news (external)

The argument: AI is wrong because scraping is wrong. Scraping is wrong because it violates privacy. The article (64 page PDF) offers a definition of scraping ("'data scraping' refers to any time 'a computer program extracts data from output generated from another program.'") and proposes a framework based on bedrock principles known as the "Fair Information Practice Principles" (FIPPs) generally accepted by information privacy law. So, I have questions. Like, by this definition, isn't a web browser a scraping engine? When I follow a blog or newsletter, is this a form of surveillance? Should I be required to notify people if I read their posts and stop if I didn't? Are bookmarks now illegal? Can I no longer store summaries of the content I actually read in a database? You can't just say "it's bad because we use tools."

Tony Bates responds to a proposal to create a new super European Open University."that is 'student-centred, inclusive, digital and green' has been supported by the European Commission with a grant of 14.4 million euros (nearly C$21 million)" with four pointed questions related to the theme, "why?" Will the university support credit transfer? help members meet targets? in crease access? or even be governed in some way?

"One small step for pug, one giant leap for pug kind." So says this fun article on how Ghost - the small open source blog and newsletter tool - is adopting ActivityPub, the protocol supporting Mastodon and the rest of the fediverse. They've open-sourced their ActivityPub module (it's developed separately from Ghost's code, which suggests that maybe it could be used by other tools as well) and you can now subscribe to their newsletter on your favourite fediverse tool.

Good article describing how neuroscience is looking to artificial intelligence to solve some longstanding issues in the field. They're doing it because they can: "They trained a battery of different language models on those same sentences, and created a mapping model between human and machine neural activity. And they discovered that the networks not only produced humanlike text, but did so in a broadly humanlike way." This, in turn, tells us some things about artificial intelligence. "When they hallucinate information, it is not their failing, but ours: We are forcing them to answer a query that is outside their narrow competence... One of the things we've really learned from the last 20 years of cognitive neuroscience is that language and thought are separate in the brain."

This article (15 page PDF) is an introduction to open educational resources, covering "the OER movement, its milestones, and its integration into educational practice" as well as arguments for OER, criticisms, and an examination of "current OER implementation in higher education and its promise of innovation." Via Alan Levine.

I really do love 'do it yourself' posts and have an entire YouTube playlist devoted to my adventures with them. This set of instructions, however, begins with the requirement for "one cluster that had 4,088 H100 GPUs spread across 511 computers, with eight GPUs to a computer." Um, that's a bit much for my home office, let alone my budget. Then you have to make sure every computer works (they don't always), set up the software, train a single node, "burn InfiniBand burn," make sure again that the machines are working, and more. I like the 'reflections and learnings' at the end and look forward to the day when home hobbiests can set up their own 70B models in their living rooms.

This is a really good article discussing where Neal Stephenson's concept of The Young Lady's Illustrated Primer works and, more importantly, where it fails. "It works its magic through manipulation," writes Andy Matuschuk, "through isolation from people or projects with real meaning, and through a misguided reliance on aesthetic pleasures. I see no way to incrementally rescue its structure from these flaws." To a large degree, I agree with his argument. A personal learning environment should be engaged with reality, should offer control to the learner, and should serve as a multi-purpose thinking and learning tool, "an ever-present conduit, providing the support, structure, and representations we need to do things we care about."

"The answer to the problem or challenge of generative AI in education is generative teaching," writes Steve Hargadon. What does that mean? "How can we help students understand and use these amazing new tools in a way that lights the fires of their intellectual curiosity and growth, rather than just filling the pails through of traditional instruction and assessment?" I think this is the right approach - it is often better to run toward a challenge than to retreat from it. But as Hargadon notes, "t we have to become capable enough ourselves with an understanding AI in order to manage the process." This is a longer post that frames both the question and answer; worth a read.

Though this is not immediately useful it's the sort of approach necessary to accomplish a genuinely decentralized internet. "The web democratized the exchange of information, but it's missing a key layer: identity. We struggle to secure personal data with hundreds of accounts and passwords we can't remember. On the web today, identity and personal data have become the property of third parties. Web5 brings decentralized identity and data storage to your applications. It lets devs focus on creating delightful user experiences, while returning ownership of data and identity to individuals." It's based on Node, which I think is far too fragile an environment for critical web applications. But it's a step in the right direction. Here's the blog, which I'll be watching.

This contribution (21 page PDF) consists of five short essays. In his introduction Peter Niesen frames the issue: the question is whether diversity confers an epistemic advantage - that is, does a diverse group make better decisions than a non-diverse group of experts? Kai Spiekermann looks at the mathematical arguments for and against (the theory, presented as the "diversity trumps ability" (DTA) theorem by Hong and Page in 2004, was first presented as a mathematical theory). Lisa Herzog points to the difficulties in assessing such a result: who counts as an expert? Does diversity mean participants are drawn from the general public? Etc. Charles Girard looks at how the meaning of 'epsitemically good' varies in contexts like search, deliberation, and voting. And Frieder Vogelman offers a 'negativist appraisal': "histories of actual epistemic communities illustrate the epistemic worth of diversity by showing the epistemically problematic consequences of homogeneous scientific communities." Readers will note that over the years I have also argued for the epistemic value of diversity (as a part of the 'semantic condition') and offered a mathematical proof in co-authorship with Daniel Lemire and Seb Paquet.

