I read a lot of Tom Clancy novels as a teenager. Something about the extremely detailed writing and overly elaborate plotting scratched an itch. I was also a big fan of the Splinter Cell games.

The 1996 novel Executive Orders—the one where Clancy’s self-insert hero Jack Ryan becomes president—has always stuck with me. Like all Clancy doorstoppers (this one is nearly half a million words), it has many subplots. The most memorable one involves terrorists backed by the Ayatollah of Iran starting an epidemic in the United States using aerosolized Ebola. The attack has two purposes: to kill Americans and to paralyze the country so it cannot respond to Iran’s invasion of its neighbours.

What makes the subplot memorable is that the biological attack just…doesn’t work that well. It kills thousands instead of millions. It is fairly successful at preventing the normal deployment of American military forces to the Middle East, but luckily for America’s allies, the available units are sufficiently badass to hold off the Iranian invasion anyway.

The epidemic fails for two reasons. First, although the terrorists successfully aerosolize the virus in shaving-cream canisters and release it at roughly twenty conventions and trade shows across the country, Ebola simply does not transmit very well under ordinary circumstances. Most of the secondary infections come from close personal contact.

I recently read this post from Paul Goldsmith-Pinkham on why the number of economics preprints submitted to arXiv and similar preprint servers is not exploding quite as dramatically as some charts suggest.

Basically, if you plot papers by the date they were last updated, you create a large upward bias in the most recent months. Old papers revised today are counted alongside papers first submitted today. The chart will therefore show a sudden explosion near the present whenever you produce it.

If you plot the papers by their first submission date instead, and recent growth looks much closer to the historical trend. Goldsmith-Pinkham still found that the number of new papers was growing faster across most of the fields and preprint servers he examined, but the apparent vertical takeoff was partly an artifact of the date variable.

This all reminded me of one of my COVID-19 pandemic hobbyhorses, which annoyed me enough that I wrote a paper about it: which date variable to use when plotting an epidemic curve.

Epidemic curves are commonly plotted by either date of symptom onset or date of public reporting. Symptom onset is closer to the actual infection, so it intuitively seems to offer a more “real-time” view of the epidemic. But recent onset dates are necessarily incomplete. When someone develops symptoms, it takes time for them to seek testing, book the test, receive a result, and have that result communicated to public health authorities. A case reported today may therefore have developed symptoms days or even weeks ago. New cases are continually added to the older part of the symptom-onset curve.

The other day I was curious what kind of hockey games were on Steam and came across this review of Franchise Hockey Manager 12 by someone who had built an AI assistant general manager for his OHL team.

The game can export an entire save as CSV: roughly two million rows across 48 tables, including rosters, statistics, player ratings, and scouting grades. The player wrote a Python script to load each export into SQLite, preserving snapshots from different in-game dates, then put a local MCP server on top of it. With this, his AI assistant GM can query the actual save, track changes in prospect ratings, build draft boards, grade draft classes, write monthly reports, and score the roster against a custom 100-point rebuilding rubric. It even noticed that a goalie prospect had lost half a star of potential because the coach was not starting him.

I guarantee this guy is having more fun playing the game than anyone else. Players of simulation games are known to be obsessive, but this guy built a full data warehouse for his virtual junior hockey league that is probably more sophisticated than the data warehouses at some actual corporations.

The interesting part is that none of it starts with AI. It starts with an “Export CSV” button. All of this is possible because the studio made the underlying game data available in a legible format; one obsessive player did the rest. Save files are often just compressed JSON and can be extracted and parsed with enough effort, but that is not the same as providing a supported export of nearly everything the game knows. As the reviewer put it: “Franchise Hockey Manager 12 quietly has one of the best features in sports sims: Export CSV”.

DAOs (decentralized autonomous organizations) are organizations governed by votes on the blockchain, typically with votes proportional to tokens held. Of course, this being crypto, any supposed purpose is usually overshadowed by speculation on the price of the underlying coin.

DAOs have a storied history of failure. The original DAO caused the second-largest blockchain to split in two over whether to undermine the fundamental promise of crypto to bail out rich token holders (the yes side won). Another DAO raised money to buy a copy of the US Constitution and failed, only to discover that issuing refunds is really expensive in crypto.

But today’s story is one of a DAO working exactly as intended.

BonkDAO governs BONK, yet another dog-themed meme coin. It controlled a treasury worth about $20 million USD. An anonymous wallet spent $4.4 million USD buying just over 1% of the total supply of BONK. This was somehow enough to meet the DAO’s quorum requirement all by itself.

The wallet proposed transferring the entire treasury to another address it controlled. Only seven wallets bothered to vote, and only two voted yes: the creator of the proposal and one other random wallet. But that was enough for the proposal to pass with 99.9% of the voting power. The transfer then executed automatically.

BonkDAO has called this an attack and contacted law enforcement, which feels a little unsporting. There was no hack and no code exploit. The so-called attacker bought the votes, proposed giving himself all the money, won, and collected.

Yesterday I used ChatGPT to generate a dumb parody logo for a post. I was never going to pay an artist to make it. Without AI, there simply would have been no image, no extra joke.

