Stephen Downes

There has been a lot of discussion about Anthropic's proposed 'watermarking' of AI-generated content. This article explains how it works, which in turn informs our intuitions about why it's so objectionable. Basically, what the system does is to subtly alter word selection when generating text, creating a difference in the probability of what it outputs as compared to what it would ideally output. This difference is essentially a 'signature' attributing the content to a certain AI. Anthropic will later offer an API which allows people to test for thaat signature. The problem is, if you use AI, the AI is messing with your text to embed hidden messages in that text, which is (to say the least) problematic. John Gruber calls it patently offensive. Tim Moon and others calls it a digital scarlet letter.

This is a clever tool that helps the reader explore deeper and deeper into a topic. Pick something to explore and it surfaces a bunch of related resources: "surprising associations, facts, links and short videos around anything." You can read the resources or pick one of the new associated concepts to dig even deeper (click on the title, then the 'dig deeper button' - the interface isn't intuitive). I guess my only criticism is that you tend to get more and more general as you dig; my search on B200, for example, led to 'neural network' and 'matrix multiplication' (it's like the old saw where when you followed the links in Wikipedia, you always ended up at Philosophy). Via Nele Hirsch.

This is "a live demonstration of everything a website learns about you before you click anything. No cookies, no permissions, no consent dialog. Nothing is stored." And yet, it can create a precise 'fingerprint' of your browser and computing environment detailed enough to create a profile that can be reused over time. It's a nice, effective demonstration that makes a compelling point. Via Grant Potter.

We've all experienced the many irritants that tech has brought into our everyday lives - everything from panes that cover up the content (aka 'dickover'), one of which you'll experience on this article as you go to read it, to anoying email subscriptions following product purchases ('limpitlisting'), self-checkout travails ('baggravation'), parking lots that require you to download an app ('parkware'), and products that require fees in order to be used ('hostageware'), among many more. Someone should write an article with an equivalent list for edtech.

Readers by now know how much I like stories like this. Here, three schoolgirls recorded a video 'news report' (using an umbrella as a microphone) to document the potholes in the road they use to walk to school. You can see the video in this report on OneIndia. According to this article authorities saw the report, which went viral, and repaired the road. OneIndia reports, though, that the filling of potholes after monsoons has become an annual ritual. That doesn't detract at all though from what was the students' best educational experience of the week.

This is a good interview with Mozilla's CTO Raffi Krikorian. Mozilla is the company that distributes the open Firefox browser, which is what I use daily. He "explains why businesses are turning to open-source and open-weight AI models for greater control over data, costs, and customization." I think there's going to be a lag in open weight models (much lss fully open source models) because the cost of running them is still too high. The Canadian Cohere model called Command A Plus requires at least one B200 GPU set, which costs $6 per hour to run (or $45K to buy). Now I remember when Compuserve cost $6 per hour to access, back in the 1980s. So we can imagine a world in which this is attainable for most people. But it won't be soon. Meanwhile, though, it makes sense to focus on the creation of the open source infrastructure we need to make sure that when the time comes, it actually is available to all. Image: Nvidia. See also: What's going on with open weight models?