Welcome to another episode of AI Goes to College, the podcast that helps higher ed professionals figure out just what in the world is going on with generative AI.

I am joined by my friend, colleague and co host, Dr. Robert E. Crossler from Washington State University.

Rob, you were just in Tennessee talking about AI.

Tell us about your trip.

One of our listeners invited me out to the University of Tennessee to be part of their college's conversation around really becoming purposeful about AI in the classroom and had a great trip.

It was fun to share with them some of my thinking about AI and how to become purposeful about working that into the curriculum and understanding what's going on.

And I got to hear about a lot of the cool things that they have going on within their ag college.

That was the group that brought me out and it was a super exciting trip.

Nice.

Well, good deal.

So we are always willing to entertain those kinds of visits.

So you can either email Rob Crossleri goes to college.com or craig@AI goes to college and we will be happy to talk to you.

So Rob, there were a couple of interesting things in the news over the last few weeks.

One that everybody in higher ed heard about and nothing made me happier to be on sabbatical than this news.

So there was a global, I think Canvas blackout and Canvas is a learning management system that basically is where we push all of our materials to a ton of people use it for testing and it was down during a lot of schools final exam periods.

So Rob, let's connect that to AI.

Were you affected?

WSU was affected me personally, it didn't affect me, but it was at the higher level of faculty going what do I do?

We become so reliant upon Canvas, I think anywhere that uses it that it's our central repository of everything we're doing in the classroom, from lecture notes to exams to grades.

Many people don't even keep grade books on their own computer anymore and trust the grades in Canvas to be available to do all those sorts of things.

At WSU we were in the grading process.

Our grades are due on the 12th of May at 5pm so we were past exam season.

But some people needed access to Canvas to even grade what students had submitted.

So there was definitely some high impact stress moments When I was in Tennessee is where I was at when this happened.

It was final season and there are final exams that had to be postponed.

Highly disruptive to students, even probably to faculty who have plans figured out for what things look like.

And if you move that exam day that takes precedence to deliver those sorts of things.

So it was definitely freak out mode for a lot of people.

Yeah, I saw something online about some school that had moved their exams to Saturday, which I'm sure created a lot of angst, although we had those when I was at Ohio University, so it's not necessarily a new thing.

About the same time I was working on something pretty closely with Claude and it just started doing all kinds of flaky things.

And I don't think these two events were connected in any way other than by general timing.

But I lost one of my skills.

I'm a big Claude skill user and I was in the middle of editing a skill when something went down.

They had a partial outage and it must have been trying to write the skill when that outage happened.

So it was just gone.

Fortunately, I had it backed up.

So here's public service announcement number one.

Back up your grades.

It's pretty easy in the LMS to just.

There's some sort of a grade export function.

I do that after every major graded event.

I won't do it after every little online two point assignment, but any test, any project, I immediately download it.

That way I've got it in case something goes wrong with the LMS.

But if you're using Claude skills or now our ChatGPT has skills for teams enterprise and I think Edu accounts back them up.

It was really no big deal.

I just had it, I re uploaded, it was fine.

But the bigger picture here is we're dependent on these things and as we start to get more dependent on AI, we're going to be more vulnerable to these little outages.

And that's the one thing that bothers me about Anthropic is they seem to have a lot more outages of some degree than OpenAI or Google do.

And maybe new deal with SpaceX will solve some of that.

But I mean it really kind of shut me down.

I couldn't work on what I was working on.

Yeah, I think that's an important thing to realize here as we enter into the AI spaces and so many of the things that we're doing have these single points of failures.

That's what made Canvas an interesting organization to go after, is they didn't go after an individual school which would have had a very small impact on a set group.

They went to a provider that was providing learning management Systems to approximately 8,800 schools.

And so it was high impact as a ransomware attack where they're trying to take the money to bring them back up or however that played out.

I think we're still going to learn more over the course of the next week or two of what really happened and what was really taken.

But if you think about how we're setting up our AI solutions, we're doing the exact same thing, is we are going to a single place where we're getting access to that.

And as more and more of our workflows begin to rely on AI tools and what AI can do, these disruptions truly will take down a lot of workers ability to create and to work.

There's a lot of directions those conversations can go.

But I think one of the things that we need to be thinking about from higher education, from an administrative perspective, is how do we ensure that we can still do our jobs?

How are we resilient to being able to do our jobs in a world where more and more of these single points of failures might exist?

