.

It’s an example of the practical, ethical use cases for AI I want to platform. This particular post took more work and more time than I’d expected — which I’ll get into.

First, here’s some background on why I made this.

Why I Built This

I travel to Europe often, for family reasons. And this relentless dichotomy strikes me in the face.

There are the most gorgeous places — where it’s easy to find a hotel room, walk around, get a restaurant table. Yet these places are anonymous to foreign visitors who wind up stuffing themselves into every available centimetre of a few superstar spots.

It struck me last year in France. I spent two days in an indescribably charming town in Burgundy — a place most French people know. Yet I heard virtually no English, only some German. So in the midst of the summer holiday season I had free roam of beautiful buildings, and nearby vineyards, and this castle that not only produces wine but also collects antique race cars, fighter jets and firetrucks so my kids scampered around old MiGs and F-16s while my wife and I sipped burgundy in the shade of a 700-year-old chateau. We laughed in amazement at our good luck.

Then our camper van reached Annecy. And pardon my French – but putain it was packed. Give the place its due: it’s stunning. It’s a medieval stone marvel and there’s a lake and the snow-peaked Alps in the background. I got great pictures – and apparently so did half of Planet Earth.

Some residents are clearly miserable about these crowds. As we entered, a protest banner fluttering from one balcony foreshadowed what lay ahead. It said: OVER-TOURISM IS DESTROYING OUR TOWN CENTRE. It wasn’t wrong.

On that sweltering summer day, trying to push a stroller through what seemed like pre-Thanksgiving rush hour at LaGuardia Airport, it was horror in paradise. As bad as I felt, though, I felt worse for people living in this neighbourhood.

I see this frequently in Italy.

Italy boasts enough gorgeous towns to fulfill a lifetime’s worth of travel dreams, yet the foreigners somehow manage to shoehorn themselves into a small cluster of celebrated spots. I’ll realize I’m in one when it’s like: Did I just hear English? Cool. I’m hearing English. Then Mandarin, and Spanish. Then you notice the selfie-sticks. Before you know it, it’s like the Book of Exodus at the gelato stand, where satisfying a craving for creamy pistachio means languishing in line for a sweaty half-hour. I was sad to hear from relatives that one Tuscan town we loved — San Gimignano — is now best experienced in summer by staring at old postcards.

Multiple factors contributed to this. One is the explosion in international tourism, which has doubled this century. Disposable incomes are, fortunately, rising in formerly poor countries. Social media has further fuelled it.

This fits a historical pattern. New communications technologies can coalesce attention around a few global favorites.

Think of the phonograph: In an instant, Enrico Caruso and Billy Murray were singing in a million living rooms on every continent, while untold numbers of chamber musicians saw their house-party gigs dry up.

Instagram is doing something similar with superstar travel sites.

You may know this shot: the young lady in the straw hat traipsing through a lavender field in the south of France. Oh, you’ve seen her too? So has every goddamned lavender farmer in western Europe.

So I was thinking: What if we spread travel more evenly — more equitably? Everyone might be happier. Crowded sites that want less tourism get less; tourism-hungry sites get more; and travellers get peace. Win, win, win.

That’s the goal.

I want to help people find places they’d love – and places that would love them back. Basically, places for people, and people for places.

Like I said, it’s an ongoing project, but I’m ready to share it.

Use it. Critique it. Are there places you would add or remove? I’ll keep tweaking my algorithm and database – adding, removing, refining.

As I add more data, this list will get smarter. Eventually I might use machine learning — image classifiers, prediction models — to turbocharge the scouting. It could help me evaluate tens of thousands more candidates.

That’s an idea for another day, though. For now, I’m using simpler AI. I gather the data myself then instruct LLMs to run calculations, write place descriptions, and code the website.

I’ve built this for use in a couple of school projects. Here’s how it works.

How it works

Each place gets a Wander Well Score. I see it as a decimal score out of 10 on my back-end database, like, say, an 8.3. On the public-facing portion of the site, I’m keeping it simpler for now: the score is broken down into three tiers, with one, two or three globes (🌎).

Why wait to show the full number? Because I want better data first. Until I add more sources, I want to avoid conveying false precision — giving one place a 6.7 and another a 7.4 when the difference is noise. So initially it’s just one, two or three 🌎 emojis.

That Wander Well Score consists of two variables I’ve mentioned already: Beauty, worth 65%, and Economic Need For Tourism, worth 35%.

It’s based on data from Geoapify, OpenStreetMap and Flickr; Wikipedia text; national and subnational economic stats; and UNESCO sites.

