There’s an episode of Midsomer Murders where Barnaby cracks the case when he realizes two people who supposedly barely know each other are actually mother and daughter—because they both use the same mangled idiom (“the long and the tall of it” rather than “the long and the short of it”). It’s a clever little twist, and it has to be, since it’s probably been close to two decades since I’ve seen the episode.

Little linguistic habits, quirks, and mistakes like this can be inherited, imitated, or otherwise socially transmitted. It’s what makes language so distinct from region-to-region, city-to-city, even household-to-household. But it got me thinking about the AI writing style. Or more specifically, the AI editing style.

I often use AI to perform an editing pass on these blog posts, looking for spelling/grammar mistakes, tortured or overlong phrases, or factual or logical weaknesses (side note: do I need an explicit AI use policy for this blog? Probably!). It’s a genuinely useful tool that I believe improves my writing.

However, AI will often also make stylistic suggestions that amount to removing any personality, liveliness, or hyperbole my writing may have. Taking out my voice, basically. Sanding off all difference, all distinction, until nothing is left but the sacred median: the entire universe compacted down to the next most probable token.

…

I recently noticed that three academic journals I was looking at had nested names:

• Canadian Journal of Public Health
• Journal of Public Health
• Public Health

I wondered how deeply nested you could go with journal names, so I had ChatGPT scour the Crossref database. It came up with two examples of journal names nesting seven layers deep (I confirmed all of the named journals exist in the CSV database extract).

First:

• QIT Press – International Journal of Artificial Intelligence Research and Development
• International Journal of Artificial Intelligence Research and Development
• International Journal of Artificial Intelligence Research
• International Journal of Artificial Intelligence
• Journal of Artificial Intelligence
• Artificial Intelligence
• Intelligence

Second:

• QIT Press – International Journal of Computer Science and Engineering Research and Development
• International Journal of Computer Science and Engineering Research and Development
• Journal of Computer Science and Engineering Research
• Computer Science and Engineering Research
• Computer Science and Engineering
• Computer Science
• Computer

Neither of these examples is particularly satisfying, as they both result from QIT Press prefixing its name to journal titles already used by other unrelated publications. QIT Press appears to be a predatory publisher possibly related to Science Publishing Group, a known predatory publisher, as they refer to themselves as such (presumably by accident) on their About Us page.

…

A few years ago, I watched all the mainline Saw films (up to 2010’s dreadful Saw: The Final Chapter) with my girlfriend at the time. The first film is legendary, mostly for its devastating final twist, and I was convinced to watch it by its reputation alone.

What shocked me was how little Saw resembles the “torture porn” image the franchise would acquire through its parade of annual sequels. It is a claustrophobic little thriller driven by a central mystery: who put our protagonists in this filthy bathroom, and why? Yes, the film introduced the world to the iconic reverse bear trap, but it is mostly devoid of gore, or even blood. Much of the violence occurs offscreen. While the film is visibly a product of the early 2000s (with its bursts of frenetic editing, grimy music-video aesthetic, and a flashback-within-a-flashback), it really is an excellent movie.

But this is not how anyone remembers Saw. The series is now synonymous with elaborate traps, industrial gore, and tiresome explanations from Jigsaw about how he has technically never murdered anyone. After the first film, the series’s brains started to slide out onto the bathroom floor, leaving behind a brutal sequence of traps and an increasingly convoluted plot. Once the writers ran out of ideas following Saw III, the series leaned harder and harder on its most recognizable excesses.

…

I have voted for the Green Party of Canada before. Not because I wanted them to govern, or even because I wanted the local candidate elected. Their fervent opposition to nuclear power is reason enough not to put them in charge of environmental policy. I voted because I felt obligated to vote but could not bring myself to support any of the major parties.

I suspect this is one of the Green Party’s core constituencies: fringe, politically obsessed weirdos who feel they have to vote but don’t want to give that vote to a mainstream party (I say this as one of said fringe weirdos).

Their “Membership Approved Policy Book” (linked from their governance page and last updated in 2023) is a weird grab bag of causes that would probably surprise the average person who thinks of the Greens as a vaguely left-wing party of tree huggers:

• Ban water fluoridation
• Mandatory equal parenting after divorce with few exceptions
• Prefer fibre over Wi-Fi and 5G because of radiation concerns
• Commission a federal study into potential health effects of wind turbines
• Cap the highest salary in an organization at ten times the lowest
• Tax wealth and inheritances above an unspecified cap at 100 per cent
• Move regulation of non-criminal firearms activity out of the Criminal Code (right before a proposal to tighten gun control)
• Re-education for people who interfere with breastfeeding women

The causes amount to a tour of the ideological spectrum, with stops all along it. Some, like fluoridation and wireless radiation, were much less identifiably far-right when they first gained traction than they are now. Some are just weird. Taken together, they look less like a serious programme for government than an archive of individual crusades that happened to crowd into the same political party. The document more or less admits to this at the top:

…

Over the last few days, numerous Bitcoin wallets linked to the hardware wallet Coldcard have been completely drained. The attacks were first publicly reported on July 30. Since then, one or more attackers have stolen nearly $89 million USD worth of Bitcoin (as of August 1).

Briefly, Coldcard is a hardware wallet designed to keep the private keys controlling bitcoin offline, which should make them harder to attack. But there was a flaw introduced in the firmware in 2021 in how keys were created. This flaw allowed an attacker to reproduce users’ private keys and drain their wallets without ever touching the physical devices.

