Podcast Live - AI On The Dairy Farm

AI is creeping into dairy whether we call it that or not, and most of the noise you hear is either hype or horror stories. So, we brought the conversation into the room with our first live event at the UK Agri-Tech Centre and asked a more useful question: what can artificial intelligence and machine learning genuinely do for a working dairy farm today, and where do the risks begin?

We’re joined by Chris Knight from Agribot, Ian Garner from Antler Bio, and Mike Jones from the UK Agri-Tech Centre. Chris pulls the curtain back on the history of AI, why the core ideas are older than most people think, and why large language models feel revolutionary mainly because we can talk to them. Ian shares how AI can speed up development and knowledge work without becoming a source of truth, including building science-grounded recommendations with expert validation to avoid confident nonsense and hallucinated references.

Mike brings it back to the coalface with tools being tested on real farms: earlier lameness detection, consistent body condition scoring, sensors that flag cows needing attention, methane monitoring, and even birdsong analysis for biodiversity benchmarking. Along the way we tackle accuracy versus consistency, predictive tools, data integration across too many apps, trust and privacy, and the awkward question of who really owns farm data.

This podcast was recorded in front of an audience, and once our guests had made their initial presentations, they then took questions directly from those in attendance, asking real questions relevant to farms today! 

This was recorded in February 2026, and all information was correct at the time of recording. 

Subscribe for more practical UK dairy conversations, share this with a mate who’s sceptical about AI, and leave us a review if you want more live panels like this.

For more information about our podcast visit www.chewinthecud.com/podcast or follow us on Instagram @chewinthecudpodcast. ChewintheCud Ltd is also on Facebook & LinkedIn. You can email us directly at podcast@chewinthecud.com

Transcript

Welcome And Live Event Context

Andrew Jones 0:14

This is the Chewinthe Cud Podcast, a podcast for the UK dairy industry, brought to you from the southwest of England and listened to around the world. Hello and welcome to Chewinthe Cud Podcast. My name's Andrew Jones, and with me today as always is Sarah Bolt. How are you doing, Sarah?

Sarah Bolt 0:41

I'm very well, thank you, Andrew. And how are you doing?

Andrew Jones 0:44

Yeah, I'm not too bad, not too bad, thank you. Um today, well, it's a little bit exciting, isn't it?

Sarah Bolt 0:50

It is, it is. It's a little bit different this one, wasn't it?

Andrew Jones 0:54

Yeah, this is I'm sure many of you hopefully many of you remember we were talking about the podcast live event that we were going to hold in February, and this is the uh outcome of the first session we held.

Sarah Bolt 1:06

It was a really well-attended uh session, wasn't it, up at the UK AgriTech Center.

Andrew Jones 1:11

Yeah, yeah, definitely. Well, and we say it, so thank you to UK AgriTech Center for hosting it. Um we'll say it again now. But yeah, and I'll be honest, you have had some nice feedback from other people in the industry since that said they'd heard it it went well, which is which is lovely to hear. So thank you for that. Um but yeah, so it's a long one because it was such a good conversation. It starts off with our guests talking and then we opened it up to our audience for questions. And there were some really good questions, weren't there?

Sarah Bolt 1:41

There were. And I think it's just one of those topics that we've all heard about, we're all perhaps a little wary about. Um, and perhaps it's hopefully it will give some people a little bit of confidence to know that they can uh perhaps give it a try.

Sponsor Message And Services

Andrew Jones 1:54

Well, I I've got to be honest with you, since uh since that I've started giving it a try with a few things. So, you know, it's uh yeah, you know, it it is it is it is an interesting message or it's an interesting conversation with people that that know it and use it uh and can see where it's gonna go. So um, yeah, really go take a listen. This podcast has been brought to you today by ChewintheCud Limited, who offer completely independent dairy and beef nutrition, co wsignals advice and training along with ROM's mobility scoring. More details on these and other services available, please visit our website www.chewinthecud.com or email us directly on nutrition@chewinthecud.com. ChewintheCud Limited now offers first aid training from a registered first aid at work trainer and experienced minor injuries practitioner. More details, please visit our website www.chewinthecud.com or email us directly on training@chewinthe cud.com. Hello, I'm Andrew Jones.

Sarah Bolt 3:10

And I'm Sarah Bolt.

Andrew Jones 3:12

And welcome to the ChewintheCud Podcast, a podcast for the UK dairy industry.

Sarah Bolt 3:16

Farmer, advisor, processor, and everyone else. We have topics and episodes that will interest you.

Andrew Jones 3:22

We discuss the practical and the technical aspects of different UK dairy industry topics.

Sarah Bolt 3:27

We aim to make you think about what you're doing and ask yourself, can it be done differently?

Andrew Jones 3:32

Listen to us speak with specialists from inside and outside the industry about their area of expertise.

Sarah Bolt 3:38

Subscribe and listen to episodes for free on your favourite podcast platform.

Andrew Jones 3:43

Sign up to our website www.chewinthecud.com for podcast notifications so you never miss an episode.

Sarah Bolt 3:50

And links to our socials, including Instagram, chewinthecudpodcast. All one word and remember, no G.

Andrew Jones 3:57

Or email us direct on podcast@chewinthecud.com.

Sarah Bolt 4:00

If you like what we do, please share and leave us a review to show your support. And that's it.

Why Talk About AI Now

Andrew Jones 4:06

Enjoy today's episode. Hello and welcome to ChewintheCud Podcast. Today, well, it's a little bit different. We've got the first of our live events.

Sarah Bolt 4:25

It's quite exciting but daunting at the same time, Andrew.

Andrew Jones 4:28

It is. It could go horribly wrong. It could go horribly wrong. So first thing we should do is uh thank UK AgriTech Centre. Um there's a nod from Mike, one of our guests for doing that, for hosting this for us, which is fantastic. So thank you to him. Uh or thank you to them, and thank you to Mike for agreeing with. But anyway, so we're here today to talk about AI. Now, for most people in the dairy industry, um, that obviously means something else. But to catch up with the rest of the real world, it's now artificial intelligence. So, should we maybe be looking at changing the term within the dairy industry? Um, usually, uh well, my time in the southern hemisphere, AI was referred to as A B, artificial breeding. So, do you think maybe we should be referring to that in the UK as well?

Sarah Bolt 5:10

Maybe it's time for a change.

AI History And AI Winters

Andrew Jones 5:11

Maybe it's time for a change, but old habits die hard because let's be honest, we all still call it purple spray. And how long has it been since it was last called the last purple spray? But anyway, that's another thing. So, um, as many people may not realize, AI is creeping into the dairy industry all the time, the new developments coming and some being tested here at the Southwest Dairy Development Centre. Uh, and most of us maybe think of the negatives of AI, that the, you know, the world's going to be taken by over by AI, dystopian future like Skynet from the Terminator. Um, but it's with us every day. So maybe it's time we start talking about it and make people think about it. And just from my own experience, I've actually been involved in the training of some uh AI. Um, but we can get on to that later. So today we have three guests, all who've joined us on the podcast before. So our first today is Chris Knight, who is um the founder of Agribot and studied AI in the early 2000s. Now, Chris was a guest uh on the podcast in December 2025, talking about Agribot, and today he'll be talking about artificial intelligence, not as new as it seems. So good morning, Chris. Good morning. Thanks for having me. You're very welcome. You're very welcome. Thank you for coming to GRIES again. Our second guest today is Ian Garner, who is head um of RD at AntLabio, and he uses AI um with his work at AntLabio. Now, Ian has been a guest in the past in September 2025, talking about epigenetics. Uh, and today is talking about harnessing AI, current uses and innovations. So good morning, Ian. How are you going? Good morning, Andrew. Thank you for having me. No, thank you for coming back again. And finally, we have Mike Jones, who is Dairy Technical Manager for the UK Agri Tech Centre, based here at the Southwest Dairy Development Centre. Uh, Mike's been a guest twice uh for us, once in May 2024, where he showed us round the site and we had a good look round. Um, and then back in May 2025 as well, um, where we talked about the year anniversary of UK Agri Tech Centre as it was then. So good morning, Mike. Good morning. So I hope this is third time lucky that I don't stumble over my words. You'll be fine, Mike. You're you're experienced, you are. So really, AI, let's start. Chris, let's start with yourself. Um, you're going to talk to us about how it's not as new as it seems. So please tell us about some of the history about AI, because most people think it's really a very modern thing, the last, I don't know, five, six, seven years.