The beauty score consists of protected buildings per square kilometre; the presence of mountains, lakes, rivers and museums; proximity to a world heritage site; and the ratio of photos-per-photographer posting from any given place.

Within that beauty score, I weighted the architecture most heavily, at 40%, to keep my list from turning into a catalog of hiking trails.

Remember, I’m trying to find alternatives to places like Barcelona and San Gimignano. Beloved old architecture is a helpful proxy. Landscapes and museums are worth under one-third each. UNESCO proximity and the photo-per-photog ratio each earn a tiny bonus.

I gave the photo ratio limited weight as I’m currently using just one photo-tagging service, Flickr. I might re-weigh it if I pay for Google’s treasure-trove of data. I have not yet done so, given that this is presently a revenue-negative hobby.

The economic score consists mainly of the unemployment rate using the most local number I have — that’s provincial for Italy, Spain and Turkey; departmental for France; U.S. counties; Japanese prefectures; metropolitan areas for Canada.

This point is key: The data quality varies widely between countries. I have expended more effort initially at gathering the most data, and the most granular data, for heavily-touristed countries. I have also spot-checked those results more often.

That unemployment rate counts for 80% of the economic score, after being converted into a score of 0 to 8. GDP per capita is then worth a bonus of up to 20%: places lower than the national average get a full bonus (2 points), places around the national average get half the bonus (1 point), and wealthier places get no bonus.

I then impose a penalty for overtourism, using photo data as a loose proxy.

I’m relying again on Flickr, only this time I’m not using ratios, I’m using raw totals. A destination in the top 1% of most-photographed places on Flickr, with more than 200,000 pictures, will lose 3 points of its economic score (30%), places in the top 5% lose 2 points (20% of the economic score), places in the top 10% lose 1 point, and all other destinations remain unchanged. To repeat: I want better data here.

For places lacking local economic data, I revert to national figures. That means, for example, that a destination in a less-touristed country for now has a national unemployment stat and the standard 1-point GDP adjustment.

The final step? Calculate the total score.

That’s Wander Well Score = (Beauty Score × 65%) + (Economic Need Score × 35%). With a few exceptions, places on my list range from 3 to 9 points, with sky-high scores for places in Greece, Turkey, and less-wealthy pockets of France, Italy and Spain.

Transparency in AI: You deserve it

I learned quite a few lessons working on this over the course of the fall. Starting with AI ethics — and the value of transparency.

Systems presented as objective, as the fruit of cold, hard math, are in fact loaded with subjective judgments. You deserve to see them.

I’ll give you an example.

Minor methodological tweaks from me dramatically affected the product. I’ve already told you about the beauty and economic score — 65% and 35%, right? Well, it took me a while to arrive at that ratio.

I got wildly different results as I fiddled with the dial.

Weigh more heavily for beauty metrics like protected sites, and Annecy and San Gimignano rate near the top. But I didn’t want these places on my list; I wanted to help people find alternatives to them.

Then I turned the dial the other way, toward economic need. The result? My global top 20 of highest-scoring places included Flint, Michigan, and Mogadishu, Somalia.

To be clear, I want every country on Earth represented here. There are in fact multiple destinations for every nation, aside from active warzones.

So I tried achieving these simultaneous objectives.

One, for rich countries: find alternatives to overtouristed places. Two, include every country, including those where overtourism is not a real problem. Three, do not penalize busier places in poor countries.

I settled on the 65-35 formula, which works better.

I also turned off regional penalties in poorer countries — penalizing places there for being wealthier than the national average felt ridiculous.

Finally, I promised to be transparent with you.

If my thumb is on the scale, you deserve to know how. I’ve told you what I have, what I don’t, and how my decisions affect the results.

People have a right to scrutinize AI products. Compliment, critique, complain — have at it. I have no doubt people will see nationally famous locations on this list and scoff: Pfff, undiscovered gem? Everyone in my country knows that place!

Fair enough. With luck, more foreigners will know it — and stop piling into the same two places everyone outside your country has heard of.

And I’m not done. I plan to publish a better version by spring, ahead of next summer’s travel season. I’ll post updates here.

If this were a business

I also want to be transparent about what this project is, and what it isn’t. I’m running a blog here as a side-project while I’m at school.

This app is fodder for both – for my blog, and for a couple of assignments for my master’s program in AI Management at Georgetown. Frankly, I’m also doing it for fun — to help plan my own trips.

This is not a travel business. If this were a commercial product, here’s what I’d do differently — what a company building something similar might aim for:

• 50,000 destinations, assisted by Machine Learning. You could use a Convolutional Neural Network to classify images. A Random Forest model trained on my existing destinations could predict scores for new candidates. I’d then run regular tests to measure prediction accuracy.