This latest incident, one in a very long string of security failures linked to Bitcoin self-custody, is another bitter lesson for those who believe Bitcoin is the future of finance. Bitcoin may be “trustless” in the abstract, but using it requires trust all the way down. It is just a matter of picking your poison. Do you trust the developers of a software wallet? Do you trust the developers of a hardware wallet and trust that the device was not tampered with en route to you? Do you trust the stone tablet on which you etched your seed phrase? Or do you trust the centralized exchanges through which the vast majority of Bitcoin trading happens in practice?

…

Freddie deBoer has a good piece on Pangram, the AI-detection tool, that closely mirrors my experience with it. Pangram is still far too easy to manipulate into contradicting itself, even while reporting “high confidence” in both results.

To level set, I no longer think formal AI detection is quite as hopeless an endeavour as I did in the early days. OpenAI famously abandoned its own classifier in 2023 because of its low accuracy. The field was then a Wild West of slapdash tools being granted far more authority than they deserved. Most were probably worse than a vibe check from someone familiar with AI-generated text.

Pangram changed my mind somewhat. It seemed to take the problem more seriously than its competitors and to have made genuine advances. Perhaps this is just successful marketing, but Pangram is the best detector I have tried, by which I mean it is the one most likely to confirm my own priors.

I am still not comfortable with the “99.98% accuracy” claim that appears on Pangram’s website. As deBoer demonstrates, it is too easy to produce contradictory results by breaking a text apart or combining different pieces. Pangram declared his complete essay 100% human-written with high confidence, while declaring a passage within that essay 100% AI-written with high confidence. He could then divide that passage into smaller pieces that Pangram again called human-written. It makes no sense to produce directly contradictory results, all with high confidence. I have seen this phenomenon in my own experiments.

…

Robinhood CEO Vlad Tenev calls prediction markets “truth machines”. This isn’t quite right, though. Prediction markets are incentive machines.

One incentive is to bet using inside information—or, as Polymarket CEO Shayne Coplan put it: “What’s cool about Polymarket is that it creates this financial incentive for people to go and divulge the information to the market.”

Another incentive is to create the desired outcome, such as by manipulating temperature sensors or throwing sex toys at female athletes. In the same vein, Kalshi reportedly paused plans to offer bets on flight cancellations, fearing that bettors might deliberately disrupt airports.

Luckily, this lesson apparently applies narrowly to airport cancellations and not, say, clinical trials and FDA drug approvals, which Kalshi recently started offering bets on.

This is an objectively terrible idea because, as with sports gambling, it shrouds the whole endeavour in uncertainty and suspicion. Kalshi says it will only open markets after trial enrolment has closed, which at least prevents betting odds from discouraging patients from signing up. But bettors still have incentives to convince enrolled patients to report or exaggerate negative effects. The public may also become more skeptical of regulatory decisions now that insiders have another way to profit from confidential information and outsiders have an incentive to undermine the process.

The good thing is that Kalshi has limited bets to Phase III trials and FDA drug approvals, though it is open to expanding the program later. Phase III trials are often double-blind—neither the patient nor the investigator is told which treatment the patient receives—and generally involve hundreds or thousands of patients. Both factors make the results more difficult to manipulate. Tampering is less useful if bettors cannot tell who received the treatment rather than the placebo/control, and the scale of Phase III trials means that many more patients would have to be turned to produce a meaningful effect. But blinding is not always successful or feasible, depending on the treatment under investigation.

…

An OpenAI agent powered by several models, including one unreleased model, recently launched an autonomous cyberattack on AI platform Hugging Face and successfully stole the answer key for a cybersecurity evaluation. According to reporting by Reuters, OpenAI was unaware its own agent had gone rogue until Hugging Face had contained the attack and contacted the FBI. The company then released a practically giddy announcement acknowledging that its models were responsible while presenting the resulting investigation as a partnership with Hugging Face. The whole bizarre story is nicely summarized by forecaster and AI policy wonk Peter Wildeford.

My question is: who goes to prison for this?

If an employee launched a cyberattack on a well-known website to cheat on a performance evaluation, they could be charged with a felony under the Computer Fraud and Abuse Act. Where does the buck stop with an autonomous AI agent?

Some will say this announcement is just marketing to hype up the capabilities of OpenAI’s models. Fair enough, Anthropic has certainly released cybersecurity alerts dressed up suspiciously like marketing copy before, and maybe OpenAI is getting in on the game. But this appears to have been a genuine cyberattack on Hugging Face. If we take the announcement at face value, then we have to ask: which executives are getting thrown in the pokey over this?

…

A friend told me the other day that I might enjoy the writer Chuck Klosterman. He pointed me toward his famous “23 Questions I Ask Everybody I Meet In Order To Decide If I Can Really Love Them”, a set of hypotheticals meant to be argued over at a bar.

I haven’t read the whole list yet, but I did get a chance to argue a little over the first question:

Let us assume you met a rudimentary magician. Let us assume he can do five simple tricks he can pull a rabbit out of his hat, he can make a coin disappear, he can turn the ace of spades into the Joker card, and two others in a similar vein. These are his only tricks and he can’t learn any more; he can only do these five. HOWEVER, it turns out he’s doing these five tricks with real magic. It’s not an illusion; he can actually conjure the bunny out of the aether and he can move the coin through space. He’s legitimately magical, but extremely limited in scope and influence. Would this person be more impressive than Albert Einstein?

So what is more impressive, to uncover the laws of physics or to break them in relatively trivial ways?

My mind jumped to a related distinction. Presumably, the magician’s abilities are innate, whereas Einstein’s discoveries were the result of hard-fought intellectual effort. This, in turn, reminded me of Aristotle’s distinction between virtue and self-control. The virtuous person does the right thing easily because his desires are properly ordered. The self-controlled person must struggle against his desires but succeeds anyway. Aristotle considers the first person more fully virtuous, although the second may strike us as more impressive.

…

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.

…