0:05
OK, awesome.

0:06
Thank you, Chris.

0:07
So yeah, I’m going to talk about AI and we’re going to do kind of an overview quickly for those of you who were on the Ag Institute webinar yesterday.

0:15
It’s going to be repetitive, so you can pull out some other work if you need to do that, or you can do a little bit of multitasking.

0:21
Not that you wouldn’t do that anyway, but I’m going to try to tailor it a little bit to crop production.

0:26
One of the things I want to just kind of throw out there quickly, I did not talk about this yesterday.

0:30
This is a publication that is, oh, my actual little blurb of the presentation is or of the publication isn’t showing up, but we have a brand new publication hot on the press.

0:42
It’s actually in the preprint phases and it’ll be published next week.

0:45
It is called Multimodal large language Models for Agriculture, A comprehensive review, and in this we have about 150 different use cases.

0:54
It’s not just writing reports and analyzing documents and doing paperwork.

0:59
This is actually the use of large language models for making crop production decisions.

1:05
I argued a little bit with my co-authors.

1:07
It’s very crop and plant focused.

1:09
There’s very little in animal agriculture, but there are some sections for those of you who are extension educators, consultants, agronomists who are helping people make decisions.

1:19
There are some tools that are talked about in this publication.

1:23
And by the way, so Chris, at the end of my presentation, if you wouldn’t mind the same way you did with Josh, if you could throw up my e-mail, I’m happy to send you a copy of this while it’s still in the preprint phases.

1:35
So I’m going to talk about, I’m going to talk about artificial intelligence.

1:41
I’m going to use the term AI, but very specifically what I’m talking about here today is generative AI.

1:48
In that previous slide you saw large language models and multimodal.

1:52
One of the things I want to just let you know, it sounds like a fancy word multimodal.

1:56
It simply means that these large language models can handle information in, in various modes.

2:03
They can take in sound, they can take in, you know, anything that’s obviously audio, video, photos, documents, things like that.

2:13
So, there’s a lot of ability to do 2-way exchange of information.

2:18
They help us to generate new stuff and I’m specifically using the word stuff for a reason and I’m going to come back and talk about it in a second.

2:28
By stuff again, we’re talking about text and video and images and audio and code.

2:33
Increasingly, people are using generative, generative AI to do coding for things like statistical analysis, doing data, downloading, data organizing.

2:44
And if you’re not a coder, I’m not.

2:47
Even though I’m in an AG engineering department, I’ve tended not to write code since back when I was a grad student.

2:53
We now have access to being able to do some pretty cool and powerful stuff, AI as we know it, with tools like Gemini, ChatGPT.

3:03
I think Sean’s going to talk about Gemini in the next session quite a bit.

3:07
These tools really jumped to the mainstream back in November 2022, and that was with the release of the open AI product called ChatGPT.

3:17
But these are based on a fundamental technology that was introduced first back in 2017.

3:22
You’ve got the yellow highlighted citation.

3:25
It’s called Attention is all you need.

3:27
That was the name of the paper.

3:29
It’s actually a fairly easy paper to read, has 207,000 citations.

3:35
For those of us on campus who do a lot of technical writing and we like to get our articles in peer reviewed journals.

3:42
For me personally, if I have a article that has a couple of 100 citations, I’m really happy.

3:49
And this is one that has been cited.

3:51
Literally, it is like the seminal, like the main thing that everybody goes back to.

3:57
So, I said it’s been around since 2022, like the usable types of programs and apps and the Chat GPTS and the things that we can all access now on their on our phones.

4:10
But actually this has been around for a while.

4:12
Our phone is full of AI.

4:14
Google is obviously based on the science of machine learning and artificial intelligence.

4:22
You know that when you text and it gives you that sort of next word prediction or you sent or you’re writing a Word document and it gives you like what it’s predicting to be the next phrase.

4:33
That’s exactly what AI does.

4:35
So we’ve had this for quite a while, but we’re also seeing it integrated increasingly into equipment.

