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.

Andrew Carnegie was the godfather of modern philanthropy. His greatest gift to the world was the Carnegie library: thousands of them, many still standing today. I’ve been to several, dotted across Southwestern Ontario, where I grew up. I’ve seen his portrait still hanging.

Carnegie was also one of the modern world’s first billionaires, a ruthless steel tycoon whose company used armed guards and state militias to break striking workers. And yet his civic legacy is one you can physically stand in.

Compare that with men like Elon Musk and Jeff Bezos, representatives of today’s billionaire class. Their vision for humanity is grand and cosmic: the settlement of space, the survival of the species, the destiny of humanity. But they leave little room for the scale of everyday life.

A new generation of AI billionaires is coming. Nan Ransohoff argues that the technology could unlock hundreds of billions of dollars in new philanthropic capital, perhaps $37–100B per year if current trends hold. That is a lot of money to give away.

Most of this money will come from a milieu infused with rationalism and effective altruism. It is natural to be skeptical of a movement that treats personally accumulating as much wealth as possible as an optimal way to do good in the world. “Earning to give” is an awkward philosophy for a world where private fortunes have already achieved escape velocity. It also does not help that the figure most people now associate with EA is Sam Bankman-Fried.

It seems that every week we read about another compromised package uploaded to PyPI or npm. The tide has become so steady that package managers are starting to implement dependency cooldowns. As a heavy R user, I began to wonder, has a malicious actor ever uploaded a compromised/malicious package to CRAN (the standard repository for R packages)?

I can’t think of any, and a Google search only turns up this weird thread on the R mailing list of an apparent false positive on a vignette bundled with a particular package. Thankfully, there is arXiv paper uploaded last year that directly addresses this question: A Time Series Analysis of Malware Uploads to Programming Language Ecosystems by Ruohonen & Saddiqa. This paper analyzes the Open Source Vulnerabilities (OSV) database for six popular programming language ecosystems: CRAN (R), Go, Maven (Java), npm (JavaScript), PyPI (Python), and RubyGems (Ruby).

Table 1 from the paper shows CRAN to be completely free of malware entries, with Maven having only 1 and Go having 8. RubyGems has nearly a thousand and npm and PyPI have many thousands each. This is not to say packages uploaded to CRAN have never had vulnerabilities: the OSV database currently lists 14 for CRAN packages. But there is no evidence that any of these vulnerabilities were introduced with malicious intent.

This article from epidemiologist Adam Kucharski describes a neat sanity check he conducted on Microsoft Copilot. In the first experiment, he simulated 2,000 free text responses and labelled them “US”. Next, he copy and pasted these responses and labelled them “UK”. He then randomized the order and passed the combined dataset to Copilot to analyze. In the second experiment, he simulated 200 free text responses but copy and pasted them five times, assigning five different country labels to the otherwise identical datasets before again passing them to Copilot.

In both experiments, Copilot returned a deep analysis of the differences in how participants from each country responded to the prompts. The only problem, of course, is that the responses were identical between all countries: there were no actual differences to describe.

Instead, Copilot leaned on cultural stereotypes to give the expected shape of an answer. This should not be surprising; indeed, the only thing LLMs can do is give the expected shape of an answer. The surprising thing is that these answer-shaped responses are correct and useful as often as they are.

Of course, some will object to the use of Microsoft Copilot, a famously weak and outdated model. But Copilot is a widely deployed enterprise tool, and so for many users it will be their primary exposure to AI in a work context. In this experiment, Copilot was asked, using default settings, to perform a task it is explicitly advertised as being capable of doing. Defaults are powerful, and such an insidious failure mode is sure to cause harm, especially in the hands of ordinary users who lack the understanding and intuition to detect the common failure modes of these AI systems.

It is fairly well known that tomatoes are culinary vegetables but botanical fruit. Somewhat less well known is that fruits like raspberries, strawberries, and blackberries are not berries, because they do not develop from a single flower containing one ovary. (For example, the yellow “seeds” on the exterior of the strawberry are the true fruit of the plant.) This article by Ada McVean and Cassandra Lee for McGill’s Office for Science and Society relays an even less well-known fact: bananas are berries, botanically speaking, since they are fleshy fruits developing from a single flower containing one ovary. As are cucumbers, pumpkins, lemons, avocados, watermelons, and grapes.

At least blueberries are, in fact, berries.

Apparently all of this recently came to a head when one angry Tesco shopper in the UK was denied his or her bonus points for purchasing berries. Reportedly, neither bananas nor strawberries were valid for the offer, meaning the store was applying neither the common definition for berries nor the botanical one.

Taxonomy is always interesting. As a biology major, I especially enjoyed cladistics, which offers up surprising conclusions like “birds are reptiles”, “birds are also dinosaurs”, and “fish are not a coherent group”. It reminds us that the categories we use to navigate the world are often practical rather than natural—and that nature, when asked to sort itself neatly, tends to shrug.

I replayed Dragon Age: Origins and its DLCs on PC about a year and a half ago, and it still slaps.

Let’s start with the obvious: the cover art is awesome. Iconic. My favourite cover art of any video game.