That.

Yeah, yeah.

And just for those of you who may not know, it was a ransomware attack on Canvas, I don't think that's what shut it down.

My guess is that they shut it down to put some patches and fixes in place or to assess the scale of the breach.

Well, and Craig, the number one response to a ransomware attack is to unplug yourself from the Internet.

So that way you can try to get the bad guys out of your machine functionally.

That's what that did.

But the other thing Craig, I think is important to think about with AI and where we're at is I like to think of what's going on with AI tools and AI solutions as the wild wild West.

I talk to a lot of people, whether they're students or whether they're faculty staff, creating agents, creating all these things that you talk about.

And many people are not trained software engineers.

They rely on vibe coding, they rely on prompting to get these things created.

And you hope they're secure, you hope they're protected.

But it wouldn't surprise me if we start seeing these one off agents, all these one off places, as being sources of hackers, or the bad guys, as I like to refer to them as finding a way to get in and then from there, what else goes on.

So I think from a security governance perspective, that 2026 could be an interesting year of crazy stories of more things like this happening, even if it's just affecting individuals.

Absolutely.

So you just gave a little generational test right there.

What is that?

So when you said wild wild west, so you probably thought of the Will Smith movie and had the theme for that movie running through your head.

I talked about the old black started out as a black and white TV show and there's going to be some segment of the listener audience that has no idea what we're talking about with any of that.

So sneaking in that little generational test there.

Well, and to give you my perspective, by the way.

Yeah, to give you my perspective, it probably wasn't even that.

I did my elementary school years in Montana and people like Billy the Kid and Wild Bill Hickok were some of my favorite characters growing up.

Nice.

Nice.

So, yeah, this is one reason I'm so interested in local large language models, because you retain a lot more control.

And I think the open source models are starting to get good enough to where they can do a lot of things that we want to have done.

So I know something to keep an eye on.

So here's the second news item and this one was pretty under the radar.

So I want to get your reaction to this and I'm going to try to summarize a lot here.

So Arizona State University, who is known for several things, one being really, really, really, really big, I think it's probably the largest public university in the country, being very innovative.

Michael Crow, their president, is very willing to push the envelope on things and this is the part that I kind of gleaned from my time in Arizona.

He also doesn't care much what faculty think about things, so he's going to do what he thinks is best for the institution.

And faculty can kick and scream all they want, he's going to do it anyway.

And that was reflected in their latest little bit with AI.

So they started up ASU Atomic, which uses something called ASU Adam, which is an AI course builder that goes out into the learning management system, pulls content from all these different courses and repackages them into something that is going to be a $5 a month consumer facing web app that I'm reading here that ingests faculty videos, lectures and assignments to generate personalized learning modules.

They did all this apparently without notifying the faculty whose materials were used.

So Rob, how would you react to that?

Yeah, when I saw this, Craig, I was like, I can't believe believe it.

But I can believe it.

But I.

But I can't believe it.

It confirmed for me something that I experienced as a young assistant professor straight out of my doctoral program where people were up in arms over who owned the video content, the material content that was put into the learning management system at the time for classes and faculty wanted to retain ownership and the Institution said, no, no, no, it's ours because you did it while we were paying you.

And I didn't see at the time why that was such a big deal, but now I do.

Right.

I can totally see where faculty feel that they own their name, image, likeness, right.

Much like our athletes deal with.

And to have that taken and packaged to push out.

It really is a threat, I think, to faculty who are afraid that there may be things with AI that will replace them.

And if the model's being trained on everything they've done, that threat becomes even more real.

It does.

And at most institutions, they invoke something called work for hire.

I think that's the principle that says if we paid you to create it, it's ours.

Which is why I'm very careful.

You probably are, too.

When we work on our textbook, I do not work on it on any university equipment.

I don't print anything out at the university.

I do all of that on my own equipment.

Now, I've never been anywhere where they would have tried to claim ownership over something like that.

And I think textbooks are largely seen as an exception because we're not paid by the university to write textbooks.

We get no real credit for it.

But that's the operating principle, is that intellectual property belongs to the university because they paid the faculty member to create it.

So even if we accept that, I think there's some real drawbacks to this, and I get why ASU is interested in doing it.

And it does have some advantages in terms of lifelong learning, public engagement.

They are a public university, so there's a lot that can make you think okay, in principle, the idea of doing this may be okay, but there are some real downsides.