On the other hand, I feel a sort of visceral disgust when I see AI art in online ads or restaurant food pictures or business signs, at least when they use a generic slop style. When deployed thoughtlessly, the outputs tend to have a slightly sickly quality to them. There’s a really weird trend of faux-claymation YouTube ads right now; before that, it was weird, overwrought AI-generated songs over generic, sad-looking slop people.

I am not broadly anti-AI. There are corners of the internet that would dismiss me simply for acknowledging any use of AI in my work. But I also understand why artists are angry. When your livelihood depends on making images, a tool that can produce an unlimited number of them for almost no money is not an abstract philosophical problem.

At the same time, I find myself making a distinction between uses that replace paid work and uses where the realistic alternative was nothing. Like, do I think Freddie deBoer, hardly an AI booster, is committing a grave sin when he sometimes uses AI to make a jokey blog header? He said it works well with his spontaneous writing style, though he checks Getty Images first. For my old newsletter, I used to get header images mainly from free sites like Unsplash or Pexels. So while I was never going to pay anyone, I guess I could be denying an artist exposure or discovery if I turn to an AI-generated image.

^{Parody logo generated by ChatGPT.}

Stake is everywhere online.

The logo for the online, crypto-based casino appears everywhere: on streams, sports highlights, clip-farm accounts, and the endless slurry of short-form video content. The videos often have nothing to do with gambling.

The online streaming platform Kick, where streamers go after being banished from the more “respectable” Twitch, is one of the most watched video platforms in the world. It was created to funnel viewers onto Stake’s online gambling site. Your teenage son’s favourite content creator is a man sitting in his bedroom losing thousands of dollars a day on the site (using the house’s money).

There are still small oases. YouTube banned the promotion of some gambling sites like Stake. But Stake is escaping into the real world too, sponsoring a Formula 1 team, a Premier League club, and esports teams. Its logo was plastered across the octagon at the UFC fight on the White House lawn.

The old promise of the internet was that creators would no longer need television networks, record labels, or publishers to reach an audience. Well, we got our wish. Gambling sites have swallowed the entire chain, from the people making the content, to the accounts sharing it, to the platforms hosting it.

That’s right, I’m mad at myself. I was Time’s Person of the Year, and so were you, assuming you were born before 2006. Yes, Time’s Person of the Year for 2006 was… “You”.

Time magazine is fond of these collective awards. Hardly a decade goes by without one or two. Some are more inspired than others, but to be quite honest, collective awards are weak and cowardly choices as a category.

I once considered studying under a professor who listed a Nobel Peace Prize on his CV. He wasn’t wrong: he was, in fact, part of the International Physicians for the Prevention of Nuclear War when they won the Nobel Peace Prize in 1985. But that doesn’t make these mushy selections any more satisfying.

The contemporaneous reaction to my (and your) selection as Time’s Person of the Year was…not positive.

What a cop-out! They couldn’t decide amongst all of the wack — jobs in the news today, so they gave up and chose YOU. What a joke. — Craig Chicago, IL

Many commenters pointed toward one of two alternative choices: US President George W. Bush (for the War in Iraq) or Iranian President Mahmoud Ahmadinejad (who was causing quite a stir at the time).

With the benefit of hindsight, Al Gore would have been the best choice. An Inconvenient Truth came out that year, and Gore won the Nobel Peace Prize the the following year. Nobody did more than Gore to make climate change a permanent fixture of Western politics.

I recently asked an LLM to infer my ideological leanings purely through my post history on this blog. It pegged me as, among other things, a technocrat.

Okay. I am all for science and expertise. I think public policy should be grounded in evidence. I want a competent administrative state. But I don’t think the label quite fits.

Getting a PhD made me more skeptical of scientific evidence, not less. The process of evidence generation is full of choices: how to frame the question, what to measure, who to cite, which limitations to admit, when to stop fighting reviewer two, and why anyone should care in the first place. Scientific results are rarely clean, at least not the kind that end up informing policy.

This is also why I am cranky about certain phrases. I prefer the term “evidence-informed” over “evidence-based”, because evidence can be an input into the decision-making process but cannot define the values that guide it.

The same goes for “follow the science”. This slogan annoyed me so much I even wrote a piece about it in The Hub a few years ago. The key passage:

In reality, “science” cannot dictate policy, because science cannot weigh the value of particular freedoms against specific risks, any more than it can tell you how to feel about a sunset. Policy is about making trade-offs, based on a set of values and goals, in the light of evidence.

…

Just before the pandemic started, I attended a talk on the ethics of genome editing using CRISPR. At the time, the main question on my mind was: What about China? What’s stopping them from using CRISPR to create an army of genetically modified super soldiers and taking over the world? And how should that threat change what we in the West were willing to do?

China was a convenient foil: infinitely ambitious, maximally ruthless, totally amoral. But now I am not so sure. This Richard Hanania post from 2024 pretty directly confronted the fears I had had a few years earlier:

Supporters of genetic engineering and embryo selection have sometimes argued that China is soon going to be creating super babies and so will eventually take over the world. Supposedly, it is our Christian morality and woke ethics that prevent us from realizing all the benefits of such practices, but a smart and pragmatic people like the Chinese can’t be expected to make the same mistake.