Open source results in better AI, for a variety of reasons - transparency, robustness, portability. So there's no reason to expect that artificial intelligence will be dominated by proprietary technology owned by large corporations, argues Ines Montani in this talk. Unless, that is, regulatory measures favour monopolistic practices, which is always a concern. Daniel Dominguez summarizes the talk in this article.

This article quotes Paul Thurman as suggesting that AI replace managers. Maybe not a bad idea? It even has the support of managers (presuming they have not lost their jobs): "Mr Rauma says that the shift towards an AI manager has not only reduced his stress levels, but has enabled his employees to work faster and be more productive. "I'm able to focus on the growth of the company and all the positive things. It's added years to my life, I'm sure," he says."

This article offers a review of a recent government report on the role of education in society "through the theoretical lens of Michel Foucault as contextualised by the scholarship of Farrell" which, says Heather Marshall, "reveals a nuanced understanding of the power dynamics inherent within educational institutions." The report argues that "the current winds of extremism, polarisation and democratic disruption combined with social and economic issues may cause even more unrest" so "how we build and deliver social cohesion must be overhauled." I'm not sure 'cohesion' is the right word, but I digress. The Foucault-based analysis reveals "the potential for educational policies not only to convey knowledge but also to serve as mechanisms of power that shape individuals' understanding of themselves, their values and their place within society." In other words, education works.

I have to pass this along, even though I'm rolling my eyes as I do it. Here's the study (this is the real title): Elevated ceiling heights reduce the cognitive performance of higher-education students during exams (7 page PDF). If you read the paper, you see the authors acknowledge that it could be because the large rooms are cold, crowded, strange, and harder to cheat in. To me, all this raises the question once again: what is it exactly that we are measuring when we assign and grade exams? Each study like this shows that it's less and less likely to be what they actually know.

Some recent work on using AI to teach math. This paper from the Learning Agency focuses on identifying a taxonomy of math errors and emphasizing "a strategic focus on the thoughtful and intentional design of assessments that provide accurate diagnostic assessments of student errors." In related work, Khan Academy provides a dataset and explanation. "This paper specifically addresses the accuracy of LLMs in performing math correctly while tutoring students." Some authors from Stanford University also wedge their way into the conversation, offering "a dataset of real-world tutoring conversations, annotated with expert decisions and responses." Via forums here and here.

I certainly agree with Scott Leslie's recommendation that this keynote address is worth a read. I gave it a nice slow thorough read while watching the baseball game after making a cycling video. Let me linger on three points:

• "for Yunkaporta everything is about relationships and flows..." This is presented as traditional knowledge, but it's basically core knowledge for me too.
• "classification, cataloguing, and ontological mapping... flattens reality into a list of predetermined categories and definitions."
• "spam... exploit(s) existing aggregations of human attention... So every commercial website is mostly spam."

My response: Knowledge is about interactions and flows, but language distorts knowledge by flattening distinctions and creating focal points that attract power and influence. Artificial Intelligence (as we know it today) is based on language - they are literally 'large language models'. That's why is appears to reify colonial inequalities, when in fact if designed correctly could do the opposite, by transcending the limitations of language.

This is how Firefox started, back before Firefox depended on Google for revenue and before Mozilla was buying digital advertising companies. Sure, maybe it'll be nothing. But maybe, it won't. "Unlike traditional business models that rely on monetizing the user, Ladybird is funded entirely by sponsorships and donations from companies and individuals who care about the open web."

It's funny because it's true. Via Mignon Fogarty. "Its predecessors already produce hundreds or even thousands of words almost instantaneously. Now GPT-5 brings PhD writing skills to the table, meaning it can generate text at a rate of about ten words per day.

Benjamin Riley posted this article a couple months ago and referenced it in a discussion today on the nature of intelligence. The focus here is on higher order skills and the role they play in intelligence (indeed, what we describe as 'intellegence' often referes to these skills directly). The main argument here is that while large language models (LLM) may surpass humans in specific domains, they cannot transfer that learning into other domains. In fact, there is a domain in AI called transfer learning, which I don't address in my reply to Riley, but I could.

According to Creative Commons, "through engagement with a wide variety of stakeholders, we heard frustrations with the 'all or nothing' choices they seemed to face with copyright...  way of making requests about some uses, not enforceable through the licenses, but an indication of the creators' wishes." In particular, they want to be able to limit the use of their work to train AI. I commented in a meeting today that it was telling that this, of all possible preferences, is the one that surfaced as most significant. I would rank 'use of content to create weapons' or 'use of content to undermine social good' as more significant preferences. I also commented that, without access to open content, AI will be created and trained exclusively by commercial entities with licnesing agreements, which will mean there is no possibility of an open artificial intelligence.