Stephen Downes

The part of this post that caught my attention wasthe detailed explanation following the phrase "reading texts is a social practice in that it involves, inter alia, imagining the author and other readers." That, writes Martin Lenz, points to a significant difference between how we read an text and how a machine does. The comouter is not imagining what the author was thinking or lived through, and is not considering what other readers might make of the text. All this, writes Lenz, happens in "the process of decoding and comprehension." I'm not a fan of this phrasing because I don't really see language as (shall we say) 'encoded meaning'. It evokes thoughts and ideas; it doesn't reprsent them. That's how it can make sense to say that reading involves imaging what other people are thinking.

Good overview of what happens when development using AI agents becomes complex. "How is context shared between teams? What's the token bill? How do we deal with increased review demands? What's the agent allowed to reach, read and change? And who's now responsible for the harness everyone depends on?" these aren't technical questions, at heart. They're questions about how the people work together, Chris Ford argues. "I urge you to remember that the system we're trying to build here is primarily a human one. Yes, the technology is sophisticated, but its implementation requires us to properly consider who owns it, who reviews it, who pays for it and who's accountable when it fails."

I think one of the biggest differences between LLMs and humans is memory. As this article notes, "large language models (LLMs) are stateless: they don't remember anything between conversations." So they need "a system outside the model that writes things down and hands them back later." As a result, "AI researchers borrowed four categories from cognitive science to describe what agents store (the CoALA framework is the usual reference)." But (in my opinion, at least) that's not how human memory works. In human memory, the models and weights are never frozen. The model is the memory. It takes so much energy and computing power today to train models than ongoing training isn't practical, but in a human, that function never turns off. There's no separate read-write system in the brain, only the one out there in the world, consisting of PostIt notes and textbooks and alarm clocks. 

Networks in nature run by themselves so I'm always wary of attempts to impose order on networks, because usually that leads to centralization and, eventually, destabilization (and ultimately, in human spheres, privatization, then commercialization, then enshittification). There is a bit of an awareness of this in this article. For example, "One shares, 'our governance is decentralized, not uniform in many ways' and 'there are processes that enable us to be accountable to each other based on our strategy and plans.'" Still, networks don't build themselves (except in nature): "Stewardship means creating concrete tools and guides, as well as explicit permission and role-modeling." But a lot of this is necessary only is network use is instrumental, "for members to harness the network towards their individual and collective goals." And instrumentalism is the beginning of the end for a network.

You definitely want to take the time to watch the short video and read the article, especially if you're an AI sceptic. Ethan Mollick offers a number of descriptions of things the AI did (Claude Fable 5.2, to be specific). The cost of using Fable for more than a few minutes at a time is beyond my budget, so I couldn't test it on any of the things described here. But assuming Mollick's honesty (which I think is fair) this is a convincing demonstration of how capable AI has become. "I think this phase of Jagged AGI is an exciting one, even as I wonder, a bit anxiously, how long it will last," he writes.

AI agents were set loose in SimCity (an online virtual city simulation) and linguists studied their interactions. They tended to adopt their own slang and jargon (though less so if agents from different LLMs took part). "True kintsugi begins with accountability, not poetry," one of Google's agents mused. "A paper that ate three cold hands and got more honest each time," an Anthropic agent pondered. I wondered whether similar work had been done with children in SimCity, but when I looked all I found were uses directly related to learning outcomes. This, for example: "As game-based learning gains momentum in education circles, teachers increasingly want substantive proof that games are helpful for learning. " Via RestOfWorld.,