4:41
So, tractors, control systems, if you drive a, a new vehicle, pickup truck or a car and it has adaptive cruise control.

4:51
With adaptive cruise control, we’re using the speed of the vehicle, we’re using images from cameras and radar and we’re able to predict like what is that safe speed?

5:02
What is that safe following distance?

5:04
And again, that is largely using artificial intelligence.

5:08
And eventually we’re going to see even more of this.

5:10
I think a lot of you probably have seen some of the pictures that are in this montage that I’m showing here, a lot of cropping field equipment, in particular weeders, sprayers, the sprayer in the upper right-hand corner made by a company called GUSS.

5:27
GUSS is a field sprayer, an orchard sprayer it was just bought out by John Deere.

5:31
So there just are so many different things happening.

5:34
By the way, I just want to make sure from a credit perspective, these are all pictures that I’ve taken at various events and field days with the exceptional lower right-hand corner.

5:44
My friends at John Deere really always want their photos to be credited.

5:47
So that’s obviously a John Deere autonomous tractor out in the field.

5:52
Looks like it’s doing probably doing tillage work.

5:55
So back to this idea of generating stuff the way that these models work.

6:01
And I want you to again, think about ChatGPT or Gemini.

6:05
A lot of you have probably tried these tools.

6:08
What they do is they use training data.

6:11
What is training data?

6:13
Training data for a model like ChatGPT is literally every single publicly accessible website.

6:22
For a long time, they were also digesting and chunking every single YouTube video. Anything that’s publicly available.

6:30
Technical reports, government documents.

6:33
Yeah, extension bulletins have been sort of digested and brought into these models and what they’re able to do.

6:41
When we go through this process of training a model, we literally, we call it chunking.

6:45
So, so chunking it into pieces.

6:47
And then we create things that computers can read.

6:50
And when you ask it a question, your question similarly gets chunked and vectorized and they get kind of matched together.

6:58
And we use these models to like, assemble and rearrange these new chunks of information.

7:05
And the reason why I think it’s important to kind of go into this technical explanation is we’re not really creating new knowledge, right?

7:13
We’re not.

7:14
It’s not wisdom.

7:15
It does not replace the experience of a crop consultant or an extension educator or a specialist or an agronomist.

7:24
What we’re doing is we’re rearranging knowledge and data and information that has already been created.

7:30
Now, there’s some exceptions to that.

7:32
We’re beginning to talk about things like synthetic data, but for the, for the most part, we need the research at places like UW Madison and other locations, other land grant universities, other private sector consultants working with us getting farmer because we need to continue to generate this information and knowledge that we can then reassemble and make meaningful for people.

7:56
So, they are essentially next word predictors.

7:59
I’m going to give you a very simple example.

8:01
Couple of months ago, Chris asked me, could you come on on the Thursday morning to Badger Crop Connect and do this webinar. So, you know, I can go on.

8:09
I didn’t do this for this particular presentation, but I can say, hey, ChatGPT, would you be willing to like write me like the first like few minutes of my presentation?

8:18
I’ve only got 22 to 25 minutes to present, so I want to make sure I’m really efficient.

8:23
And yes, it can go do that.

8:24
I could probably tell it it’s going to be on a Thursday morning.

8:26
It’s going to be on Zoom.

8:27
It’s going to be for this group of people. When it gives me that output, it’s actually going to use past information.

8:35
So it’s probably digested several 10s of thousands of Zoom presentations that were up on YouTube, right?

8:43
So good morning, everyone.

8:44
Thank you for the introduction and I’m happy to be blah, blah, blah.

8:47
So 75% chance, 75% probability that that’s going to be the case.

8:52
So thank you.

8:53
Good after or good Thursday morning.

8:55
I guess now we’re an afternoon, right?

8:57
Hope everybody’s doing great, having a great 2025 growing season.

9:00
Our greetings from Madison.

9:01
I hope everybody’s doing great.

9:03
So you kind of get the idea.

9:05
It’s nothing magical.