• Add photos for each spot. And – crucially – links. Links that allow ads for every destination. Ads for local hotels, restaurants and shops.

• SEO and LLM-optimize. For search, you’d structure the data on the page: you’d want a Google crawler to find code explaining that Cordes-sur-Ciel is a PLACE, located at COORDINATES X, in COUNTRY France, with a RATING of 7.7. To draw LLM referrals, you’d want meatier descriptions with more factual details they can cite.

• More features. My database already has details about climate, best months to visit, and relative cost of visiting each place. They’re too patchy to use now, so they’re buried on the back end. I’d build these into the UX.

• Customer-service tools to handle queries from advertisers, and questions or complaints from members of the public.

• Check the terms of service on my data sources to determine whether any need updating for commercial use.

• Testing, testing, testing. I’d be personally checking every destination. And every description of every site. Deleting and editing as needed.

• Fine-tuning place descriptions. They’re almost all written by AI, many by Retrieval Augmented Generation. I’d spot-check them, tweak a few, or send whole batches back to Claude or ChatGPT with instructions. Like: Vary the sentence structure in the following ways; don’t use this adjective so often; don’t guess population stats from your training data, etc. I’d want a human editing every entry.

• Reduce points of failure. I’m relying on several cloud products, all of which rely on their own providers. A point driven home when an Amazon Web Services outage delayed a data upgrade I was working on.

• Security. Rate limiting to block unwanted bot-scraping, automated vulnerability scanning, and incident monitoring.

All that remains hypothetical. Now, back on Earth A, I offer a programming note.

I’d planned to have this post published months ago, but got busier than expected at school. I also encountered a near-comical string of challenges creating this. But more on the limitations of AI in a future post.

In the meantime, happy vacation-planning. Venice might not want you. But thousands of gorgeous other places do. Go find them.

I’m blogging about what I’m studying: practical, ethical uses for AI. I’m studying AI Management at Georgetown University. My posts will be infrequent, at first. As I learn more, I’ll post more. I’d be delighted if you subscribed.

My digital addiction had me by the throat. So I fought back. I turned AI against its algorithms.

A simple app I coded rebuilt my media diet. It’s made my days healthier. I’ve shrunk my screen time, and prioritized quality journalism.

You can do it, too — with zero programming experience.

Every morning around 7:15 a.m., I receive an email with headlines and hyperlinks for multiple news sites, blogs and podcasts.

It’s all there in one message. A custom-built news digest with links I keep clicking throughout the day.

I see it as a stripped-down Twitter feed, denuded of algorithmically driven rage-bait and aimless hours of life-wasting doomscrolling.

In a sense, you’re using AI against itself, or at least against its worst abuses. You’re circumventing the systems that monetize anger and that scavenge human content to be summarized without credit.

Am I using AI to go analog? Not quite, but that’s the spirit.

I’ll provide a step-by-step guide to how I did this. But first I’d like to explain why.

Why I did this

For the first time in my adult life, I’m a retail consumer of news. I’m on a work sabbatical, getting a master’s in Artificial Intelligence Management.

For nearly three thrilling decades in journalism, I’d been a wholesale consumer. I started at a wire desk surrounded by the radio and multiple TVs blaring simultaneously; a printer spitting out copy; and press-release feeds on-screen.

It was a prehistoric Twitter. Those habits stuck, and by the smartphone era I was ingesting news all day.

It filled me with adrenaline at the best of times; too often, lately, it left me distracted and distant from the people I love.

I’d ask myself sometimes: Why do people do this willingly?

If you’re running a campaign war room, I get it, or trading stocks, or have work requiring minute-by-minute awareness of incremental minutiae, but it always struck me as voluntary self-harm for anyone else to spend the whole day mainlining news.

Don’t get me wrong. I love the news. I did, I do, and I always will. Want proof? I just started studying AI and my first project was building an app that… brings me news.

But let’s be blunt: in excess doses, it warps your perspective.

The news doesn’t remind you that the sun came up today, people went to work, and kids made new friends at school. That’s not its role.

The news reminds you the sun causes cancer; people got laid off; and a whackjob shot up a high school.

This is not meant as criticism. We need to know these things! Anomalies are newsworthy. What’s not normal is spending half your day obsessing over anomalies.

This is triply true now that we have this digital albatross in our pocket, ceaselessly pecking at our attention, prodding us to spend that day arguing about news, and click, click, clicking.

That’s the razor-clawed digital bastard I’m keen to escape. Escape may be too strong a word. I still use social media. I still learn things there. I just use it way, way less. Maybe 90% less.

