news (external)

We all agree that websites should be accessible, but what's the best way to accomplish this? Chris Coyier considers two options: either building accessibility into the browser, or building it into the website. Both options have a cost, he suggests. The better option is to build it into browsers, since the reach would be universal, but this means charging the browser users who need the feature, and they tend not to have a lot of money. So companies have chosen to focus on making websites accessible, a much more expensive and less universal option (and one that needs to be backed by legislation), because people who create websites tend to have more money. Browsers are expensive to develop and maintain - there has been a lot of discussion about Mozilla's acquisition of an advertising company (and the addition of the Privacy Preserving Attribution API in its latest release (see this thread)). And yet they are essential public infrastructure - the sort of thing a public service should provide, so we don't have to depend on corporate sponsorship to (say) get the news. Or pay extra if we want to browse the web while blind. Image: Intuit.

When researchers study something through the "lens" of some theory, what are they doing? They are investigating one property by studying another property that some theory (or more specifically, model) says is covariant with it. For example, we don't study pressure directly; we use an instrument, such as a barometer, which varies as pressure varies. That's fine for physical theories where we can easily trace a causal relationship between the two, but what about, say, mental phenomena? A person can feel 'under pressure', but we can see the signs - sweaty forehead, for example. How do we know what we are measuring even exists? Or as Mahmoud Jalloh puts the question, "what if my measurement process is measuring something other than the intended measurand?" How do we calibrate for error? Jalloh's discussion applies to high-pressure physics, and he argues that two theories can calibrate against each other if they are measuring across the same dimension (aka the principle of dimensional homogeneity). But what even are we measuring when we measure mental pressure? We have things like the perceived stress scale (PSS), but what values do the numbers represent? All food for thought. This is a dense technical paper that can be a difficult read, but the first half especially is valuable if you are deeply interested in the concept. Image: DALL-E 2.

According to this article, "Researchers from Monash University have surveyed 1000 primary and secondary teachers across the country and their results reveal what health and wellbeing measures educators consider essential at work." The result (13 page PDF) is a nice little list of things like safety, respect, autonomy and trust. It's important, I think, not to misconstrue what is being said here. There is no discussion of compensation or pay whatsoever, so the results apply only "to foster thriving educational environments for all teachers," as the authors write, and not "solving the teacher shortage crisis through recruiting and retaining teachers," as the Education Review summary suggests. It's a conceptual exercise; wellbeing is explicitly contrasted with 'fear' in the instructions. P.S. I like the use of the CRediT authorship contribution statement at the end, which makes it clear that the paper had one author (Duyen T. Vo) and two supervisors (Kelly-Ann Allen, Andrea Reupert) all of whom appear in the byline.

What might AI do for us personally (as opposed to generating generic content)? Not a lot so far but there are hints. Matthias Melcher explores the topic. "For me, the most important 'mind-numbing' (@jarango) activity that AI might do for me, is re-reading stuff. Stuff that I have already read or written myself, but which has grown too much about a given topic. So I would want AI to sort it into subtopics." That sounds useful. Also: "find texts or code on my hard drive that I remember only too vaguely." So useful!

This is from a few days ago but I left it on my computer before I went on my short vacation because I wanted to make sure it was noted. Donald Clark writes, "In a rather astonishing blind trial study (markers were unaware) by Scarfe (2024) (33 page PDF), they inserted GenAI written submissions into an existing examination system... (the results): 94% AI submissions undetected; AI submission grades on average half grade higher than students." Unlike Clark, I don't actually consider that shocking, not simply because AI is good, but also (mainly) because assessments today are basically language tests. We give them a bunch of language (direct instruction, readings, other content) and then ask them to perform a generative language task (multiple choice, short answer, essay). What do we think the result is going to be?

Readers my age will remember Darrell Huff's How to Lie With Statistics (73 page PDF), the examples in which were out of date even when I was in university. This article (54 page PDF) reminds me a lot of that slim book, detailing as it does many of the ways the presentation of data can be manipulative or misleading. Some of the examples are even the same as Huff's: the distorted axis, for example, or combining data. Even so, this would be a valuable addition to any course or program about data literacy or working with data. P.S. Sage Journals brands this as 'Available access'. I don't know what that means. It isn't 'open access'. So maybe download a copy while you can. Via Data Science Weekly, another of the newsletters I troll regularly fo content related to online learning.

What is a 'sneaked reference'? "Some unscrupulous actors have added extra references, invisible in the text but present in the articles' metadata, when they submitted the articles to scientific databases." Why is this a problem? "Citation counts heavily influence research funding, academic promotions and institutional rankings." The article calls this a new form of academic fraud, which I think is a bit much, but it's certainly unethical. Personally, I don't think much of citation counts; I have often seen an idea referenced in one paper that gets a few cites and then restated in another paper, usually by a more prominent researcher, that gets a much higher citation count. But it might take AI to track 'idea origin' and attribute real, not merely counted, influence. Via Scott Leslie.

Just for fun, this is an AI app that will take an image (in this case, me) and a sample video and create a video of the image mimicking the sample video. Watch mine here. I can't imagine it will be available for very long. It was created using an service called Gradio, which will help you create fun apps like this. Here's the GitHub. I found it on the AI Tidbits Weekly Roundup. You should visit this at least one; this is a list of the new models, services and applications released this week, which shows the remarkable activity in the space. I follow it every week to keep an eye out for anything relevant to online learning.