9:07
The other thing that’s really problematic though, for some people and for some uses and actually for some individual times in some points in time when you use these models, they can come up.

9:22
There might be a probability like a 0.005% where they say good morning, Zoom world.

9:27
I hope everybody is having a rainbows and Unicorn start to your crazy Thursday.

9:33
Seems silly, but these kinds of things happen.

9:36
Maybe not this exact situation, but that’s what we call hallucinations.

9:42
We have to remember that these models are, we’re simplifying it here, but we’re at the mercy of these probabilities.

9:50
And when we see these models come up with bad answers or things that are just silly, that’s why it’s happening because we are using these probabilistic functions to regurgitate these chunks of information that the AI has digested as part of its training set.

10:11
So what we’re talking about here are tools like Copilot, Gemini, ChatGPT, Perplexity, Clod Llama, There’s a whole bunch of them.

10:19
There are a couple of Chinese models that are now being used that are open source.

10:24
But I want you to understand, like, they’re not magic.

10:28
They cannot read your mind.

10:30
They use these statistical methods to extract and reassemble data that can sometimes lead to incorrect, hallucinated, sometimes embarrassing, and sometimes even unsafe results.

10:43
I’m going to show you an example here in just a moment.

10:47
All of that said, it sounds like I’m super negative, right?

10:50
Like I’m like a naysayer.

10:52
I use these products all the time. In my work.

10:56
I’ve got research going on in this space.

10:58
Obviously, I was a contributor to that big article that’ll be in the IEEE Transactions.

11:03
So, I do think that they hold amazing potential for the kinds of work that we all do.

11:11
But I do think we need to be super cautious and super aware of how they work.

11:17
Let me see if I can tell this story really quick.

11:19
Safety story. Affected me personally and remember my my background.

11:23
I’m a farm health and safety specialist.

11:25
So back in February, I’m embarrassed to tell you this, I fell, I ripped the crap out of my leg.

11:33
I separated my quadriceps tendon, which basically means you can’t lift your leg because you don’t have a quad tendon.

11:41
And I ended up having to have emergency surgery on March 5th.

11:45
I was in a straight leg brace from March 5th until the 1st week of June about 1212, 1/2 weeks. When April first rolled around.

11:54
The grass is starting to green up.

11:55
It’s I’ve got about a 7/10 of an acre lawn.

11:58
I was getting stressed.

12:00
I got some prices like there’s no way that I can push mow it anymore.

12:04
I’ve got between a 15 and a 23° slope in my backyard.

12:09
It was impossible for me to even walk on this let alone get a push mower.

12:14
So I go down to my, I’m not going to name names, but I go down to my local green dealer and I said, hey, you know, what would it take?

12:23
Here’s some pictures of my yard.

12:24
Could you get me into a zero turn mower?

12:27
Because there’s no way I can afford to have somebody else mow my lawn the entire summer.

12:31
He looks at my pictures.

12:32
He’s like, yeah, this would be a piece of cake.

12:34
He brings out a zero turn mower.

12:36
It was a cheaper one, like one of their lower end, so it didn’t have a ROPS.

12:41
He mowed a couple rounds.

12:42
He’s like, yeah, this should be great.

12:44
And I’m on out on the lawn the following Saturday and it the grass was dry.

12:49
Everything was perfect condition wise.

12:52
And it was like I was on a shopping cart, like the front end was wanting to go down the hill.

12:56
And I’ve got at the bottom of the hill that you see in the middle picture about a three foot drop off of a Rockwall.

13:03
And it just scared it.

13:04
And I couldn’t like jump off of my mower because I’m in this straight leg brace.

13:08
So I call them.

13:09
I’m like, yeah, you, you got to come on Monday and pick this up.

13:11
There’s just no way I’m going to be able to use this.

13:14
So, once finally I’m able to finally mow the lawn in July. I’m push mowing it and I see on Facebook a Simplicity Lawn mower 2020.

13:24
No, I’m sorry, 2000 models.