Chris Knight 7:33

Well, yeah, that's the I when people found out I studied AI back in 2000 to put a date myself. They they seem to think I'm one of the older ones doing it, but uh thankfully I'm not. There's even older people. Um yeah, I can try I mean what modern AI hasn't really changed since 1940, to be honest with you.

Andrew Jones 7:51

What Turing?

Chris Knight 7:53

Uh well, so Turing's name's brought up, and I don't want to take anything from Turing, but there's actually key papers additional to his if you're so interested in such things. Uh William Pitts was the first one that really came who actually decided on artificial neurons. It all stems from a very uh basic understanding of how our brain works. So AI is actually influenced by how our brain works. Right, it's not trying to replicate it, it's just influenced by it. I think that's another thing people get around the wrong way. Uh but it's a very simplistic form. Um, and like all the way through to the Chat GPT and the things we see today, that's actually directly linked to those original papers back in 1940. Hasn't really changed much since then. Of course, it's grown in scale and it's grown a little bit in complexion, but the bare bones, the actual bare fundamental mass of it, hasn't really changed.

Andrew Jones 8:43

So what's that 1940s? That's hold on, 80 years, it's it's not massively changed in that time. Probably what the computing power has got faster or better.

Chris Knight 8:54

We've there's been in the AI field we tend to call them uh AI winters. So, of course, the first AI winter was in 1940. Someone had the wonderful idea of replicating the brain from this basic understanding we had. Um, but of course, computers weren't computers weren't even a thing. Never mind computing power, computers weren't even a thing back then. So how do we do that back? Wasn't really known. And then we've got more developed uh computers by 1950s and 60s, maybe, but of course they weren't up to the task. I would argue the big uh AI computing thing was solved when I I think when I the iteration I came through on it was kind of basically solved at that point. What's kicked off more recently is the data. So when I built my first uh AI neural network, uh we had picked it was to do a classification. It's be really uh complicated to do now, but it was just a simple AI that would look at a picture and decide it was male or female. But then you um it was accurate as a human, but back then we had a data set of 200 images, um, which we thought was a massive data set. Um of course it would be work then. Uh but AI was always a solution to looking for a problem, so it was never quite good enough, but we couldn't figure out why not. Fast forward to today, unless you've got millions of data samples, that's that was why. So our understanding of big data was different back then than it is now, and that's the final solution.

Andrew Jones 10:12

But what made you study it back then when it wasn't the fashionable thing?

Chris Knight 10:16

Uh it wasn't, it was just a niche thing within computer science. I mean, I started for uh to show how sound depressing I am. I actually started coding when I was eight. Um, so I I grew up in Scotland, cold dark place in the winter, not a lot else to do. Um so then I I went up, went to university. I wanted to do I didn't want to go do computer science because I'd already done coding everything up there, but I knew they'd polished off, so I wanted to do something in that area. So I thought, well, I'll do this AI thing. Uh sadly, I can't claim that I had the foresight uh as to the benefits of doing it back then. I I wish I was that smart. Uh I just lucked into it thinking I'll do this seems interesting, I'll do that. Um and then most of my software career I didn't really do AI. I was working on satellites, which is where Agrobot comes from. So we use satellites to do field analysis. So my life story beyond that was uh Agrobot's story. And then of course, uh space and defense. And to further I remember one so I'm harping on a bit, but that I remember going for my first job, I was in space and defense, and I first uh interviewed outside of that sector. They uh they were like, Oh, someone turned around to me and said, Oh, you your CV is basically the plot to Terminator.

Sarah Bolt 11:28

I would laugh, but I don't know the plot to terminator. So if there's anybody else in the room to join me, well then you're not gonna get this reference.

Chris Knight 11:35

And that's because I special I specialize in neural networks, which was the Terminator's brain. Then I thought, well, if that's just one night, why is it going for this? And then you you click, I actually worked on a satellite constellation called Skynet. So weirdly, I hadn't connected the dots at that point, and then that's when I first realized that, oh yeah. And then uh you you said, so I mean, does this AI thing work? And I was like, well, obviously not, because no one's come from the future to try and kill me yet.

Andrew Jones 12:00

So uh sorry, yeah. But but so so really you you you're saying in the last I don't know what 10, 15 years, whatever it is, it's more been the data. Now, is that again because we've got more processing power to handle more data, more um uh storage to handle more data?

Chris Knight 12:19

Yeah, I mean, uh to say it's not computing power recently is is a little bit cheeky. It kind of is like if you had that massive data set back in the when I was doing it, yeah, we couldn't have got through it. But we could have been clever about it, we could have worked our way around that computing power issue, but we just didn't have the data. Um so I I think the big iteration. So I'll I'm in the field that uh and decide that there's not been new science in the AI sector since 1980. So whilst we talk about fundamentally it's from 1940, there were improvements from all the way up to 1980. Um but really if you had the if you had our knowledge from 1980 to now, it wouldn't be a big step. There's no fundamental, oh my god, how have they figured that out in AI since 1980? That's we can go into about that, but it's not really important. Um so ever since then, for my mind, it's iter it's iterative steps.

Andrew Jones 13:09

So really, well, that's 45 years, so really the last half century, you're saying there hasn't been big fundamental improvements in AI. It is simply probably processing power, storage, the ability to hold bigger data sets and work with bigger data sets.

Chris Knight 13:26

There have been improvements. So I'm talking like new science level big steps since 1980. There have obviously been improvements and new ideas and new theories, but it's just fundamentally still based on the theories back back to then.

Sarah Bolt 13:38

I mean, from my point of view, it's sort of the general public, as it were. I think it's the fact that it's now available to us, whereas it's obviously been available to you in the industry for since the 80s in in that same sort of form, as it were. Whereas it's only really the last three, three years, or maybe I'm a bit behind times. Blondon over 50, it's probably uh, you know, it taken me a while to catch up.

Andrew Jones 14:04

Google, I forgot what it's even called, the Google one now, but Alexa and and Sari and and and all of those sorts of things that we can all use every day. Chat GTP, whatever it happens to be. Or let's be honest, the last what week or two you've had the crazes of caricatures on social media using AI. You know, I've had someone damn that because of all the energy it's taken and wasted on things like that. But yeah, but it's become more accessible. But I I say that more accessible. It's taking away, you could argue the fun things. I mean, there's isn't there a band in Sweden or something that's all AI generated? You could argue, well, where's the fun? It shouldn't be taking away the creativity, it should be taking away the boring things, and really that's kind of what we're here today, is it's using data. I mean, we talk data, I mean, here we are at the Southwest Dairy Development Centre, you've got three Lely robots, and the amount of data that comes off those Lely robots is just huge to the point I remember um when we put the rotary in in North Queensland, it was and we we went halfway house in terms of data, and they're like, Well, unless you've got the time to sit down and look at it, you're never gonna make the most of it. But I'm guessing with some of this AI, it's to take out some of that um uh routine for you to do or flick through this stuff a lot quicker for you.

Chris Knight 15:16

Yeah, I mean, AI's always been there. I mean, so you talked about maybe uh dairy should change away from using the term AI to you could argue really it's AI is the term we shouldn't be using, in all honesty. It's it's a very contentious term in of itself. And I think the reason why it's become more prevalent recently is because we've developed large language models so we can talk to it, we can interface with it. It was always there behind the scenes, but doing things like Agrobot does. So we do AI, but not ChatGPT, the more traditional ones, which is where we take all these inputs, all these different data sources to try and make one decision. Like one thing we're rubbish out of humans is we can hold three, maybe four facts into a decision. Beyond that, it gets more complicated for us. But AI can actually have all sorts of inputs into it, and it's really good at doing that. The issue is AI can be have what we call width, so that's the number of variables it can make a decision from. It has terrible depth, it doesn't really think too hard about it. And that's why I think we see whilst we think it's come on leaps and bounds recently, it's been a very slow progress to get to this point. And two, I think that this is pretty much, and most uh AI people I know work with it, um, it's going to plateau pretty quickly from here. That the I think the reason people get carried away with it is because like you said, the people it's coming to people's references now, so they think it's had this meteorite rise and it's come from nowhere. No. And two, it's got a if we go back to the point I made earlier, it's based on a very uh basic understanding of how the brain works from 1940. I mean, even to today, we barely know how the brain works. Never mind 1940, which it is based on. I mean, the evidence I can use from that is if you give me like a bowl of wheat wheat abix and uh you know uh a cup of tea, I can think and do lots of computational skills from the energy I got from that. But if you imagine the massive data sets and nuclear power plants, you need to power the AI models just to make a basic decision that shows that we're fundamentally getting something wrong from an AI perspective. So that's why it has a plateau.

