Stephen Downes

I've kept this item in a browser tab for a couple of weeks waiting to give it the proper attention it deserves, and while I never did, I still want to pass it along. There's a whole debate under the surface of the AI revolution on the relevance of knowledge structures from good old fashioned AI (GOFAI) such as ontologies, graphs and namespaces. The suggestion in this article is that new AI (that we see in large language models, for example) will not escape four issues that dogged (and still dog) these activities: open world (multiple 'truths') vs closed world (single 'truth') models; reification (statements about assertions); context graphs and the primacy of time; and the relation between 'authority' and namespaces ("there is no neutral graph"). Many of the issues people have with large language models can be traced directly to these four issues: questions of bias, authority, the meanings of words, and the passage of time create reasons to doubt generative AI, and AI in general. Well, and human knowledge too. Anyhow, this is a great article. Do spend some time with it.

I haven't added xAPI to CList yet, but I do plan to when I get back to it in the fall. This guide will help. The focus is on xAPI extensions. An xAPI statement "may describe who did what, to which activity, and with what result, but many real-world use cases need additional information, such as environmental conditions, system state, scoring details (etc)." Hence, the need to add an extension to include the additonal information. (My favourite statement from the whole document: "There are only ~150 prepositions in English.")

As HESA comments when introducing this item, the same concern applies in Canadian research finding as well. " the excellence trap: the effect that arises when the majority of research is funded through a project-based model in which many projects require co-funding. Such requirements mean that the more successful a university is in securing competitive funding, the less room it has to invest in the long-term development of excellence." Martin Nilsson Jacobi argues, "society is not best served by allocating resources in ways that are too narrowly targeted and too short-term. Academic freedom is not only a fundamental principle underpinning democratic values; it is also the most effective means of maximising the value that academia delivers to society." Governments are able to enter into longer term contracts and infrastructure agreements with companies; I see no reason they can't do so with universities and consortia.

I think it's really important for people who lead with 'theory' and see it as a 'lens' to rethink their understanding of the scientific method in the era of data science. "In data science, you don't start with a hypothesis or prediction," Weber said. "You start with the data that already exists - maybe numbers someone collected years ago, or information gathered for a totally different purpose - and you work backward. You look for patterns, connections or surprises in the data, and those clues help you figure out what questions you should even be asking. So, instead of testing a hypothesis, you're discovering one." This article is based on a paywalled paper by Eric Weber, et al., though there's an archive version available (nor now) here.

There's some nuance here that the headline doesn't really capture. Google was found liable by a judge because its AI-generated search output produced some false statements about the publishers it referenced. So the core argument here is that the person who publishes the statement is liable, even if it was an AI that generated it. As it should be. The nuance is found in the fact that the search engine would not be held liable for false statements produced by other sites that showed up in the search. So if the NY Post published a lie, and it showed it in a search result, Google wasn't liable. Again, as it should be. Google tried to use this as a defense, saying the AI output was produced in the context of producing search results. Which would be a defense if you needed AI-generated text to produce search results. But you don't. Hence, the headline.

Grant Potter pulls out the relevant quote from this article (which should be read in its entirety): "Here's a good rule of thumb: if you go online - especially if a social media platform is involved - assume there is an attempt to exploit your fears, grievances, beliefs, personal identity, and sense of community. Whether you are an elitist music nerd or an aggrieved Alberta separatist, assume you are being manipulated." The 'hottest band of the year' was a band named Geese, but "their stratospheric hipness was all due to algorithmically manufactured buzz. Fake fans. Fake comments. Fake reviews. Bots pushing social media posts." That's what buys fame these days (in our field too). I'm proud to say I had never heard of them before I read this article.