So, first of all, it pulls things out of context with AI that may or may not be able to judge the context effectively.

So if a faculty member has a lecture that calls back to another lecture or to a prior course, I'm not so sure AI is going to be able to put all of that together.

So you may have this disconnected set of chunks of knowledge that aren't really flowing in any coherent way, and maybe that's addressable.

I don't know.

I think the bigger problem is it's going to lead to course materials being less accessible.

So if I'm worried about this as a faculty member, the first thing I'm going to do is just not post things to the lms.

There are already faculty doing that because of the accessibility guidelines, which got pushed back another year.

But I mean, why, if you're worried about this, okay, I'm just not going to have my lectures recorded.

I'm going to minimize what I put on the lms.

I might put it on my own Google Drive or something to let students get it.

But that's going to make course materials less accessible.

And I think that's going to have a detrimental effect on the most vulnerable students, the ones that are trying to work and go to school, ones that are trying to take care of families and go to school.

I don't know, maybe I'm off base there.

What do you think?

No, I think you're spot on.

And what I think is an interesting tension that's being wrestled with, and I think AI is just amplifying this incredibly is between what's in the best interest of our students versus the financial impact of various different decisions that are being made.

And when you prioritize one over the other, then we create these unintended situations where the students ultimately suffer or certain students do.

And I see it especially as we move into this world of what can we do for kind of that lifelong learning?

How can we take what we're already doing and make it accessible to people outside of higher education who may need to upskill and to do some of those sorts of things?

I'm not sure that the interest of who our core undergraduate, in person student is is completely aligned with that model for others.

But this is now an easy way to take some of that material and make it accessible that way.

And so I'd hate someone who's at a land grant institution myself.

I would hate to see us make decisions in how we do things.

That kind of takes our eyes off of that vision of the student population that we're trying to meet, at least in the in person, on campus section of classes.

Yeah, absolutely.

And I think some of this could have been solved by doing a smaller scale and engaging with the faculty and paying the faculty.

You know, if somebody came up to me and said, hey Craig, we want to use your course to make something available to the public and here's how we're going to do it and here's what we want you to do, they said, we're going to pay you X number of dollars, I mean, I might be okay with that, but I don't think upper administration at ASU cares much about that.

The other thing that's kind of a buried concern here.

I think one member mentioned this in one of the news sources I looked at is this is going to expose faculty, especially that discuss certain topics to substantial harassment.

So let Me lay out an example on both sides.

Let's say that I'm teaching a class that deals with diversity, equity and inclusion.

And I talk about the criticisms of DEI programs, and there are legitimate criticisms of those programs.

Well, somebody on the left gets a hold of that and it's taken out of context.

The criticisms of DEI module that anybody with five bucks a month can now go access.

Well, they pull that out.

It's completely taken out of context and they don't know that.

I just spent 30 minutes talking about all the benefits, and I'm going to spend more time talking about how the benefits outweigh the risks and how the risks can be addressed.

None of that's there because it's pulled out of context.

Or on the other side, I could be talking about the benefits of dei and we have the exact same situation going on in reverse.

And you will see faculty get attacked over these kinds of things.

We've already seen it with somebody with a, you know, that just has their cell phone out recording something.

This is just going to do that at scale.

Yeah.

And the other thing that I think's an interesting side of this is they put these into TikTok, like videos that from doing all this is they're short snippets.

So again, you miss that context.

But it's also, with those shorter videos, it's very much easier to take another device and record what's going on and to capture that.

So then is the protection in place of that intellectual property from the sense of does this just make it easier for that material to get out and go viral and have no level of control about those sorts of things?

Well, and you throw that in with deepfakes and forget about harassment, you could have considerable misinformation and disinformation campaigns built around legalization, legitimate faculty members.

So I can see the upsides to this kind of program, but it's going to be a mess, is my prediction.

And it is just a pilot.

So it may end up dying out, but I wouldn't bet on it.

Yeah, welcome to the wild, wild west, Greg.

That's right.

That's right.

So, you know, one of the things we were talking about the other day, Craig, you and I was about some things you're doing with Codex and programming sorts of things.

How are you playing with this in a way that is helping you to be more productive?

So, Rob, I have a list and so I'm going to go through that list with you because it is fairly impressive.