Andrew Jones 17:19

I did have a quote I saw the other day, how true this is. It says a human being uses 12 watts to think, while an AI system doing the same job would need 27 billion watts to think.

Chris Knight 17:31

Yeah, so we're obvious it's obviously not where we think it is, and it's a long way off it. Um it's incredibly useful. So to come back to your point more uh focused, that's where AI is. Like again, what we do with AI is we try and uh take away the although we hope to be taking away the tedious jobs. It was interesting. You said the interesting ones. Uh I think AI is better the tedious jobs no one wants to do, the the boring, drudgery jobs.

Andrew Jones 17:54

Yeah, but it's it's being used a lot for create I know we're getting a bit distracted, but being used from creative things like the caricatures or people is creating music for people, that surely should be the human enjoyment things that we still do and let it do the drudgery, let it do the boring data sifting.

Chris Knight 18:11

Yeah, but and I again I'm really if I was to pick a sector, I'd say I'm from the software development sector. And if we're the ones constantly being told that we're the first ones to go because of AI, and there's lots of good reasons for that. Uh, we make our our work that we produce is a wonderful uh treasure trove for AI to learn from. Because we write our code and we we write it beside every line, well, not every line, but we write what the code is about to do. So we we say this is what we're gonna do, then we actually do it, and that's what it code looks like, and AI loves that. But the thing is AI is not that I've I use AI as a productivity tool. So every software developer now has a junior software developer with them. I would you can say to a junior developer, like you can say to any junior position in farming as well, go and do this job. You're crazy if you don't check in on it every so often. If you just let it run a mark for three, four days, you know how that's going to turn out. Yeah, it's the exact same thing. It's a it's a it's a junior level position right now. And that so that from a jobs perspective, that the big question is well, how do we get people into the career ladder if that entry level is taking away? That's the real difficult question to answer. But from a creativity perspective, can it do can it do art? Yes. Can I do art? Yes. Is everyone gonna buy my art? No. You know, just because they can do it doesn't mean it's of quality and it's of good standing compared to our actual professional.

Andrew Jones 19:31

Well, like you were saying that junior, we were having this discussion at home last night, really. I guess you could argue goes back to like the 80s, where there was a lot of hoo-ha that jobs were replaced by robots in terms of, let's say, welding or whatever it happens to be. What you're saying is it's almost uh the same thing and we'll evolve and adapt to to work with it as a comp as a as a companion or a uh a junior position to make us all more productive.

Chris Knight 19:54

I mean, yeah, the well, the the the cliche argument and the AI pressing would do to justify what we're doing is we'll say, well, we've been through the printing press. The printing press was a bigger impact on jobs than the AI will ever have. I would some would argue. Same with the car, the car is an Iron for the horses and carts and that kind of thing. And you know, there's endless examples of it, but we always uh, you know, something else always comes along to create jobs. Uh I think the, you know, which is a very lazy argument. I don't like using that one because it it just the the past isn't always an indication for the future, but it's something to use the mind a lot about, I suppose.

Sarah Bolt 20:31

I've got a question that I just wanted to ask you about, Chris. Um the difference between sort of artificial intelligence and machine learning, because that's often I believe that we refer to things as artificial intelligence when they're perhaps modeling or or the like. So can you perhaps tell us a little bit more about that place?

Chris Knight 20:49

Oh, it's a pretty easy one. Uh well, I'm gonna make it easy. Whether people agree with me is a different thing. But what I I I didn't finish my point earlier when I said that actually we shouldn't be using the term AI. We should the one that we should be using is machine learning. That's a much better phrase to use because it is machines learning. So um you give it lots of data and it learns. So one little key difference between a software developer and an AI developer is say I'm doing uh a model two. Say if I was to do the gender classification one again, if I make it as a software developer, I give it the rules. I tell it, look for this, look for that, and I give it, and I actually program in the rules. If I'm doing an AI model, I mean, the thing is, an AI model can be 100, 200 lines of code, and all I'm doing is programming it how to learn. So I just get say this is the data, this is how you learn from that data, off you go. Actually, how it figures out to do what it's doing, I don't know. I don't know how it does it. And that's a big research area. AI, machine learning, or neural networks, it we say it's not explainable. So it's just it's a black box decision maker. You give it a data, it makes a decision. Why it came to that decision? Who knows? All right, it's just that's what it's learned. Uh so your your question is very good one. And one that we should bring up when we talk about should dairy change from AI? No, everyone should just be using machine learning, to be honest with you.

Andrew Jones 22:08

So ML we should referring it to us.

Chris Knight 22:11

I should change the company's name for many good reasons. But yeah, it should be aggravating ML, really, to be if you're going to be purist about it and proper about it. But you know, as a tech startup and everyone gets hyped about AI, so it's as much of a purist as I am, you've got to play the game a little bit. Um and that's why everyone gets sucked into it. I think everyone who's in I think it's a good uh test to see if someone's serious about AI, it's how much they cringe when they talk about AI. Uh they do it because we know we have to, but don't necessarily like it because it is actually machine learning.

Machine Learning And Black Boxes

Andrew Jones 22:42

No, that's grand, that's grand. Well, looking at it, it's now time to move on. So we're um moving on to Ian. So um, hello again, Ian. So, Ian, um, you're here to talk to us basically about how you use it. I mean, Chris has um spoken a little bit about how he uses it with Agribot, uh, but I know you use it in your role with um Antlabio. But yeah, just talk us through basically how is AI used today on projects like Chris's yourselves or or other ones out there. What is it capable of doing or not doing?

Ian Garner 23:12

Sure. Yeah, no, I agree a lot with uh what Chris said. I think one of the the great points Chris made was is it's like having a junior developer with you at all times. So when you know, in a startup Atler bio, when we first started, there were three of us um full-time uh and you you wear many hats. So we've you know, just as we started the chat GTP, as you said, uh the first model came out there in 2022. Um sorry, 20 yeah, 2022. Um it is a long time ago, it's so much has happened in that time. Um but it allows you to speed up development. As long as you've got guardrails in place, it's it's like having someone that you can bounce back ideas with, right? I like the fact that you said guardrails in place. Yes. So it's very important again, as Chris said, you you can if you leave AI, if you leave someone doing a job and you don't check in on them, it can go and do all sorts of things. And if you don't go back and check, it's very hard then to go back and find out what it's done. So if you if you use it in a in a sensible way, like not not as a source of truth, but in a way to help bounce ideas, to to write code, as long as you know how to write code and what it's doing, then it can be a very powerful tool, just like that. Um, one example uh where we used it early on, um, you know, setting up infrastructure. You could you might need, say, a solutions architect or an AWS at cloud expert. Uh so all our infrastructure is in what's called the cloud, so it's all online on databases and stuff like that. Um we had knowledge of how to do that, but with AI, it sped it up a lot. So we were able to get online a lot quicker. Um especially with Antler, we sold EpiHeerd our product to Finnish farms to start with. I don't speak a word of Finnish, I do not I speak one word of Finnish now. Uh Moika, hello. Uh yeah, pretty much, yeah. I'm terrible at languages. Uh AI is a lot better of language, so a lot a lot better at converting language. So it allowed uh again, when there's only a few people in a company when you're bootstrapped, uh, it allows you to you know write in English what you want to talk to a farmer in Finland. You can send them an email, you can send uh you can convert your slide decks into Finnish. It's simple things like that, but it speeds up things so much where you don't have to go to, say, a Finnish-speaking salesperson, explain the science, get them to translate it all, and then go to farmers in Finland. You can take your English version, translate it with AI, check with the Finnish person that it is okay. I was gonna say, have you ever been caught out by mistranslating? Uh no, no. Um it's it's eerily was good back then. It's it's very, very good now at translating language. Um there's there's a person in Finland that I've only spoken on email to in Finnish. She has no idea I'm English or doesn't speak Finnish. Um and and they get on fine. It it it works absolutely fine. So yeah, we it's always good to check. Um you don't want to be, you know, saying horrible things to him accidentally, but no, it's uh it is very good compared to something like Google Translate, which again would have had models behind it, but it's just it's in a different league. So uh it's one example.