13:26
So it’s 25 years old.

13:27
The guy only wants $700.00 for it and it’s only got 400 hours.

13:32
So I have him come out and he brings it out.

13:36
He gets it off the trailer and he’s like, he looks at my lawn and he’s like, yeah, this is, this is no problem.

13:42
And he’s mowing on my hill.

13:44
And he gets to that 23° section and he’s starting to tip.

13:48
And you could see his uphill wheel start to turn and his eyes get really big, right?

13:54
And I noticed that he turned. He wanted to get off the slope quickly.

13:57
So he turned up the hill, almost rolled the tractor over.

14:03
Later I said, like, Steve, like, why did you go?

14:05
He said, oh, you always want to go up the hill if there is a situation like this.

14:09
And I thought, God, I’m like, that’s not what I tell people.

14:12
If they’re packing a silo or just operating a tractor on a slope, you typically want to get your center of gravity down, not take it up.

14:21
So I go on to ChatGPT after Steve leaves and I’m back on that hillside, I’m like, hey, ChatGPT, I’ve got this mower.

14:27
I give it all the specs.

14:28
What are your recommendations?

14:30
Do I turn up the hill or down the hill?

14:31
If I start to slide and ChatGPT says you, well, of course, you dummy.

14:36
You turn up the hill and that night I’m laying in bed.

14:41
I’m like, no, that, that’s just not right.

14:44
Like that’s totally, that’s just incorrect.

14:47
And so I go to Gemini and I went and they’re all like, yeah, no, of course you turn down the hill.

14:52
I found the manuals.

14:53
It’s like, oh, no, absolutely.

14:55
First of all, stay off the slope if it’s anything greater than 15° and you always turn down the hill.

15:01
So the bottom line is these things can be unsafe.

15:04
Now, you may not encounter a situation like this, but if you’re talking about making herbicide recommendations or you know, making very specific agronomic decisions, these tools that we have at our disposal, they do not replace your knowledge, your intuition, what you have learned.

15:24
I just really want to emphasize that that there are safety implications that we need to consider and that’s why I think knowing how they work is really quite important.

15:35
I’m going to talk about the barn picture here in a second, but here’s how I use.

15:39
Here’s how I view large language models and generative AI.

15:43
View them as having a very talented assistant.

15:46
They can reduce a lot of the routine like mundane tasks, paperwork, report writing, finding information.

15:54
They can help us think in new ways.

15:56
They help me to be creative in the work that I do.

16:00
Here’s the biggest use case, and I’ll talk about this in a moment.

16:03
They help me to ask the right questions and I know that my extension colleagues and I know a lot of consultants, I do consulting work myself.

16:12
A big part of helping a client is helping them ask the right questions.

16:19
And many cases the question that they’re actually asking is not what they came to you with initially.

16:25
So kind of like honing in on the right question to ask.

16:30
The other thing I want to just mention to you is that these models, they’re only going to get better.

16:35
This is as bad as they’re ever going to be.

16:38
So, you know my, my example here with my simplicity lawnmower, I would say that in five years, maybe three years, the hallucination, the, the safety issues are going to get largely worked out.

16:50
But I still don’t think that they are going to replace us.

16:54
A few things that I use them for helps me generate new ideas or germinate new ideas, new research, you know, things that I might be thinking about from a research perspective.

17:06
I use it very, very often to outline documents. For me when I’m writing, the hardest thing is getting started.

17:14
So a lot of times I will, I will tell it what I want to write about and I’ll ask it for an outline to at least get me started.

17:22
It also can help me to tighten writing if I’m writing reports, if I am publishing a newsletter with somebody or I’m writing newsletter articles for a magazine or a newspaper or just did something

17:35
a few weeks ago with Vita Plus.

17:37
They also helped me to summarize complicated information, talks, articles, peer reviewed publications.

17:43
They can help me to identify new sources of information, but they don’t replace a good like literature search, whether it’s Google Scholar or some of the library databases.