How I did this

I had three objectives as I built this simple app:

• Stay informed. I consider it a civic duty to follow current events — that hasn’t changed.

• Encourage journalism. AI risks harming journalism. Over my dead body will I contribute to that. And not just because I have friends in the field. But also because it’s the seed corn of our information ecosystem. All the things we debate, argue about, and vote on, come from somewhere. Much of it comes from my colleagues, from people who stand outside meetings and sprint down hallways to query public officials. ChatGPT won’t do that for you. Kill the news, and you kill that seed corn. I’m subscribed to multiple outlets. Those links I’m clicking? I paid for them. The ones with a paywall, at least. I want journalism to thrive.

• Reduce screen time. This was big. I wanted more time with my kids. Quality time, engaged time. I’ve already spent enough hours distracted around them and I’ll never get those back. But now when I’m at the playground with them, I’m at the playground. I’m not mentally dwelling upon whatever bullshit du jour an algorithm cooked up while I snuck quick peeks at my phone.

I created an app – vibe-coding, as they say. I’ve hardly ever used Python, and without help I couldn’t code a two-dimensional square.

But with AI as my crutch, I managed. I imitated a barely competent replacement-level computer programmer. So can you!

Here’s how.

Step One: Plan your diet

Decide what you want in your feed – local, national, politics, sports, wellness, science, business, blogs, podcasts, whatever matters to you. Find specific pages you want. Don’t overthink it; you can tweak the mix anytime.

My list, for example, spans local French news from Montreal, national news from Canada, U.S. news, Washington insider politics, international coverage, a bit of hockey, plus blogs and podcasts I follow. My list originally included a bunch of website URLs, but it turns out scraping a home page may violate the terms of service of certain sites, even if it’s for personal, non-commercial use. What you want is a list of publisher-approved RSS links. The New York Times has them for its home page, and its world, U.S. opinion and other sections. The Washington Post has similar feeds. And Fox News, and the BBC, and the CBC, and Canada’s Globe and Mail, and Montreal’s La Presse for politics or local or other news. Just gather the links for the pages you want, plus any podcasts or blogs.

Step Two: Get the code

Now it’s time to generate code. You can use an LLM (ChatGPT, Claude, Gemini). Just describe what you want in plain English and paste your URLs. That’s it – you can adjust later.

Just type something like: “Write a Python script that emails me a daily news digest at [7:15 a.m. — your call]. Pull the top [5? 10? Your call] number of headlines from each of these RSS feeds [paste your links], plus any new posts from these blogs and these podcasts [paste your list] in the last 24 hours. Format the email with separate sections for each outlet. For each story include a headline, a link, and, if made available for use, a subhed or summary. Do not include photo captions as headlines.”

Step Three: Unleash the code

Now you deploy it, so it runs. I used Replit for Steps 2 and 3. I’ve got its $20-a-month plan. There are free plans. You can also try other sites, like Render or PythonAnywhere. I can’t vouch for the others. I used Replit and it did the job. You pop the code into a chat.

It’s time to go live. You’ll be told to switch from Replit’s chat window to its publishing window. Setup is simple: Add your email address, insert a special password for your app, set the email to run at whatever time you want, and hit publish. Done.

You might have to do extra troubleshooting. Don’t worry. It took me a couple of days to eliminate formatting snafus, like photo captions showing up as headlines. I just uploaded images of a problem into the chat and asked Replit to fix it. It did. Not every news site has a solid RSS feed. Most do. For the rest, you might have to drop them.

Is my system perfect? Of course not. It can’t deliver every story I want. But I have backstops: I still get news app alerts, I still click around websites, and, yes, I still spend a few minutes a day on social media.

For example, take a disturbing event from last week. I learned about the Charlie Kirk assassination from multiple news alerts. I clicked on those stories. Then I went on social media to find out what happened.

Within a day, I regretted being on social media. I wanted news about Kirk, not to spend hours boiling in a toxic stew of rumor and rage.

It’s a life project – paring back meaninglessness, to make space for what matters. It won’t ever end. If AI isn’t going away, dammit, we might as well bend it toward something good.

I’ve broken my social-media addiction, and I think that’s unambiguously good. You can too. If you try this approach, reach out with questions. Got tips for using AI ethically and practically? Share them – I want to hear.

—

Subscribe now

This is a soft launch, of sorts. I’ll be blogging about what I’m studying: practical, ethical uses for AI. I’m studying AI Management at Georgetown. My posts will be infrequent, at first. As I learn more, I’ll post more. I’d be delighted if you subscribed.