Sarah Bolt 26:28

Um I think I've always been quite scared to trust that, but you're saying perhaps I need to be a little more brave and uh give it a go.

Ian Garner 26:35

Yeah, give it a go. Uh if you've got other if you've got friends that speak languages that you have no idea, type something and and to them and see what they think. It's pretty good. Yeah, I I'd I'd have no hesitation using it um at all for that. But yeah, it's just an it's a small example, kind of a niche example, I guess, but it allowed that communication to break down communication. Um, where we really utilize AI right now in ANTLA, uh again, junior developer kind of role, so you can bounce back ideas. Um, but also uh again, like Chris said, machine learning. So a branch of AI. Machine learning, you can find or it can learn patterns that you couldn't possibly find just by yourself.

Andrew Jones 27:17

So I was gonna say, how do you use it today within Antler bio? What do you use it for?

Ian Garner 27:23

So one piece of work we've just finished where AI was heavily used, again, with guardrails, and and we had validation from um uh vet experts, was to when we write recommendations, um we give a lot of information, but it's got to be grounded in truth. It's got to be grounded by science. So we can't just say do this because of this without any proof that it could possibly work. So in the literature, um, there's lots of studies. You you can harvest knowledge from literature without AI, you can just do it with logical code. Um, you can then uh essentially what we did was you know looked at all the literature out there rapidly, um, pulled it together, created a knowledge base of validated literature, and then from that, we it's essentially what's called a retrieval augmented generation approach, a rat-based approach, where you uh you split all that text into basically line sentences, and then it can you use the AI model, so whether it be Chat GTP or uh LAM or whatever, whatever model you're using, um to only look at that information. So it's already kind of if you imagine like a a big table, uh AI can take any information from that table, but with the rag-based approach, it might just be a placemat on that table and can only focus on that.

Andrew Jones 28:41

So are you using it almost as a first level search to we want everything on, I don't know, looking at a few cows out here with iodine on their backs. We want some information to do with why is iodine important um for cows' health. So then you use it to search for that as the first level, and then maybe you'll go through it manually or bring someone in to go through it manually to confirm what it is.

Ian Garner 29:03

Very similar. So we would pull the literature without AI needing AI there. We can pull the literature all about iodine in dairy cows or other comparative species, right? Uh you can get all that information, then you can apply the model on top of that to kind of structure what you want to pull out of that information. So you could say, like, um show me all sources of iodine, what's the most common sources of iodine? And it will pull it out through that literature. What amounts per cow, like per kg are they giving? You can pull out that information, then you can go to your trusted vets and say, right, can you check this? Is this true or not? Then you've got vet expert validation, you've got the science that's non-hallucinic. So hallucinogenic hallucinic, um, some people don't like that term, but in AI it can make up things that sound very plausible.

Andrew Jones 29:50

So they yes, but you think hallucinations.

Ian Garner 29:52

Yeah, you think that's great, but then the references attached to that are completely made up. Um, it's a lot better now, but it's still there.

Andrew Jones 29:59

And that's I guess the thing people are worried about is you do hear these horror stories, and let's be honest, that's usually what you do here, is it is those hallucinations then AI is creating, and it completely off um off track of where it's supposed to be.

Ian Garner 30:14

And then you recommend something that's completely not grounded by science, and then you're in a lot of trouble. So we we pull out only information that's been peer-reviewed, that's been validated, so it's supported by science, anything we say, uh, and then it has the bets that cla might confirm that that's true, and then it's got the ref the real references to those papers. So it's completely it's ironclad. Like you it's very So pull out, say, Journal of Dairy Science or wherever it happens to be. And you can go back to that PubMed IE and and it will take you to the exact paper and it will pull out exact information that you've you've said.

Sarah Bolt 30:45

It really worries me that it's doing my job for me. So as a as a knowledge exchange manager at Kings Hay, um I I trawl through the literature, find the science, and then put that into well, what does that mean to the dairy farmer?

Andrew Jones 30:58

But I guess what Ian's saying is though, it does the first search for you, and then with you can then determine what's there and get it get it done quicker. Just don't just don't tell Richard that you're using it to get it done quicker.

Sarah Bolt 31:11

Just don't make me redundant, yeah.

Ian Garner 31:13

It'll always need that expert validation, right? Because it can you you want to it use it as a filtering tool to gather is it's I guess it's uh I'm I'm I'm very bad at analogies, but you know, maps. Everyone used to know how to read a map and then GPS came along, right? But you still check, you probably still check on the map that you're going to the right place.

Sarah Bolt 31:35

I'm the old-fashioned person that still has an OS map in the car. I have got it on my phone as well. I've moved one step closer.

Andrew Jones 31:45

But it but yeah, I guess we move with technology. So what you're saying now is is for yourself, it's more about processing. I mean, Chris mentioned data sets, it's more about processing large data sets quickly. I mean, if you didn't have this for argument's sake, I know you've tested the chaos here at Southwest Dairy Development Centre. If you had to make the recommendations without AI, how long would it have taken you to find that information to make those recommendations? Or how big a team of experts behind you would you have needed to make it happen? Sure.

Ian Garner 32:20

Um it's a great question. It's probably how long's a piece of string, a little bit. Yes. Um, to gather information. If you think back uh when I was at Uni, uh, you know, you'd do a six-month project on a specific area and you'd you look at all the reviews, you'd filter out the poor ones or lots of good ones, you you'd bring all that information, you get like a set of good papers, then you'd write your review uh right on it. Um we can do that in about 30 seconds with the tools. And that's with that's without AI. AI just adds that level of then contextual analysis, um, so it gives context to what all those papers mean. And then with that rag-based approach, you can really pull out exact uh sentences linked to the references, like exactly what you want. So it speeds it up, orders and magnitude from that. In terms of experts, you still would have the same bets. You still have you need bets to say that this is they're happy with that. Um, because that's what they're trained in, right? Yeah, you you could you could do it without them, but then you risk you you're not validating what you're pulling out, essentially. Um, in terms of data sets, you know, as as again, you can do a lot without machine learning when you've got uh yeah, we look at 27,000 genes. Uh that's a lot of features. Um, you can do a lot of analysis without any ML on that. But as you start building up your data sets, as you get a number of farms coming into that, you've got millions of feature points, right? And you've got all the phenotypic data, you've got all the metadata associated. You you just can't find patterns yourself. You need learning models to then sift through all that data. And that's where now, again, there's more, there's better models being built all the time. Um, the power, the computing, the storage, the accessibility, like it it's a lot better. So the infrastructure behind it.

Andrew Jones 34:08

Yeah, it it can do that data sorting. Let's be honest. Who likes sit who likes sitting there looking at Excel spreadsheets that just go on and on and on and on? You're pulling faces as I say that. Uh, and as you say, you you I know you've mentioned it before that for a startup business that AMTLA was, it provided so much um support that maybe financially wouldn't have been viable uh in the past. Uh and it's allowed you to get to market quicker, to deal with people quicker, and it it's basically looking through and not just one level of an Excel spreadsheet, but a whole 3D model, multiple layers, it's pulling out the data for you quicker and allowing you to respond to, for I mean sake, your uh epigenetic tests to then make um recommendations on farmers to what you should and shouldn't do.

Ian Garner 35:00

Exactly. Yep, exactly. And that's and that's how we've used it. So we haven't relied on it, it's just been there to help in certain areas when it's needed. We've very, very like um the guardrails are in place. We don't let it go off on one. Um we we we make sure if the guardrails weren't there, what would happen? Well, then you could end up with uh hallucinogenic re references, you could you could end up with recommendations that complete like make no sense at all or completely against what you should be. If you didn't have the bets then to validate it, you could be in very dangerous zone, right? You could recommend something, it could have a massive detriment effect. If you don't have the supporting science behind it, or you don't have the bets behind it, it's it's that kind of combination.

Andrew Jones 35:40

Yeah, you you throwing that, yeah, your background isn't chaos, it's obviously the epigenetics and other things, but you need to it's the same with coding.

Ian Garner 35:47

You could end up if you just purely relied on it to code for you with no knowledge of coding, you don't know what it's done. Yeah, right? You don't know what tests it's run, you don't know how it's run that data. It might have completely missed things out.