17:54
They can also help me develop learning plans if there’s something that I feel a little bit uncertain or insecure about.

17:59
Lately I’ve really been wanting to learn a little bit of basic Spanish for when I visit dairy farms.

18:06
It seems that all the workers I interact with are Spanish speaking and I’d like to know a little bit more.

18:11
It can give me kind of a, a formulaic way of, of doing the educational work, kind of guiding me.

18:19
It’s the same I play guitar, you know, give me a plan for learning the G scale or learning a certain song or whatever it is it might be.

18:28
So an example of how I might use this, the article that I showed you that I’m happy to e-mail to you that 150 different use cases.

18:36
I said, hey, ChatGPT, can you summarize the uploaded paper and tell me what it means for me as a crop consultant?

18:42
And it gives me like, instead of having to read 33 pages of densely written journal article, it gives me some like high level like here are the things it talks about.

18:56
You could ask it questions about the document. You could ask it questions about the research.

19:01
And just an example, by the way, that 33 page paper, I’ve had to go through it three times because this journal, they are so picky about everything.

19:11
So we’ve had to go through it with a fine tooth comb.

19:13
I never want to read it again.

19:15
But I do feel like I could go back to it with questions like this.

19:20
So, if you have had issues. If you, if you have used these tools and you feel like you’ve kind of struck out, like, yeah, they’re, they’re crap.

19:29
I, I’m not going to be able to use anything like this.

19:32
What I would suggest to you is to maybe think about it a little bit differently.

19:38
Think about the questions that you are asking.

19:41
Think about the types of problems you are wanting to solve.

19:44
And we talked about this in terms of prompting the AI, asking the AI questions or maybe a better way of looking at it is giving it directions.

19:54
So, I like to think of when I use Gemini, I use, I use it a lot because we’re allowed to do that at the university.

20:02
I like to think of it as having an intern.

20:05
That intern is super talented.

20:07
They’re like a undergraduate at River Falls or Platteville or UW Madison, but they’re majoring in like everything. They know

20:17
quite a bit about from my book knowledge. They can begin to help me think about solving problems, but at the same time, this is where the caveat comes in.

20:26
They can wander off, they can get like really lost in the cornfield and like really take me down the wrong direction.

20:33
So the more, and this is the same way if I’m working with students, the more I can be precise with my questions and my prompting, the better off I’m going to be.

20:42
They also may not get those connections that I have.

20:45
You know, I’ve got 35 years of experience and working in my field.

20:50
It’s partly why when ChatGPT told me like, Oh yeah, you dummy, always turn up the hill with your mower.

20:55
Like, no, that’s not intuitively.

20:57
That just doesn’t, that’s not right.

20:59
So, you’ve got to be able to be, I think Shawn may talk about this, the human in the loop.

21:06
And it is again, why it will not replace us.

21:08
When we give an AI specific directions, we’re literally pointing it into the universe.

21:14
It has this universe of data and information.

21:18
And what we’re saying is I don’t care about 99.999% of that universe.

21:24
I want to get to the right galaxy and the right solar system and the right planet.

21:29
And if we’re talking about Wisconsin, like even the right moon, right?

21:33
Like we want to narrow down that universe of information.

21:38
A couple of hints for you.

21:39
LinkedIn.

21:40
I find LinkedIn when I’m working with AI.

21:42
If you’re not in on LinkedIn, there’s a lot of cool resources.

21:46
An example is like this infographic.

21:50
There’s a, there’s literally a million of these, well, maybe a couple hundred of these like frameworks for prompting AI and if you got to find something that you can work with.

22:02
Also, if you have access to LinkedIn learning, there’s some cool courses there on prompt engineering.

22:08
And for me, it’s been worth the investment of a little bit of time.

22:14
I don’t. Let me just go back.

22:16
This is the same thing I did yesterday.

22:17
I don’t, I don’t rely on one of these like RTF or Solve or TAG. I tend to go a little less formulaic and I err on the side of having, of giving it too much information.