^{Actually, the game’s soundtrack has an even better version of the cover art.}

The music is also great, especially its GOATed title track by composer Inon Zur. Imagine being a high schooler firing up a new game and being greeted with that. RPG players talk about being stuck on the character creation screen, but that music kept me glued to the main menu.

I still like the graphics, too. I know they are “ugly”, but they are ugly in a way I find comforting. DA:O belonged to the last era of gaming before graphical fidelity basically saturated and every big-budget game started to look the same. It wears its dark fantasy pretensions in various shades of brown and sprays of blood. It looks like 2009 in the best possible way.

Of course, good presentation means little for an RPG if the world is boring or the characters are unlikable. Thankfully, DA:O works because it is so completely unselfconscious about being a dark fantasy RPG, and the characters are better for it. They are prickly, well-voiced, and rooted in the setting. Each character carries a piece of the world without being reduced to it.

I recently came across this article “Is logistic regression regression?” by data scientist Richard Vale. The post is a response to an apparent view in the machine learning community that “logistic regression” is a misnomer, because it is used for classification problems (with categorical outcomes), not regression problems (with real-valued outcomes).

As someone who learned statistics before machine learning, by way of biology and then epidemiology, this question would never have occurred to me. I was taught to group models by outcome type—logistic regression for binary outcomes, linear regression for continuous outcomes, multinomial regression for unordered categorical outcomes, Cox regression for time-to-event outcomes, and so on. Of course logistic regression is regression. It is regression with a binary outcome and a logit link. And besides (as Vale points out), logistic regression does not output a category label. Its raw output is log-odds, which may be transformed into a probability. “Classification” is a separate step layered on top, usually by imposing a threshold of 0.5.

This point goes to the statisticians.

I was not taught to think in terms of two types of predictive modelling: regression and classification. In fact, I’m not sure I ever came across the term “classification” until I picked up The Elements of Statistical Learning (okay, it was actually An Introduction to Statistical Learning). Rather, I was taught to approach modelling based on what I was trying to achieve: the type of research question. In this framework, there are three basic types of questions: descriptive (how are things distributed?), predictive (what outcome can we expect?), and causal (what difference would an exposure/intervention make?).

I was listening to Ezra Klein’s latest interview with Yuval Noah Harari and was interested by the section on AI relationships (emphasis mine):

HARARI: AIs, which are experts in pretending to be conscious entities that have feelings for you. And it’s relatively easy for them to do it because maybe the most important way for people to kind of build relationships is language. So, you know, when an AI tells you, I love you. It’s not like a science fiction movie from the 1960s when it does so in a very cold, mechanic way and doesn’t really understand what love is. No, it does so in the most seductive voice possible. And then when you ask the AI, do you really love me? Do you even know what love means?

The AI can give you the most amazing description of how love feels like because it has mastered language and it has read All the best love poems in history, all the psychology books about love, all the blogs, it’s have seen all the Hollywood blockbusters about love. It can describe love better than almost any human poet or psychologist or lover.

KLEIN: In this respect, it’s able to sever language from meaning. […] When an AI says, I love you, it does not mean what it means when a human says, I love you. There’s not an I behind that.

…

I saw this post a while ago, “Fellow Artists, I’m Begging You to Pull Your Heads Out of the Sand About AI”, but I decided to write about it today because there’s a passage in Freddie deBoer’s post today on LLMs that almost perfectly encapsulated my own feelings on the subject.

The post from Substacker funplings exhorts artists to abandon the idea that AI art can be reliably distinguished from the real thing purely by inspection. This part I agree with—much like with yesterday’s post on Google AI Overviews, the underlying technology has been quietly improving to the point that it’s no longer quite so easy to dunk on obvious flaws like too many fingers or garbled text or the piss filter.

To illustrate this point, the article points to an experiment by a random Twitter user who posted the following:

The gotcha here is that this painting is, of course, a genuine Monet. And yet commenters were happy to describe in detail why the supposedly ersatz image was inferior to the real thing.

Sometimes you read a piece that annoys you because you wish you wrote it yourself. “I’m tired of talking to AI” by anonymous developer Orchid is one of those pieces. The short post ends like this:

I’m tired of talking to AI. I want to talk to real people. But even when I talk to people, they forward my questions to AI and send me the AI’s answer.

Every company is shoving AI down our throats now. One of the most prominent examples of this is Google’s AI Overviews, which pops up on damn near every search you do. In the early days, it was pretty bad and easy to make fun of. But annoyingly, it’s now gotten pretty good. Arvind Narayanan from AI as Normal Technology put it well in a post from earlier this year:

I think our disdain for companies “shoving AI down our throats” is largely a selection effect — when one of these AI integrations is new and experimental, we tend to notice, but over time the kinks get worked out, it becomes a part of our workflow, and we stop noticing it. Reminds me of the classic quip that “AI is whatever doesn’t work yet.”

But Google AI Overviews is also Reddit.

Reddit has annoyed me for literally decades now, which is why I never made an account despite for years appending “reddit” to the end of my Google searches when I wanted to get a usable answer to certain types of questions. It annoys me because I was participating in Internet forums since long before I was old enough to legally do so. Reddit centralized discussion on the Internet and replaced all the old forums, homogenizing their unique cultures into a single grey blob.