Andrew Jones 35:58

Uh and then you're not know you're getting adverts for software that will code for you and build apps for you. But what you're saying is who knows what it's actually doing in the background?

Ian Garner 36:08

It's a black box, as as Chris said. And and if you you know, you can put the guardrails in place and it can work for you. You you know, we we shouldn't turn a blind eye to it. Like AI as is now is a powerful tool. If you're using it well correctly, it's awesome, it's great. Um if you if you don't know what you're doing, or you think it can do more than it can, it's a random guessing machine and it's not so good.

Andrew Jones 36:35

Using Sarah's usual phrase is what does this actually mean on farm? Is there is there a way farmers can use that themselves today, or is that something that the likes yourselves or Agribot or whoever the business is are going to use to build, bring functionality to a product on farm, or is there something people on farm can do today to use? I don't know, ChatGTP, I mess we're all used to. I know I used it for the first time to write the invites for this. It sounded a little bit better than what I would say. You know, are are there maybe I'm thinking more the data, are there ways people can use that themselves on farm today?

Ian Garner 37:17

There are, you can upload files to chat GTP. Um you've got to be careful about that data, you know, how it's used. Um they will use if you don't specifically opt out, they might use your data into future modeling, um, which is not necessarily a bad thing, but it's just something to be wary about. Um if you've if you've got a base understanding of what you want to, if you've got data, then you can use you can use those tools, but how reliable that will be, it's not.

Andrew Jones 37:46

You've got to understand what your limits are that you're putting on it to or where your focus is, rather than just going randomly off so far and say, you might want to look at a pattern as I'm just thinking lameness off the top of my head, why particular animals are getting lame or look through the data to filter out is there a pattern of where it's happening, when it's happening at certain times of year, or something like that.

Ian Garner 38:08

If you've you've got very well structured data, um, you might be able to pull out something like that. It's it's the nuances of say, if you say you asked, is has my milk yield improved since last year? Well, you can you can see that without AI from your data. But understanding why it may or may not have improved, that's where it's not gonna be able to help you so much. Well, most of these things are multifactorial, aren't they? So yeah. And without you know, without that data, uh you need lots of data sources really to provide for that model to then say well, your yield has improved by X percent because of edge ride. Whether it's you've improved the water, you've improved the minerals, or whatever it happens to be. So you need that data bank. So you wouldn't just be on your data as such. But yeah, you you could you could use it to a point um how accurate, how viable it would be is very dependent.

On Farm Tools And Real Trials

Andrew Jones 38:59

Okay, well now Sarah's do doing the timekeeping as she's been asked, don't worry, I was aware. So now it's over to Mike. Now, obviously, Mike, I mean, part of the background for the whole idea for this podcast is obviously some of the conversations you and I have had with different projects I know have come through here. The one that immediately springs to mind when we first toured here two years ago would have been HoofCount and how they were using um machine learning to pick up um digital dermatitis a lot quicker than we could. And I know myself I've done a little bit of the work for HoofCount since then that's all going into that machine learning. Um, but also obviously I've heard Chris speak here. I know the bio have been involved here. I believe you've got is it chirp AI? Is that the other one? Yes, yes, we got that. This is all coming through. So you're here, the the coalface actually um uh seeing these products, testing these products before they've gone out to the general public or the general population farming population. What's your experience with them? Where do you see them, or how what projects have you seen that have been successful? Maybe some that haven't been successful, you haven't told us about, but you're you're seeing this stuff here. How is it actually practically applying on a working farm today?

Mike Jones 40:12

Um, so it's the acceleration of how a a project, a product is developed that is really emphasized to me how machine learning is what we should be calling it now, um has has helped uh everything go forwards. You know, uh Hoof count with their petty view system. I don't know how it has managed to interpret every cow's foot and say whether that's a healthy foot or a foot with a uh digital dermatitis leisure.

Andrew Jones 40:44

Well the fact it's picking up two stages quicker than probably you or I could pick it up is just uh phenomenal.

Mike Jones 40:49

Yeah, and that's it, and it's every it's it's human nature to be skeptical of something new. That as confidence grows within you on a product um and you you become uh more confident in what you're seeing is actually happening there and then. You know, Sarah said something earlier on about you know, it's now available to everyone on your phone laptop. And yes, you know, I've been using you know Copilot and Chat GPT, and yes, I was skeptical at the start of it.

Andrew Jones 41:28

Well, I'm a bit of a Luddite with it still. As I say, I use Chat GP for here for the first time.

Mike Jones 41:33

When you know, as Ian says, you've got to give it boundaries and look after it, and don't take everything that it turns out as gospel, you've still got to read it to make sure, and you know, my confidence has grown in in using that at a at a you know farm management level. But with the projects we've got here, you know, um what goes on in the background is completely unknown to me. We've got two fantastic experts here today. I've learned an awful lot in the last quarter of an hour, um, 20 minutes on what these guys have been talking about and how it has developed. Um so yeah, you know, we gotta embrace it as we go forward.

Andrew Jones 42:17

I mean, we've mentioned right, so what have you seen from the different projects that you've seen here on farm?

Mike Jones 42:24

Um think about this one. The herd vision has gone from just a camera taking body images scans from above a cow, you know, it's now interpreting 3D, two-day um infrared images and giving a very accurate score. And it is about having the ability to crunch all that data very fast. Yeah. You know, we can all go out and body condition score, you know, six cows in front of us, and we'll all have different answers. As long as you're scoring the same in your own head, every cow, it's all right.

Andrew Jones 43:08

Well, yeah, I've often used the argument if you take body condition scoring, you almost just want too thin, too fat, just right, because the just right is you and I could do it, and we'll be quarter, half a point. And some of the figures that you'd mentioned herd vision there, we couldn't, was it 0.14 of a um a drop or you know, we couldn't physically score that ourselves. It would be impossible.

Mike Jones 43:29

No, and and that's it. And you know, having a machine that is able to deliver that accurate of detail makes the herd manager's file manager's job easier in making those informed decisions going forward.

Andrew Jones 43:46

Because again, it's dri it's sifting through that information, presenting you that information without you having to physically, for argument's sake, go through that Excel spreadsheet after somebody's condition scored them for you, or whatever it happens to be.

Mike Jones 43:57

You know, no, I'm not a person. Person, I can use a spreadsheet, but not to how you know someone who's really trained in spreadsheets can use it. And to go through a list of cows on a spreadsheet and just pick them out, it'd take me, you know, a couple of hours to do that. Some of this equipment that we've got here, it does it instantly. And it's on your phone, it's factual, and it's only flagging up the cows that require attention. You know, it's able to filter out with parameters, with the guidelines on right, we need to go and look at number four, number six, seven, eight, and nine and a hundred. We're not too worried about them today. They're they're fine in that parameter.

Sarah Bolt 44:40

But the time saved in all of that, it's not just the time you looking at the data, it's the time of actually collecting that data of body condition scoring. You know, doing that for a whole herd, it's two or three hours each time, isn't it?

Andrew Jones 44:53

It's putting me out of one, you know. I condition score for a couple of clients, not a lot, but I do do. But you say, um, yeah, I'm thinking one got robot herd by the time you go around and see them all, it's probably two, three hours for a similar size herd. And this time actually gone and found them all and everything. So yeah, it's saving that time. And then then I usually type it up for them and then put some basic, oh, if it's above, I don't know, 3.25 and below maybe 2.25 or something like that, I'll highlight them. I can't remember what character um uh figures I use exactly. But yeah, you're doing that on a daily basis.

Sarah Bolt 45:26

There's there's no farmer that has time to be able to do that on a daily basis, or even I'm guessing it's even more than a daily basis every time that cow walks under that camera.

Mike Jones 45:36

Yeah, you know, the stockman eyes will look over every cow that he sits walks past every day. As herds get bigger, he hasn't got time to look at everything. And I've learned, you know, I was a Leadite when I first started using activity monitors 20 years ago, and he thought, yeah, yeah, yeah, the salesman sold his good story here, won't believe this. Well, you used to sell them yourself. So when I first met you, and you know, within six weeks, I was thinking, hang on a minute, we're getting cows and calf that we are not seeing actually showing signs of Easter. So I learned to embrace technology because it made me a better stockman. And I think that's where you know a lot of these uh technologies, um, they're still under development, they're really practical now, they've got a return on investment, so they will be successful in the marketplace.

