Stephen Downes

"The narrative of our field is very much a before-and-after Gen AI story." That's how this article starts, and how it's framed - though the interviewee, Reach Capital's Jennifer Carolan, is very firmly fixed on the traditional classroom environment. "I'd tell every edtech builder to become a substitute teacher. Shadow a teacher. Sit in a classroom for three days in a row—spoiler: it's boring and exhausting, and it isn't the teacher's fault, but the factory-like structure of it all." I don't think that ed tech ill succeed by solving the problems faces in the current classroom environment. Students - just like the startup founders being coached by Reach - should get out of the building. That's why the field isn't a before-and-after GenAI story. Not so long as the AI is intended to emulate a teacher

This is a chapter from a larger book, the whole of which is worth looking at. This summary doesn't do this complex chapter justice. The first paradox is power: "The central issue is no longer only access to scholarly outputs, but who controls the infrastructures through which scholarly communication is organised." This is why I have historically argued for decentralized open access, rather than platform centrlization. The second paradox is 'reciprocity', where proponents of openness have opposed text and data mining (TDM) because " the same openness that dismantled subscription barriers also created conditions under which scholarly content could be recomposed as a scalable input for platform economies" (readers should remember my many years of advocacy for a non-commercial clause to prevent just this). The third is governance: "smaller data infrastructures now find themselves overwhelmed by automated scraping requests, forced to absorb the operational costs of large-scale harvesting while lacking the resources to govern, limit, or benefit from such use." This is why (and my colleagues from EDUCAUSE three decades ago may remember) why I argues for distributed resources and aggregation, rather than federation.

It's as though people believe there's a learning 'off switch' (though it only exists in other people). Like this: "A junior who made a mistake is one step closer to being a senior; a junior who let an LLM make a mistake (and had the LLM fix it for them) has probably learned nothing." What would justify this conclusion, authored by John Collinsworth (Nielsen doesn't link to it, though he quotes it extensively - boo, hiss)? The junior will still learn, but will learn a different thing. Nobody 'learns nothing' - human brains don't shut off like light bulbs. What's really happening here is that we're mking a value judgement, specifically, that the lesson learned from doing it by hand is more important than the lesson learned by doing it with AI. This 

The first think I thought of when I read this post from Julian Stodd was simulated annealing and Boltzmann machines. Why? Because Stodd describes crystals as structures that "represent the lowest energy configuration," which is what Boltzmann machines try to do with neural nets. Stodd compares crystals to organizations, and here the metaphor needs rescuing a bit. He writes, organizations "build structure for greatest efficiency, value, predictability, replicability, and of course, ease of control, and potential for oversight and measurement," but that they are "systems trapped at one energy level, dreaming of the next." Sure, they can change state - they can melt into fluid and merge with something else, or they can sublimate into their individual atomic components (a.k.a. people). But they can change energy level, but (just as with matter and neural nets) they need to go through an annealing process, a 'hardening through fire', as it were. 'Move fast and break things' is a clumsy attempt at just such a process; in neural networks they just increase and decrease the bias (ie., sensitivity, not prejudice (that's a different meaning of 'bias')) of individual entities.

What's great about this article is that it really clearly defines the distinction between Mastodon and Bluesky, arguing that the question "where are the instances?" is a Mastodon question, and doesn't really apply to Bluesky. But I think it's misleading in three major ways. First, there are instances in Bluesky, but there are different types of them: 'atproto hosting' and 'apps'. The fact that there are so few 'atproto hosting' instance is an issue. Second, it under-describes the ATmosphere architecture by leaving out the 'personal data store' (PDS) entirely. And third, it argues that atproto is just like RSS and Google Reader. But RSS is a simple text file any person can create in a text editor (so, for that matter, is a PDS, sort of), while a 'atproto hosting' instance is a very large and complex piece of software that consumes massive resources. So too are the apps, which for some reason, must handle "the whole Atmosphere". And the whole argument raises the question: why can't we simply have PDS readers as apps? What is the whole 'atproto hosting' infrastructure buy us, except control over our identity?

This is a longish discussion of McLuhan's idea of a tool as an extension of the self, which then of course can be applied to media. Today, it can be applied to AI. So when we say 'AI is a tool' we're not limiting our imagination - "the ceiling of what tool literacy can imagine" - we are instead inviting ourselves to explore how a human and an AI can be (dare I say it?) a system. Now I'm not sure Levine meant his discussion exactly that way, but this is what follows from the text. And I question both McLuhan and that reading. I used to consider the question of whether a social network is an extension of the neural network (which is essentially the Siemens interpretation of connectivism). But I see them as two separate networks, with a process of communication (essentially, emergence and recognition) between them. The McLuhan argument is a nice metaphor. But it should not be taken literally.