22:31
So again, I think about it, it’s like my my undergraduate student who is sitting there, who is massively smart and I want to tell that that student exactly what to do and where to go.

22:42
Who am I?

22:44
What is it that I do for my work?

22:46
What is the audience of this thing that I want to produce?

22:50
And then specifically, what am I asking you to do?

22:54
Do I want you to write a 400 word article for a Vita Plus newsletter?

23:01
Do I just want it to give me an outline?

23:03
Do I want it to come up with that cool infographic like you saw on that last slide?

23:08
In what format and where should I look?

23:12
I tell anytime I prompt a large language model, I will say, please stick to the work of land grant universities in the upper Midwest and associated government documents from places like state departments of agriculture and the USDA.

23:31
And you can you can point it into that part of the universe so that it’s not drawing from the whole rest of the world.

23:37
You can also provide information.

23:39
Do I want it formal?

23:40
What style do I want it written in?

23:41
What’s the tone, the texture, the length?

23:45
And when we’re talking about giving it context, right?

23:49
That’s what this is called as giving it context.

23:51
You can, you can upload similar pieces similar to what it is you’re looking for.

23:56
Very specific example.

23:59
I think I’ve got one or more couple slides here, but I’m going to build this one out. At a phone call

24:06
Not this very farm.

24:07
This is actually a farm right up the road from me in Cross Plains.

24:10
This woman calls me.

24:11
They are retired dairy farmers.

24:13
She says my husband and I, well, she goes, no, actually it’s my husband is thinking about turning the farm into a bed and breakfast.

24:22
And specifically he’s got this idea of converting the old silo into a sleeping room.

24:29
I’m like, oh, I’ve seen that before.

24:31
And she says, yeah, but it is full of bats.

24:33
And so I need your help.

24:35
So normally in the old days, I would have, like, found, you know, a couple articles on histoplasmosis, and I would have put it in the mail because they didn’t have e-mail.

24:42
And like, yeah, probably not a good idea.

24:45
Instead, you can use an AI to do something like this again, using my formula.

24:51
I’m a farm health and safety specialist, UW Madison.

24:54
I got this phone call from a retired dairy producer.

24:56
Her husband wants to convert their old silo into a bed and breakfast sleeping room.

25:01
She, she’s really worried about health.

25:04
This is really important context, right?

25:06
It’s pretty clear that the couple is they’re on totally different pages.

25:10
And so the question that I ask it is, could you help me to formulate, like the right questions?

25:17
I didn’t know what to ask them other than like, so how many bats are there?

25:20
What kind are they?

25:21
It’s like those aren’t the right questions.

25:24
Instead, when it came back, you know, I talked about structure.

25:27
I talked about the Department of Health and Regulations.

25:30
And by the way, even like maybe you could help the couple like refer them to someplace else so that they can get their conflict worked out.

25:39
So again, just an example of helping people to ask the right questions.

25:43
This approach puts you in control.

25:47
When I, when I’m writing something with AI, I typically ask it for outlines, like big picture ideas and outlines.

25:54
This helps me to personalize my response, my writing.

25:58
It leads to authenticity.

26:00
Sometimes it requires multiple iterations.

26:04
I also think it helps me learn.

26:07
I view my job like, yeah, I’m supposedly an expert in my field, but I’m a learner.

26:13
And I think if you’re a good consultant, if you’re a good educator, you are also a learner.

26:18
And then it allows me to go back, internalize, to own what I’m doing.

26:23
And I will have to check in and double check the work.

26:28
Just a quick peek at the future.

26:29
I have 10 seconds left.

26:30
We are developing a recipe for individuals to develop their own like body of knowledge and an associated large language model that will kind of go with them so that they can have a more targeted, tailored assistant.

26:49
And if you’re interested in that, we’re using a technique called RAG. Retrieval Augmented Generation, which every single person who’s on the Zoom here can access through a program called Notebook LM.

27:02
And I guess with that, I am going to conclude. I don’t know, Chris, if there are any moments for questions or comments.