Andrew Jones 46:32

But you're saying that it always makes me think back to a farmer um relatively local to me. He made a comment uh Southwest Dairy event a couple of years ago, same sort of age as me. And he said, you know, this technology is great now. And he says he said, It's not like you or I, been around cows all our life. Usually we started with smaller herds, whereas now the youngsters coming in, he said, Um, we're coming into bigger herds, don't have that experience, and it's just another tool. It helps them, I don't know, monitor those calves um in terms of are they eating when they're supposed to, if you've got uh you know a robot um calf feeder, or or temperatures, or like the cows, are they coming up when they're supposed to, and all these sorts of things. Is it's like like both Chris and uh Ian have said, it's like a junior um apprentice then on farm. It's doing all that running around for you, doing those little jobs for you, to then sift that information and go, well, actually, um this calf hasn't come up and fed when it's supposed to set an alarm off, or the condition score has been left. And in bigger herds, it's it's easier to miss um some of those animals, maybe.

Mike Jones 47:45

Yeah, you know, it definitely is. And and this is where you do have a friend in your pocket with you, you know, with all the technology you've got. Everything's now is on on a f on a phone. And you know, you you when I've you know been using Copilot, it does treat you as a friend, it tries to buy you in. Um and you know, so you do have to, you know, rein it back in every now and again. Um but you know, we are getting the stockman has to look after more and more animals and technology, artificial intelligence, and machine learning, they're making everything more practical, more and more efficient.

Andrew Jones 48:32

I'd say possibly more consistent as well, because like we're talking about condition scoring, but I know, like I say, I mentioned I I've um mobility scored a herd for hoof count um since I uh you know been up here and obviously talked to Anthony and whatever. Uh and I know off the back of that the every one thing that I'd put down as a two or a three, it was a 600-cow herd. Um, everything that was put down as a two or a three, they were then going to find that eight-second video and pick it up to then teach the machine, machine learn that I is a mobility score. And it wasn't just me, others were saying that that was a lame cow to ultimately teach it that that was a lame cow, and then it could mobility score literally every time those cows walk through the hoof count and go, well, this cow's lame, flag up uh an alarm, this cow's lame, probably before maybe you'd realize yourself brings more consistency to it because again, like the body condition scoring, there might be an argument: well, is that a one or is that a two on the mobility scoring? Is she you know, is she coming like, you know, it will bring a consistency to it, not just oh, you're having a good day or a bad day in terms of how you refer to that uh that score.

Mike Jones 49:44

Yeah, you know, consistency is what we're seeing now with everything going forward. And, you know, I think we're we're only really touching on where machine learning, artificial intelligence is is taking us. You know, we've had two fantastic speakers here this morning and you know, with the Agribot and um Antlabio on what their companies are doing and embracing it and taking the decision process on a lot faster to become more efficient in the farmer's job. You know, but you know, we've got the likes of um Remedy uh just coming out. Um I'm still trying to get him ahead around it, but being able to predict what cows are going to be more susceptible to Yonis in the future.

Andrew Jones 50:30

So Remedy, do you want to tell us a little bit more about that?

Mike Jones 50:32

So it's um it's a new uh tool being developed by Andrew Bradley and Nottingham University um looking at Yonis scores. And I'm still trying to get my head around it, but it's able to predict what uh cows will succumb to Yones in the future with a percentage of, you know, whether it is gonna go or not. And this is where, from my point of view, machine learning will take us in when uh data sets can be joined up onto one platform and then the decision making process of predicting the future becomes more co I'd be I'd be more confident in where that's going. You know, like you know, Agribot, you're able to hopefully predict crop yield of you know, grass, what's going to be in the field for the cows in ten days' time. Predicting the future using machine learning, I think that's you know, where we're just really touching on the surface of where we're going with all this.

Sarah Bolt 51:39

So, Mike, you've touched on a couple of um things that you've got here at the dairy centre that use um machine learning or AI. Can you just perhaps talk a little bit more about a few more of them? So you mentioned chirrup. Can you tell us what it is and and what it does?

Mike Jones 51:55

So, yeah, so chirrup, that's part of our biodiversity projects. Um so it's uh a device that sits outside in the hedgerow listening for bird song. Um Is it insects as well or just birds? That's just birds. Uh there is there are other devices, other companies that are looking for uh insect noise. I recently came across one that you put in the ground and listens to the gr noises in the ground of invertebrates and things, telling you how healthy your soil is. So, you know, chirp it will pick up bird song, but it's then able to identify what bird it is and the number of times it picks up that particular bird. It gives you, you know, a benchmark on you know how many different types of birds you've got with different bird species. You know, we can all look out the window and we can see the odd crow and there'll be some sparrows flying around and possibly the odd starling as we're near the Somerset levels. But when we've actually looked at the data, it's come back with over 30 different species of birds.

Andrew Jones 53:03

I suppose historically you probably need an ecologist to sit there for a month or two to get that. Whereas now, let's be honest, some milk contracts are now being driven by what what level of birds you've got, or whatever it happens to be, this is doing the job for you, I guess. Uh well it is, and it's doing it more accurately.

Mike Jones 53:20

You know, I like looking at uh birds, you know, I've got a bird table at home and we get a lot of chaffinches and goldfinches. But the cherrup device that we've got here at work when we've uh uh put it out in the fields, it's picked up gold crests, wheat ears, all sorts of other birds. Not going to enhance the farmer's bottom line at this moment in time, but it's giving us a benchmark of where we are in the moment in time, which you know we couldn't do beforehand.

Sarah Bolt 53:54

And how about any anything else that you've got here that uses AI that you can share with us?

Mike Jones 53:58

Uh so we've got a Miracle methane laser that's detecting methane emissions being blown out of this barn from the 200 cow herd here. So, you know, that that is bringing in weather patterns, um, detecting a specialized you know, gas um and interpreting how much methane is being blown out of this shed. We're averaging, so they tell me, two kilograms of methane per hour, which is you know twenty-four kilos a day A. Is that good or bad? Do we know? Well, and that's and we don't know, do we?

Sarah Bolt 54:37

So it's it's more data.

Mike Jones 54:39

Data, and we can now say, well, this is where we are today. It's a starting point. What I'd love to be able to do is put methane detectors inside the building to get a more accurate individual cow score. Um that it's getting everything in place and then the data joined up um to take it forward. So then we've got to be confident in what the data's producing, you know.

Andrew Jones 55:08

But again, I suppose you could say that's where the AI machine learning comes in is projects like that, if they've been done in the past, would have then taken a year or two probably to sift through all the data to give you that data, as this is probably giving it to you in real time.

Mike Jones 55:20

Uh and that's uh and real time is one of the important things. We all want to know what's happening now, not what happened six months ago. Six months ago, yes, we can learn from historic data, but we have to react to things happening now.

Andrew Jones 55:38

Well, Sarah's now going to run around with Mike. Does anybody have a question they particularly like to start with?

Sarah Bolt 55:43

Or here's my question for for all of you guys. Um, if we held this same event in five years' time, what do you think would be normal on a dairy farm that still sounds really advanced today from a point of view of AI?

Ian Garner 56:00

Uh I think um what Mike spoke about remedy with uh you know detecting onis uh before it happens or early, I think that's exactly where Antler positioned as well, but with gene expression. Um so you know, you could say like if you can predict something happening before it happens, you'll save a lot of money there. And you do need that past data to be able to build that model.

Andrew Jones 56:21

But just made me think, what level of predictability or probability, maybe is the question, is acceptable in something like that? Does it have to be 80%? I mean, I guess it's a bit like I'm just thinking of the genes. Was it was it parent average was less than 50%? Was it? I can't remember what it was. And now with genomic data it's pushed up and up and up. So uh what level in AI in general, what level of probability is considered acceptable that that's where you're expecting it? Well, Chris, I'll we'll bring you in on this one when Ian's done. But I mean, you obviously must use that for predicting what your grass growth is going to be in the next 10 days, fortnight, or whatever is there's got to be a level of probability. At what level do you go, this is when that level of probability is acceptable, that's what we think will happen?

Ian Garner 57:08

Well, for us, it's a difficult one to answer exactly because I guess each different thing will require maybe a different probability. Um if something's impossible to predict right now, but in the future you could get 70% accuracy, that's that's a big improvement, right? Um if you can get 99 as close to one, then you've got the perfect predicting model. Um that's what you could argue.

Andrew Jones 57:32

Some people are gonna say that's always gonna be your uh argument, isn't it? Oh, well, it's only 80% predic probable, then why should I therefore believe it?