He goes on to point out that China is more restrictive than the US on biotech policy. Research connecting genetics to IQ is restricted. The first Chinese scientist to use CRISPR to create a genetically edited baby was thrown in prison, while his foreign collaborators went unpunished. Surrogacy is de facto illegal. The United States, and even Europe, are generally more permissive on genetic technologies than China. The fears about Chinese super soldiers have not come to pass.

Looks like that probably AI-generated story that won the Caribbean category of the Commonwealth Short Story Prize last month has won the overall prize, too.

Writer Vincenzo Barney predicted as much in his piece for Unherd shortly after the original controversy around the Caribbean category winner:

And if all of these finalists are real, and their work is their own, their writing operates by the same rules, and is still trash. Of the five finalists — one of whom is the verifiably real Lisa-Anne Julien, who has been published since 2021 and appears animated and otherwise alert and not-computer-generated in videos — the overall winner will be announced on June 30. For my money, I like Nazir’s work the best.

If you’ve ever tried using AI for fiction, you’ll know that it produces text that is often overwritten, with metaphors that sound deep but don’t actually mean anything. Several of the winning story’s passages in the quoted article have this unmistakable shape.

Barney comes to the same diagnosis for the degradation of writing as I did for the degradation of Internet culture: optimization for algorithms.

The writer was not hooking a reader anymore with an anecdotal lede, but hooking Google’s algorithm, which prioritizes algorithmic styles of writing, styles of writing that a computer can understand and parse. We have been reading and writing in this medium for three decades, watching the birth of the one-sentence opening paragraph — a contradiction in terms — in journalism, and the extinction of voice and style across all genres.

…

If you’ve ever had to fill out a digital form for the Canadian government, there’s a good chance you’ve come across an XFA PDF. These are essentially fancy fillable forms with features like validation, dynamic fields, and barcode generation. But if you’re a Linux user, you’ve probably seen this message instead:

If this message is not eventually replaced by the proper contents of the document, your PDF viewer may not be able to display this type of document.

You can upgrade to the latest version of Adobe Reader for Windows®, Mac, or Linux® by visiting http://www.adobe.com/go/reader_download.

Adobe may respect the Linux trademark, but they don’t respect Linux users. Adobe Reader for Linux was abandoned in 2013. So what’s a guy living in the eternal year of the Linux desktop to do?

Let’s consider our options. Alternative PDF readers? No. Chrome? Nope. Firefox? It can open the files but not save any of the changes. The online Adobe Reader? Nope. Adobe mobile apps? Nope! It’s only Adobe Reader for desktop. But the Linux version is so old it no longer appears on Adobe’s website.

The solution? Use Wine to run the Windows version, of course! The website linuxconfig.org provides a nice tutorial on how to do just this. I can confirm that it works, at least on my Pop!_OS 22.04 install.

By now, p-hacking is a familiar research vice. It can be done with varying degrees of intent, from the naive beginner poking around in their data to the celebrity professor who makes a career of it. Basically, p-hacking occurs when researchers turn their datasets upside down and shake them until p < 0.05 falls out. This is generally accomplished by slicing the data in different ways, adding and removing variables, and running different kinds of tests until statistical significance is achieved. Usually it is accompanied with a healthy dose of selective reporting and HARKing (Hypothesizing After the Results are Known).

Less often discussed is seed hacking (or seed optimization/selection/scanning), as described in the humorously titled paper “torch.manual seed(3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision” by David Picard. Basically, Picard built models on two of the classic datasets in computer vision (CIFAR and ImageNet), varying only the seed between training runs. He found that while overall variability between seeds was low, he could find “lucky” seeds that produced increases in validation accuracy that would be considered important improvements in the research community, despite being entirely due to luck. He concludes:

I am definitely not saying that all recent publications in computer vision are the result of lucky seed optimization. This is clearly not the case, these methods work. However, in the light of this short study, I am inclined to believe that many results are overstated due to implicit seed selection - be it from common experimental practice of trial and error or of the “evolutionary pressure” that peer review exerts on them.

…

Alvin Chang has a lovely new piece in The Pudding about what the pandemic did to American relationships. It is based on “How Couples Meet and Stay Together”, the Stanford survey by Michael Rosenfeld, Reuben Thomas, and Sonia Hausen, which followed couples in 2017, 2020, and 2022.

One of the piece’s main conclusions is that the pandemic widened existing gaps: strong relationships often got stronger, while weaker ones were more likely to deteriorate. I suppose this pattern repeated itself across other domains, such as the huge wealth gains that flowed to the already wealthy during the same period.

It is a remarkably engaging piece of data journalism, with little claymation figurines scrambling to rearrange themselves with every story beat.

There are some caveats. The stylish data presentation uses absolute numbers rather than relative proportions, which can occasionally make it hard to see some of the relationships being described. It also obscures some of the more sophisticated analysis behind the piece, which uses survey weights rather than raw counts. The main visualization only includes participants who were reached at all three time points, which was about half of the original sample. The original manuscript this paper is based on accounts for attrition using survey weights, and argues that attrition is not an important source of bias in any case.