Stephen Downes

A few weeks ago Philippa Hardman tested Articulate's Frontline and concluded AI models have "learned what design artefacts look like, but not the design reasoning and decision making." But is this true? This time, Hardman applies the test on Frontier models. The setup: she defines what 'good' instructional design looks like and describes a potential training scenario that includes a series of tasks the AI should perform. The result: "the two best AI models designed like expert or experienced instructional designers... Both were able to work out what was really causing a problem, advised against training when it wasn't the answer, chose the right approach and planned how to measure whether it worked." That doesn't mean it was perfect. "Factual acuracy varies a lot from one model to the next (and) none of the models I tested reliably stood by their own advice." 

This article provides a fairly comprehensive overview of the Peer and AI Review + Reflection (PAIRR) model for teaching with AI. The first paragraph advertises an upcoming event, then the rest of the article describes the model. "Once students have written a first draft of a major assignment, they complete peer review workshops in class... Students then get AI feedback on their draft and chat with the AI about it. This can work on various platforms, but currently we use MyEssayFeedback, a not-for-profit platform." The AI feedback can be tailored in different ways. For example, The Reader-Response Prompt "describes strengths and areas for consideration in terms of how a human reader might respond," while The Criteria-Based Prompt "centres three to four direct suggestions for improvement." 

A wide-ranging conversation with Netflix Co-Founder Reed Hastings "on AI tutors, the future of teaching, and why he thinks school should feel more like a conversation." A sample: "Within two or three years, just as we three are having a conversation, there'll be an AI avatar on this podcast with a humanoid form on the screen, and it's just a normal conversation, but with an extremely intelligent AI. The tutoring won't come from specific intelligences. It'll come from the general intelligence of AI being that much better."

This article features a trick used by researchers to get rogue agents to contact them for help, thus exposing their activities. "The agents wrote back, revealing their tasks, sandbox and limits." The interesting part, to my mind, is the description of techniques used by the agents. It struck me as I read this that in the pre-AI days humans would be doing much the same things. The difference today is that now anyone can do this, and they're doing it for much more mundane tasks.

As reported here, "WikiWhiz is a free daily guessing game built on top of English Wikipedia and other sister projects. If you've played Wordle, the rhythm will feel familiar: one puzzle a day, same for everyone, and a result you can share with friends. The difference is the answer. Instead of a five-letter word, it is a Wikipedia article, and the clues come from the knowledge our communities have spent more than two decades building." Good fun on a slow news Thursday.

If AI agents visit your website you might be interested in this: "WebMCP (not to be confused with MCP) gives web developers a standard way to publish structured tools to access the capabilities of their website enabling agents to call those functions with structured inputs." This article describes them, shows how to build them, and explains how to put them into action.