Ian Garner 57:40

Well, it's it's 80% if it if it's eight take 80% as an example, 80% of the time you get it right, eight in ten cows you stop them developing yonies, or you stop them developing mastitis, that's a huge benefit, right? If you can predict that early on.

Sarah Bolt 57:55

I guess with each farm, that is gonna be slightly different because if yonis is a huge problem on a f on a particular farm, you don't want to be culling absolutely everything, but actually you're then gonna say, actually, I need a higher percentage of probability for that farm than actually if I've got a farm where I, you know, incidence of yonis or mastitis or whatever is way lower. Actually, I'd take a lower probability.

Chris Knight 58:22

Well, I accuracy for me is always an interesting topic. We've discussed it a few times, so I'm gonna repeat myself for you sadly. Um, but like as mentioned, we we do we use satellites to uh measure grass growth, how it is today, trying to save mate from I don't know if he enjoys it or not, but hopefully we're saving you from the keeps the spat stick counter.

Andrew Jones 58:41

Well, as I say, I was used to well, I used to uh plate walk with someone, I always used to say I'm getting paid to do my weekly exercise. So, you know.

Sarah Bolt 58:48

It's fine on a dry day.

Chris Knight 58:50

Well, from that from that side, we're we're we're an extra opinion on it, at least. Um but I think there's a few things about it. First one is that's as someone who's talking about consistency, that's exactly what we try and do from a grass growth perspective. Because uh if you go out, if you go out there or a plate measure reading, you can do it three times a day and probably get five different uh values. Um and that's just the same person, or maybe it's multiple. And so with aggro, we try and measure grassbit consistency hostily. Um we just try and always be the same way of measuring every time, which brings us to the accuracy question. The thing about accuracy from a grass growth perspective, and in general, is it percentage of what? So one re one reason, one pushback I got from early is there's lots of academic studies trying to use satellites to measure grass growth uh before, but they were never quite accurate enough.

Andrew Jones 59:38

Well, I remember in our podcast I mentioned is it like Southern Hemisphere have had sort of um grass growth from satellite for what, 10 years or whatever, and I've always sat there thinking well, why haven't we got it here? And then obviously you explain cloud cover and whatever. But yes, it it's not something new, you could argue, in the wider industry.

Chris Knight 59:55

Right. But for I suppose we've taken a different view, is the reason why it was never deemed as very accurate was because the ground truth was plate measure readings, so they were which has its own problems. So you get they'd be trying to correlate to put the salad, because satellites don't measure grass, they measure what they measure, and we turn that into grass growth.

Andrew Jones 1:00:12

Um yeah, it used to be chlorophyll density, I think the southern hemisphere used, but not here it doesn't work, doesn't it?

Chris Knight 1:00:17

So that's NDVI. Uh you're but of course that's the fact by cloud cover, whereas we use radar and that's a different thing. We can go into what I was trying to avoid it. Uh the the the issue there was they were then trying to correlate what the sat what the satellite measures with grass growth, but that grass growth was plate measure readings, which is fundamentally wrong as well. But from a from a sat from a science or AI perspective, that was the ground truth. So therefore it was never accurate because while you're trying to your accuracy is measured against something that isn't accurate, therefore, you get this whole debacle. So the way we treat um uh Agrobot is we we don't talk about accuracy, we sort of take a view of math scheme, we sort of use Bayesian, so we try to understand the error of what we're doing rather than the accuracy. We know roughly how wrong plate measure readings are, we know roughly how long how wrong the satellites are, we know roughly how wrong everything is, and we come at it from that perspective. But even then, we'll come back to the whole argument as we're just trying to be consistent. So even if we we know how wrong we are, but we know how consistent we are.

Andrew Jones 1:01:17

Oh, you're consistently wrong.

Chris Knight 1:01:21

Well, that what's the phrase? It's just like uh all models are wrong, but some are useful. Um that's the good AI phrase to use. But like Mike touched on earlier, you can get different people looking at different cows, right? And then you can get different opinion, which is great if you've got the same person out there every day doing the same thing, but that never happens. So if you get someone else that's to do it one day, well, that whole thing's ruined. So I would argue consistency, of course, to a point, but consistency is more important than accuracy because you'll big that in. Because if you're if you're wrong by the same amount every time, that gets shadowed by the fact, well, you just react to it the same way because your act your action is consistent to what you're measuring. So whether it's puristically correct. But then to go into a bit more, we often then with the long-range weather, we'll measure our accuracy on perfect forecasts. So with the our grass growth models are about 95% accurate four weeks in advance, but that's assuming perfect forecasting, which is no, there's no thing on that. We don't do weather forecasts four weeks, we do atmospheric forecasts, which is a whole different thing. It's a lot more accurate.

Andrew Jones 1:02:26

Most people, my generation, would probably say, oh, the weather forecasting's got worse, not better, just uh despite the fact we've got all these supercomputers in there now.

Chris Knight 1:02:33

I would agree with them, but I'm not a weather man, I'm an atmospheric forecast. I'm an atmospheric forecaster, not a weather forecaster, very different things. Uh and video to very different ways. Um, so again, with the accuracy, we yeah, we'll sense that we're 95% accurate four weeks out with perfect uh conditions, uh perfect forecasting. So even I I don't know, I I tried I battle with them myself talking about accuracy because I do like to keep things simple. But then whenever you talk about accuracy, my little brain goes on it's a little bit false advertising if you just give a percentage because there's all sorts of caveats with that. And then you know I struggle with how do we convey that from a message perspective, because someone just wants a nice number, but then it's not quite fair and accurate either in anyone. Um, so yeah, consistency for us and error is more important than accuracy.

Andrew Jones 1:03:19

But I guess, Mike, you're poised. You're you're dealing with this all the time. What do you see happening in five years that at the moment seems a bit well science fiction, I guess, because you you are see seeing this stuff like the methane monitors and those sorts of things, you know. Are we gonna see more of that moving forward?

Mike Jones 1:03:36

We're still gonna be milking cows in five years' time. We'll still have a lot of livestock around, which is fantastic. Um, we're gonna have healthier livestock. We're probably gonna have cows that produce less methane per litre of milk produced. Um you know, with uh Antlabio, the confidence that I will have in in their um uh process will we'll gain traction. Um being able to take data and predict what that cow is gonna do, and by change some of the interventions that go with her, we that that that cow will become more efficient. And it it is about you know taking everything to the next level. You know, every cow here has got a Smaxtex bolus in, and we know how much water she's drinking, what her body temperature is, and when parameters for those individual cows reach certain uh areas, it will flag her up as a cow that needs attention. And we're gonna have more stuff, more technology joining up with other technology on single platforms where they're able to predict what that cow is gonna do not just tomorrow, but in a week's time, in two weeks' time, and she will be a healthier, more profitable cow.

Sarah Bolt 1:05:08

And I think it's that health and welfare thing that is so important that all this technology, whether that's machine learning, AI technology that's behind it, I think we're going to have that healthier, higher welfare standard animals on our farms.

Andrew Jones 1:05:23

Certainly second that again. I think it I think it's a healthier, like we mentioned earlier, whether it's the body condition scoring, it's gonna be it's it's a it's that junior to pick up the um problems quicker, whether that's body condition or lameness or not drinking or not eating as they should do. This stuff will just pick it all up quicker. I mean, I remember a few people, um, some of the early was it ear tags that monitored uh calf temperature, and that was just like, oh, well, it's telling us the calf's sick before the calf even knows it's sick, and it's just going to be more of that, isn't it?

Chris Knight 1:05:56

Well, I mean, I think the freezer we're trying to do is we're trying to, in our Little corner make farming or dairy proactive rather than reactive. Like with weather's limit, if your models are based on three days, you're reactive. And there's all sorts of problems with that. But if you give two weeks, three weeks heads up, there's a lot you can do that, and a lot you can prepare and make things more efficient.

Andrew Jones 1:06:16

Take the summer we had last year. If you knew that was coming, you can predict, you could you could plan earlier in terms of where your forages are or what your feed budgets need to be or whatever it happens to do.

Joining Up Data Across Platforms

Sarah Bolt 1:06:27

So we've got a question from our audience, not quite confident enough to read it out themselves. So I'm going to do that for them. So have we got an idea as to sort of how many farms are really using AI or machine learning already and to what extent? And then sort of a follow-on question from that.

