Commentary on Introducing Nano Banana Pro by Stephen Downes. Online learning, e-learning, new media, connectivism, MOOCs, personal learning environments, new literacy, and more

A prediction about Florida's coastline is wrongly portrayed by people engage in climate change denial -- but AI Mode does mess up the summary in a way that r...

Once I thought more deeply about what people had been asking for it made a lot of sense

"I had one friend who told a colleague that he was going across campus to an Al workshop, and the other professors said, 'Don't, we're leading a boycott against the workshop.' Okay. I mean, I don't remember that kind of thing happening with Wikipedia or other tools for online learning..." For me at least, it's pretty simple. People are using these tools, and they are using them poorly. We are educators and if we can teach them to use them more effectively we should. If we refuse to do that, where we end up as a society is at least a little bit on us. But I disagree with Bryan a bit. We went through this before in miniature. In 2010 I was trying to convince people in civic education conferences we should teach people to use social media more effectively, including checking things online. The most common response "We shouldn't be teaching social media, we should be telling students to subscribe to physical newspapers instead." Those students we could have taught that year are thirty-five now. We could have had 15 cohorts of college students knowing how to check the truth of what they see online. Our entire history might be different, and maybe we wouldn't be seeing this rampant conspiracism. The thing is those professors who said we should just give students physical papers will never realize their role in getting us here. I wish others would consider that history before they treat boycotts of AI workshops like a noble act. When you engage in politics you are judged by results, not intentions. And the results of this approach are not risk free.

We reason about a a study showing chocolate improves memory.

Is consumer sentiment really near an all time low? And what does that mean? We get to use (at the very end) our definitional follow-up, which reveals an asto...

A while back I released my "get it in, track it down, follow up" framework for teaching students to use AI to assist with critical thinking. The framework is meant to address three major issues: * Help students mitigate confirmation bias and sycophancy, * Make sure that their answers are grounded in appropriate sources outside the LLM, * and turn LLM sessions into interactive critical thinking exercises that not only mitigate the harms of cognitive off-loading, but scaffold their critical thinking development. “Get it in” reflects two principles. First, just make the first step — as I say, the most important part of a gym routine is walking into the gym. You want to make it as fluid as possible to start. But the second part deals with sycophancy and confirmation bias. I’ve found in general that a practice of just putting the claim in, either bare or with a dry “analyze this claim” is a good way to avoid the pitfalls of inadvertently signal you want it to take your side. Track it down reflects my observation that when we use AI for information-seeking it is best conceptualized as a “portal, not a portrait”. LLMs don’t return answers, exactly. They return knowledge maps, representations of discourse. For anything with stakes, you are going to want to ground your knowledge outside the LLM. You need to follow the links, you need to check the summaries. I sometimes use the metaphor of those little mapping drones in science fiction that fly into a ship or set of caves and produce a detailed map before Sigourney Weaver (if you’re lucky) or Vin Diesel (if you’re not) goes in. Like that little drone (which I guess is science fact, now, isn’t it) a search-assisted LLM goes out and maps the discourse space, providing a representation of what people are saying (or would tend to say) about certain subjects. It’s a map of the discourse “out there”. But it’s still just a map. You’ve ultimately got to take it in hand and venture out, click the links, check the summaries, see if the map matches the reality. You’ve got to get to real sources, written by real people. Track it down! The final element, follow up, captures at the highest level that you have to steer the LLM as a tool or craft. Many people don’t like the idea of of LLMs as “partners”, being too anthropomorphic. Fine. This undersells it, but sometimes I think of them as “Excel for critical thinking”. What do I mean by that? Just as if you know the right formulas in Excel (and understand them) you can model out different scenarios and shape presentation outputs, with LLMs you can use follow-ups to try different approaches to the information environment. This can all seem very abstract, which is why I've created over 25 videos showing me walking through example information-seeking problems and showing how these "moves" are applied. Check out the link in the comments for links to the videos, and more explanation.

Mostly a cautionary tale about trusting AI Overview with any photo based investigation.

In this video we show, for what feels like the hundredth time, that AI Overview is a product that should be taken off the market but that AI Mode is fairly r...

We spot a new Google Lens "questions about this content" feature in the wild and muck around a bit with it, unsure how much it really improves things.

I investigate using the foundational question "Is this what people think it is?" as a follow-up -- and get strikingly good results. For people needing quick,...

I usually encourage people to jump straight to AI Mode, but you don't have to be an absolute robot about it. In this example we get such a great set of sourc...

I see some instructors on here joking how they are going to add prompt injections into assignments as a defense against agentive browsers. I get the frustration and maybe it's just jokes? Maybe I need to lighten up? But just in case: prompt injection applied to people to whom you have a duty of care is not funny, it's not resistance, it's deeply messed up behavior that involves using your power to hijack the computer of people you force to consume your compromised materials, and if that seems reasonable to you, you need to touch some grass.

I show a bit of how I make these examples, and then we get to the real question, which is "What did we do to upset Mister Rogers so much?"As usual, everythin...

A video exploring the etymology of "hoe down" (while debunking a folk etymology). A bit of "track it down" in it, as usual. And a good result!As usual, every...

Do we have two moons? We explore that question while using a follow-up about definitions and measurement and singing badly.As usual, everything I say here is...

Another video primarily about tracking it down. I use both follow-ups to request better links and use the little citation links to get me there. I eventually...

A simple one today, tracing down a quote by Emma Goldman. Since it turns out not to be hard we mostly use it to show how to use the little quote icons in the...

Sometimes the best thing an LLM can do for you is misunderstand you

I've been holding back on sharing this until I could present this more fully, but let's YOLO it.

A little venting about what educational explanations are for...

A very minimal test with some fairly clear results. (The answer is yes, it's bad)

Many "errors" in search-assisted LLMs are not errors at all, but the result of an investigation aborted too soon. Here's how to up your LLM-based verification game by going to round two.

Just as there is "search" and there is "search process" there is a process around AI-assisted investigation too -- and we should teach it

Just as there is "search" and there is "search process" there is process for AI investigation too -- and we should teach it

Search might be a better approach for learning about the world, but the right comparisons matter

Anna Mills prompting session at MYFest

An abbreviated version of SIFT Toolbox is now available to all as a completely free GPT