We've all heard those stories about the government web app that crashes as soon as it's launched. Corporate and institutional web apps can be as bad. The reason is that serving a billion requests is very different from serving a few thousand. This article describes the process of building a billion-request website that did not crash. That doesn't mean the process was flawless - they came to regret using React (for these reasons) after a few years. Using gRPC (a request protocol) "was not as great for us as I'd hoped."

When I read a title like "How to do P" my first question is whether, after reading the article, I can "do P". In this case, the answer is no. It's not a bad article, but it's badly titled. It should perhaps better be called "things to consider when building an educational app". In that there is some good advice, like, "The difficult part is not the technical building, but the interaction design," said Devan Walton... "Being able to pick thoughtful design hypotheses that you can test and iterate is much more valuable than dumping all your money hiring developers." That's true, of course, but you also need to know how to actually write the application code, which unless you've done it is generally what stops most people.

This is exactly why the content on my site uses a Creative Commons Non-Commercial license rather than CC-by. "Many indie hackers that contact us about our experience of running an open source startup tell us that their main concern about open sourcing their code is the risk that large corporations will resell that code and take advantage of their project." This article explains why the developers of Plausible, an open source website analytics tool, are changing their business model. Via Ben Werdmuller, who as the veteran of similar experiences is sympathetic.

Ben Williamson explains his "metaphorical labouring of the 'oblong' as a model of education" as a way to illustrate how "despite the rhetoric of transformation, all these AI tutors really seem to promise is a one-to-one transactional model of learning where the student interacts with a device." At times he seem to suggest that it is this pedagogy that's the problem, but the main issue seems to be his belief that AIs simply won't be successful: "AI tutors are simplified models of the very complex, situated work of pedagogy. We shouldn't expect so much from oblongs." I think he's drawing too much from the metaphor, from corporate rhetoric, and from Apple's widely criticized 'crush' advertisement, and not enough from what AI is actually doing in education.

This survey sounds like that famous restaurant review: "The food is awful! And there's not enough of it." It reflects, as Bryan Alexander suggests, "the varied purposes Americans think higher education should fulfill." One side complains that higher education is indoctrination, while the other focuses on value for money and cost. I think that so long as higher education - public or otherwise - is only serving an elite, public suspicion and doubt will continue to escalate. Opening up is at the same time the best way to counter accusations of bias and the best way to improve utility and access.

This (17 page PDF) is Mary Meeker's first public report in four years, and it is a pale shadow of the comprehensive coverage she has offered in the past. The focus is mostly on the opportunity AI brings to education. " bringing AI to learning and teaching requires what Sal Khan calls 'educated bravery'," she writes. "AI creates once-in-a-lifetime opportunities for evolution, creativity and leadership... Our universities and regulators have a responsibility to rapidly and deeply understand the global stakes that AI presents for freedom, democratic values, good and evil…and take strong stands." All this is true, but I really miss the depth and subtlety of the old Mary Meeker.

There's a good lesson here. "Ask almost anyone how Australian students are going in tests of basic skills, and the perception is that results are getting progressively worse," Dr Larsen said. "In reality, student achievement is only declining in the PISA assessments." As is well known, Pisa creates its own standards for assessment, and doesn't assess what national curricula are actually teaching. This makes Pisa more than just a testing program - like most evaluations of this sort, it's actually a form of advocacy for a certain set of learning outcomes. But you wouldn't know this unless you compared Pisa results with other forms of testing, which is what this article does.

Here's the question: "But there's one misgiving that I share with a surprisingly large number of other software developers. Does the use of generative AI increase the gap between entry-level junior developers and senior developers?" Based on my own experiences with AI, I would say "no". Instead, what it does is this: it raises the floor of what we would expect a junior programmer can do. It's similar to giving a junior programmer a higher level language like C or Lisp instead of having them write in Assembly or Machine Language. They're both equally hard, but you can get a lot more done with the higher level language. Wrangling AI to get it to produce decent code is similarly difficult, but it will use all those obscure functions or libraries you won't have learned as a junior.

I have been a non-academic philosopher for the last 30 years - indeed, I don't think academics would even recognize me as a philosopher ay more. I didn't get the terminal degree and I mostly eschewed academic publications. This article tells a different story about a different philosopher, but a lot of Pranay Sanklecha's reflections echo my own. And much of those reflections apply to all branches of academia, not just philosophy. "Let us go back to the world, to the modern equivalents of the Greek agora, let us do philosophy in places and with people where we are not protected – and mummified – by the sophisticated conventions and intricate rules of the institution of academic philosophy."

This is a great article where, after 20 years working in instructional design, Christy Tucker reflects on her first ID job, career journey, and what's changed. The job search part is a great story - after accumulating 200 rows in her job attribute spreadsheet, she attributes getting an interview to "luck". What's changed? The tool, she says, are a lot easier to master today than at the start of her career.

"Media should be a force for social connection, a convener of people across differences and a facilitator for what to do after the facts are laid bare." As usual, we can easily substitute 'education' for media. The current state is the same for both: "Trust in journalism is at an all-time low, engagement is declining, and the business outlook for the industry is uncertain at best." This article looks at why we have journalism at all and how we should redefine its mission. "As journalists, we should care about the people we report for, right? Otherwise why bother?" Via Michelle Manafy.

Cloudflare is a service that acts as a proxy between a user request for a website and the server that delivers the content. It was originally designed to shorten the distance content has to travel globally by creating a network of caches worldwide. The same technology is also used to protect websites from spammers and distributed denial-of-service (DDoS) attacks. Now it's offering to protect websites from AI bots. "This feature will automatically be updated over time as we see new fingerprints of offending bots we identify as widely scraping the web for model training." The article also offers an analysis of bot activity (pictured).