So for listeners who may not be familiar with Codex it was developed, as the name implies, to generate computer code so you could give it a goal and it would help you plan out how to achieve that goal.

And then it would write the code for you and test the code and do lots of stuff like this.

It's more autonomous than using a chatbot.

Chatbot is going back and forth, back and forth, back and forth, and you kind of guide everything the chatbot does.

This is an exaggeration because the line is getting a little bit blurred with these coding systems.

It goes out and just does things.

And you can set it up to where it has to ask you permission.

Like, I've got one that we found out about the ASU thing is, I've got a briefing.

This one uses CLAUDE code.

It's called a routine.

Every morning at 9 o' clock, it goes out, scans the Internet and finds news items that might be of interest to people that are interested in AI and higher ed.

And then, rob, I send them to you.

And literally that's how we found out about the ASU thing.

And so it's great, but it just goes out and does it.

All I have to do is I have to click on one button because I'm not comfortable when I set it up, I wasn't comfortable having it right to my machine without permission.

So I have to click on a button that says allow, and then I've got this briefing, but it just goes out and does it.

I don't know how it does it.

It just goes out and does it.

Codex is OpenAI's version of that same thing.

It's CLAUDE code with anthropic.

And then Gemini or Google has one called Anti Gravity.

So I took a bunch of transcripts, messy transcripts, from a set of focus groups that I'm doing with some colleagues.

These are in different formats.

They're in different, not just formats in terms of the transcript, but in terms of the file.

Some are word files, some are text files, some are actual transcript files.

There's a VTT file format.

I wanted to get all that put together in some reasonable structure that would have taken me hours to do.

You've done this kind of thing where it's all this copying and pasting and it's just a nightmare.

I said, hey.

Codex pointed to a folder where I had all these transcripts.

I basically gave it a little prompt and said, go to town and do this.

And then it was just a couple of minutes.

I had de identified transcripts.

They were in a CSV file, kind of a spreadsheet file, where it was sentence by sentence, identified the person, the focus group.

Just beautiful.

Exactly what you want in order to be able to process this further.

Plus it was de identified.

A lot of these were zoom transcripts where it would say Rob Crossler on there, Craig Van Slyke.

So all that was pulled out for confidentiality.

It was great.

But then I pushed it further and I don't think we're going to publish any of this part of it, but I just wanted to see what it did.

I had IT run a number of really sophisticated text analysis algorithms.

I mean, these are kind of cutting edge things that helped us understand what's going on with the transcripts.

It did it in minutes, had fully documented Python code, so it's repeatable.

I could do the same thing with another set of focus groups.

If anybody wants to audit it, they can audit.

And just as a test, again, I would not publish this.

I had IT write the introduction to a practitioner oriented article about these focus groups.

It was pretty good.

All of that took place in a matter of minutes with minimal direction for me.

I had IT go through and look at all of these briefings to pull out the best topics for us to talk about today or in future episodes.

It helped me prep our episode plan.

Here's one that instructors may be very interested in.

So we do a summer seminar that helps our doctoral students prepare for their comprehensive exam.

Do you remember what comps were like?

High stress.

Everything's a mess.

We decided a number of years ago that we're going to spend the summer helping our students refresh on these things.

They still have to do a lot of work, but it's a big help.

But we do these every two years and the seminars change a little bit.

New papers might come in.

We lost a faculty member to another university, so I want to pull out things that were really particular to him.

It would have taken me five or six hours to do this by hand.

I also wanted to audit the slide decks that we have against the new schedule.

We move some things around because of faculty schedules, so the order's not the same.

Anybody that's taught before has done this kind of thing when they move from term to term.

So it's that kind of stuff.

I just gave it some instructions, let it go in 30, 40 minutes.

It had all of this done and I went through and spot checked it.

I'll do a more careful check.

But it did all of that kind of thing, including I had forgotten to change 2024 to 2026.

It found that pointed out where the slides needed to be updated.

There was one slide deck that the former faculty member handled.

It was thin on articles for them to read.

It went out and found some really, really good articles for us to include and I could go on and on.

The best one though, was it went through over 130 PDFs that had no cohesive logical naming scheme.

Yeah, these are.

You download something and it's 0524361 PDF.

It went through, looked at the PDF and named all of them with first author et al, year, first few words of the title, which makes everything easier.

And it did all of that more or less on its own.

It even renamed the folders to align with the week.

So what we're going to talk about in week two now has 02 dash in front of the folders where all I had to do was upload it into OneDrive.

That's just phenomenal.

So listeners, if you're not using codecs or Claude code, start looking into it because it is not just for coding.

I didn't write one word of code and the only code that I would ever even use out of this is the Python documentation on the text analysis.

I'm out of breath, Rob.

I've got so many questions, Craig, so many questions from all that.

So the first one is it sounds like you went from not being very trustworthy of the process because you click that button to tell it to go off to do something, to where you developed a level of trust where you're letting it modify things on your computer and so on and so forth.

What kind of precautions are you taking from a backup perspective so that if Claude or Codex, whichever one, decides to go off the rails and just start deleting all, it's not going to be a giant headache that makes you your own victim of a ransomware attack.

Well, no, and this is a good.

I'm glad you brought this up because this is a big caution.

All of this was what we might call sandboxed.

So I created a new set of folders where I just copied everything from what I really worked in over to this new set.

So if disaster happened, I still have all the stuff I had before.

And so you absolutely want to do that.

If you're working with these coding agents, you could do it on a different drive.

I just have it in a different folder structure and don't give it permission to my main set of folders.

So that was the big thing.

And I had worked with it, you know, it had not gone off the rails on these briefings and some Things like that.

So I did trust it more, but I'm not going to trust it completely.

Sure.

No.

And I would encourage anyone, if you don't understand what Craig's talking about with sandboxing and putting things in a separate space, have those conversations with your AI agent about how to do that and it will give you some pretty clean instructions about how to go about doing things.

This is one of the places where I think there's some real danger in making the power of AI available to everyone is if you don't have that technical background of thinking about how this stuff works, you might inadvertently make a poor decision because you lost access to data, you lost access to things, and you kind of learned the hard way what some of the degrees in the technical space may have taught you for how to think about these things.

If you're really worried about it, back it up onto a thumb drive or some sort of an external drive that you then disconnect that.

And I also don't ask it to delete a bunch of stuff.

I mean, some people use these agents to go in and clean up folders and delete stuff.

I actually did that with my downloads folder because there was nothing in there that was really all that critical and it did a great job.

It's like, yeah, I'm not ready for that.

You know, I'm just not ready for that yet.

So I'm sorry to use a technical term like sandbox, but it just means make it work on a copy.

Is the short version perfect?

Another follow up question to this is you talked a lot about that first part of how you did things with those transcripts and the coding and those sorts of things that you wouldn't publish them.

And I'm curious, why wouldn't you?

It seems like it was a great help and it was something that would be good.

Why would you not use that to help make you more efficient as a researcher in that space?

Well, so a couple of reasons.

The biggest one is pragmatic, because I'm not going to lie about the fact that I used AI and journals will reject it just because you used AI.

And so that's the pragmatic piece of this.

Until I do it a lot, I'm not going to trust AI to do it well.

Now the spot checks that I did indicated that it was pretty good and maybe better than what a human would have done, but I'm not quite there yet.

So I'm getting ready to do another set of papers that use a fairly sophisticated technique that involves a lot of human Labor.

So what I'm doing there, and I will publish this is I'm having it do the first step and then I'm auditing the first step, and then I'll have it do the second step and I'll audit the second step.

I think it can get rid of a lot of the drudgery of doing some of this kind of text based analysis.

For example, I might have it go through and extract codes, I'll give it codes, have it go through and extract codes, and give me a confidence level and then I'll look at the low or medium confidence level codes, spot check the others, do that sort of audit, and then move on to the next phase.

So right now a lot of people have this feeling that there's AI slop and AI has never touched it.

But there's this huge middle ground that we're going to have to figure out where we're using AI to let us do better research and produce knowledge more effectively and more efficiently.

But it's not AI slop.

It's still something that was done with human oversight, kind of like we've been doing with GAS for a long time.

You've worked with gas, I've worked with gas.

You don't just say, hey, go do this thing and then try to publish it.

You say, do this piece of it, let's sit down, go over it, we'll talk about what you did, right, what you might have done a little bit differently.

We'll work through it, do the same kind of thing with AI.

And I think that's the middle ground, but who knows?

I don't know.

Yeah, that's very interesting.

And I want to bring this back to our undergraduate students and that experience as well as I heard you talking about all the really great things that these tools are doing, I can see our students becoming much more efficient in what we're asking them to do in the classroom.

And as part of that learning process, how do you see yourself keeping your undergraduate classes still resilient to the point where when student comes into your class having paid to receive some sort of teaching and learning, that they receive that, and then how does this also begin to maybe even change what that process of learning looks like?

I don't know.

What are you asking questions like that for now?

I think what we need to do, we talked about this before, maybe not in these exact same terms, is we've got to help students learn how to learn with AI and how to, to use AI in ways that let them do more and better.

I mean, that's a terrible sentence construction, but I mean, that's what it is.

Here with me is some of the stuff I just wouldn't do because I don't have the time.

So now I can do it, and it's going to have some interesting results.

So I think those are the two things we need to really be working on.

And it's not easy.

It's going to require some big rethinking.

One of the ideas that came to mind as you were talking about the things that you're doing and what's possible with undergraduate students is oftentimes we create fairly well controlled canned experiences for them where there'll be projects or various different things that are small in nature with some good boundary conditions on them.

And what I heard you describing of getting all that data ready to be able to use with the graduate students and renaming files, those are real problems that real organizations face all the time.

And so are we creating or do we have the ability now to create in class experiences that are going to be more closely mirrored to the issues that organizations face so they can not so much as come up with that final product of the study guide for a comprehensive exam or the course design for a comprehensive exam, but really that process of organizing and synthesizing and bringing together big, ugly, messy things, you know, that process to me sounds like one that's very repeatable in so many different situations.

If you kind of change the way you think about your approach to that.

Yeah, absolutely.

Things like updating policies, you know, a new something comes out and now you've got to update policies.

AI is great at that kind of thing.

It's going to be more detail oriented than most humans would be.

I'm going to go out on a limb here.

So this may be one of those things that gets cut out, but in my ideal world, a lot of that stuff, that's just us standing up talking about it to students and them sitting there listening.

I think we should offload a lot of that to AI.

Well, I think what you're saying, though, is a move towards experiential learning.

I heard this when I was visiting Tennessee.

It's kind of, you know, we need to be less the sage on the stage and more the guide by the side and helping students work their way through problems as opposed to just getting in front and sharing our espoused knowledge of everything we know about everything.

Yeah, we've been saying that for 30 years.

Maybe we'll actually do it this time.

Maybe our hands being forced.

But you know, the other thing Is, I mean, I'm not.

So I don't mind giving lectures and I think I'm at least in, okay, lecturer, but it's not the fun part of the job.

It's the easy part of the job, but it's not the fun part of the job.

So I don't know.

The other thing I would put out there for our listeners is you need, once again, we've said this before, you need to be playing around with these tools or you're not going to understand how your students are using these tools.

And so by pushing the envelope myself, I can sit down with, in this case the doctoral students and say, look, here are some things you can use.

Codex, Claude code could have done the same thing.

Here are some things you can use Codex for that will help you get organized for your comprehensive exam preparations.

And for those of you who haven't gone through this, take an average doctoral seminar.

Let's say it's 10 weeks worth of readings at five papers.

So now you've got 50 papers at a minimum and you've taken at least six of these seminars.

Was that 300 papers to try and get them organized?

I mean, that's a lot.

We have one doctoral student who we tease about how organized she is and I think you should do a side hustle here and just get everybody else to pay you to help them get their stuff organized.

But I mean that's a huge win and we all have this when we teach, right?

You know, got a new edition of a textbook.

What do we need to update?

Help me figure out some out of date test questions or what needs to be changed in my PowerPoints.

There are a thousand things that we do that are like this that have to be done, but they're kind of a pain.

So here's the big message.

Try Codex or Claude code or on something.

I, I have not played with anti gravity for this kind of thing.

I suspect it could do it, but I'm not willing to say that it can yet.

Oh, I mean it was impressive.

It was several days worth of work in total in a couple of hours and work that I hate doing.

I think that's a great point to wrap up on, Craig.

You've kind of hit the nail on the head with that takeaway of the play with these tools.

Maybe it's Codex, maybe it's Claude code, but if there's other ones, you're like, oh, I wonder what's possible.

Or that bothers me.

I'm not quite sure how I'm going to teach with that.

Tool.

The first step is to understand how it works and just do something with it and begin to wrap your head around it.

I think that's a great idea for any of these tools as they continue to change and be new things being deployed with them.

That should be a strategy that everyone can take away from the show today.

Absolutely.

And you will likely find some pain in the rear end things that it'll do for you that you don't have to do anymore.

But to make sure that we're fair and balanced, I have to talk about a pretty spectacular AI failure that I'm pretty disappointed in.

This one.

So do you use Gemini at all?

So Gemini has always been really good for me at all.

That tech support kind of stuff.

I mean, we talked about when I figured out what was going on with my propane tank during a winter storm using Gemini.

So I'm upgrading my mesh network at home to one that will fail over.

So we finally have fiber, but it's strung along the power lines through the woods.

Nothing can go wrong there.

Yeah, look, I'm not complaining, but if it goes out, if a power line comes down, so does the fiber.

As soon as the lines are back up, fiber comes back up.

So I've kept my Starlink subscription, and I want something that will automatically fail over to where if fiber goes out, it'll kick over into StarLink and the TVs and all that stuff will just keep working because they're all looking at this mesh network.

So Amazon's got these eero eero 7 devices that are pretty good and they will do this.

So Gemini helped me find those.

That was great.

They came, I set them up.

Really easy to set up, but trying to get it to where it would automatically fail over when the fiber went out was a little bit of a challenge.

Because the fiber's down, but the router's still running.

The WI FI router is still running.

I'm not sure if that router is still running.

Is the mesh network going to keep trying to use the fiber or is it going to kick over?

So there's a thing called bridge mode that you can put these routers into where it just is a pass through.

Not a big deal.

Except it is.

And so I spent about an hour and a half yesterday, right around dinner time, trying to figure out how to make this work with Gemini, and it just couldn't do.

Kept giving me bad information.

And I eventually had to get on tech support with Xfinity, which, ironically, once I got to the right place, its AI assistant was quite good.

I mean, it couldn't solve the problem, but it tried.

And what it was trying made sense.

And it finally kicked me over into somebody, and we got it all resolved.

So the point of this is not to regale everybody with my irritations around this sort of thing, but here's the big message.

The reason Gemini was failing is that there are hidden things in the fiber provider where they change stuff, and it doesn't really get documented because they don't want normal people to be messing with this stuff and screwing it up.

And so Gemini did not have access to that.

So what it was telling me probably would have worked a few months ago, but it didn't work now.

And it's not going to work if a tech has come out to my house and changed something from the defaults.

So it was trying to tell me how to log in.

So most routers, the administrative user ID is admin, and the password is literally password, which you should always change, but people don't.

And that's what Gemini had.

Well, that didn't work.

And I tried a couple of variations.

Didn't work.

It locked out.

And this just kept going on and on and on.

But the bottom line here is these AI tools are great until they're not.

And it's a limitation of this hidden knowledge that every organization has, which is why you still need humans.

Yeah.

And I think that's a great example, too, of these machines, if you will, are only as good as the data that they have.

And so the training that they receive, where they get that from, There are limitations in that.

And the fact that you've played the customer service game long enough that you knew if you could get on the phone and talk to a human who could help you through this process, is that critical thinking?

I think that's important in all these different ways we do things.

We can't become paralyzed when we don't have a solution because the machine doesn't know.

That doesn't mean there isn't a person out there who does know.

But we need to have a series of ways we triage through these various different things.

And it displayed my own limitations because, I mean, I know in enough about networking to kind of know what's going on, but it's not my thing.

So I didn't know whether what it was telling me was entirely reasonable or not.

The other piece, and this is kind of bad on my part, is I was hungry, so I should have stopped earlier and didn't.

So, I mean, I really should have kind of bailed and said, let me get on with a tech support person.

What ended up having to happen is they had to reset my modem remotely.

Resetting it locally wasn't going to do it.

So once they did that, everything was fine.

But I should have known that sooner and given up.

But my stubbornness, and I'm going to blame my lack of food, would not let me do that.

Yeah, so don't ask Craig to do tech support hangry is what I'm hearing.

No, that is true.

I'm well known for what happens with that.

All right, well, I think that's enough for today.

Rob, any last thoughts?

No, I think we've plugged them all into the show today.

Hopefully this is a helpful episode for everybody.

All right, once again, we'd love it if you get in touch with us.

Rob Crossler, C R O S S L e r@AI goes to college.com or craig@aigostocollege.com we'd love to help you out and love to talk about whatever you want us to talk about.

All right, that's it.

And we will talk to you next time.

Thank you.

Bye.