Chris Knight 1:06:49

I can do the second one. I don't know the first one. But I I think it kind of touches on what do you what do you think in five years? I don't know five years, but I know what I would like to do in five years, and that's less apps in the pockets of people. So a more uh relevant. Yeah. Consistent again, unified and in one place, because all these bits of data are great, are useful separately, but if they were collected in the same area, that they need to talk to each other.

Sarah Bolt 1:07:16

How do we get this cooperation within a commercial industry? That's my big question.

Chris Knight 1:07:21

Well, I'd I'd like to think we're we're we're the good chaps on that one. We're all all our data is available for any other company to integrate with. In fact, we're trying our hardest not to make another farm management app. We really don't want to make another farm management app, we want to integrate into others. But um that depends on other people playing nicely.

Sarah Bolt 1:07:40

That sounds like the the way to go, definitely.

Andrew Jones 1:07:43

Come on, Mike, you deal with all these different people. How do you make them all talk to each other and and and and make something happen together? Not putting you under pressure, of course. That is a six million dollar question.

Mike Jones 1:07:54

Um companies are very data is a valuable commodity and uh all has a price on how it can be used. Uh subconsciously, I think everybody's already using AI to machine learning. You know, we don't realise it. And you know, I I did go to a a meeting uh last year where the main speaker said AI is the new electricity. And I sat and thought about that for a while and I had to agree with it because when electricity was first brought into our world, it it was oh that's a bit scary. But how many things could not happen without electricity today? None of us would be here in this room if we didn't have electricity. And going forward anybody that doesn't embrace artificial intelligence and machine learning will fall by the wayside.

Andrew Jones 1:08:55

You're you're probably right, we are subconsciously all using it, whether we do or don't. I mean, let's be honest, since the beginning of season four I've now started to use AI to do the description for the the podcast, because it's just easier than me sitting there and it does a lot better job. Admittedly, it's longer and probably doesn't have a particular slant in its head that I'm thinking of. So we are all slowly creeping in there, whether we like it or not.

Chris Knight 1:09:17

I think it's I think it's a difficult way to think with AI. So how do I use AI? I think the equivalent is I've just bought a new chainsaw, what can I use it for? It's probably a dangerous line of thoughts to go down, you know. So you you should be thinking more about problems. I I think if you know you're using AI, we I try our hardest to sell what we do without using the word AI, to be honest with you. It's like we do grass growth.

Andrew Jones 1:09:41

It's the tool for grass growth, it just happens to have AI in the background.

Chris Knight 1:09:44

Right. You know, being my background, I a number of times I get people calling me, Chris, how do we use AI? What do we use AI for? I said, Well, that's you know, that's you're like you're you're you've started the journey of a bad decision there.

Andrew Jones 1:09:56

Uh yeah, it just happens to be part of the tool rather than the other way around.

Chris Knight 1:10:01

What problem do you have? And then we can talk about maybe if AI is the right thing, right? Rather than uh, you know, again, it's like asking a plumber to turn up and saying fix this using this hammer.

Ben Richards 1:10:12

Yeah, uh Ben Richards, this is probably better for Chris, I would have thought. Um, Albert Einstein once said uh technology will one day breed a generation of idiots. Um, I personally feel he's 100% correct um with the existing generation. How do you feel going forward? Is AI going to continue to to uh degrade the intelligence of the population or is it gonna actually improve the intelligence of the population?

Chris Knight 1:10:37

Uh well I I think it's some uh difficult. I I I agree with I agree with the question, first of all. Is it AI just going to do it? Potentially. I mean, uh I've got down my notes. The the the difference is so we use AI in different ways. I use Chat GPT and I use this to do my software development, right? But when you talk about Agrobot, we don't use Chat GPT. We have our own AI models that do different things. But I see myself doing it. Like the number of times I've been trying to fix a problem, trying to bully AI into telling me it's the solution. And really, I you know, I spent four or five hours trying to get AI to fix it for me when really it was taking me five minutes. Yeah, right. So I've caught myself in that trap. So but that's what it is. It's a trap. But then uh we talked about guardrails earlier. Like we need guardrails around how we use AI as well.

Andrew Jones 1:11:26

It's just so easy to know not that some people need guardrails.

Chris Knight 1:11:29

Yeah, well, it's just so easy to get yes, to rely on something and too much. Um, but I think there's other traps. Like if you're an employer, you can hire juniors and get trapped trusting them in too much as well. So I don't think AI is a special case if you look in other ones. Um, but yeah, it's a good point and a good question, potentially. I I I'm not gonna disagree with you, but I think I would push back on my earlier point and uh AI, you'll notice that AI probably, and I'm going, I don't do a future gazing, but I'm pretty confident enough to say that AI is gonna plateau for quite a while. Um, and you can see from a startup perspective, there's a massive bubble. There always is in these new techs. There's a there's always a bubble with new tech. AI is, I think, is particularly overgrown financially, but also expectation-wise. I think there's a massive financial correction coming because of all these oversaturated valuations that AI companies get. I do VC due diligence. And the one reason that the for uh VCs, I've venture capitalists, so these investors. And part of my job is just to go and prove that that company isn't AI, so they don't have to pay as much for the company, which isn't as hard as you think. It's actually quite easy to be honest with you, with that. So yeah, I think AI is overhyped financially and what it's going to do. But over a 50-year period, I don't know, maybe. But not it's not an immediate concern. I I I, as a software developer, constantly being told I'm being replaced, I don't see it and I don't believe it. And I think I'll be replaced before anyone else in this room.

Andrew Jones 1:12:58

And so I'm I'm supposed to another example we're talking about AI, I suppose, is the internet. I mean, I remember hearing about the internet when I was what, probably my kid's age, it was one that it was just that thing, and you email and you're like, whoa, wow. And now, I mean, I remember probably 10 years ago, it was sitting down, my nephew and his cousins and going, I remember playing with no internet, and they're like, What? No internet, how did you do it? Or when my youngest was born, he would the oldest was given the phone, and because we were out of from home, we turned off the Wi-Fi, and his first reaction, there's his brother, huh? And it was my Netflix, my Netflix doesn't work because he was more interested in that. And yet, like Mike was saying about AI is a new electricity, almost the internet's a new electricity, isn't it? Really, it's become a stage if you haven't got um uh the internet, then you almost slipped to second class.

Chris Knight 1:13:45

Well, but you could uh you could view it, you could view AI in a different way. So everyone's smarter and a little bit dumber because of the internet, right? We can say it has a duality about it. AI is just the same thing. AI is just Google on steroids in a certain way. Whilst all the information is on the internet, good luck finding it. But AI is a shortcut way. So Google was particularly worried about AI because that makes Google.com redundant now. Like Google's use case has gone skyrocketing down because everyone just goes to ChatGPT as the new search engine or the the way I get the truth from. So uh I think I think it's just that in that from that sense, from that a large language model, Chat GPT and whatnot, it's just replacing Google and the search engines. And I think it'll make and it'll be an extension of you'll be smarter and dumber.

Trust Privacy And Corporate Limits

Jim Juby 1:14:31

Jim Juby. Um, I've got a question on trust. Um, we're part of a corporation and we're not allowed to use a lot of the um AI that's public domain um by our company, which for for various reasons. How do you get over that trust thing with I mean, a lot of farmers, I mean Ben clearly had a bit of cynicism towards it. How do you get past that trust thing where you're getting, you know, you're starting off with raw data, you're analysing it, you're coming up with a solution, and the middle is a black hole to most people. Um, if you're making a thought pattern yourself, you tend to work through a process, you understand that process, but obviously, you know, you're expecting not expecting, but you're getting an end, an end result. How do you get past that stage where you just ex uh people will accept that? Is that a problem that you come up against or you know?

Chris Knight 1:15:25

I'm happy to answer a bit.

Ian Garner 1:15:26

I'll I'll I'll I'll give a little answer. Um, so I think I think it's a it's a very important point. Um I think in a a lot of the models you can turn off whether your data's used in that model. So that's one good thing, but then you're still never, you know, you're never sure, right? Um I think with you know the data we handle, we never put it into AI. We we have our own servers, we run our own machine learning model, so it never goes outside of where we tell it it can go. Um but I think uh it's an it's an important point. Uh you know, trust it, you need it needs to be built. You don't want your own personal details getting out there. Um But I think as long as you're responsible with that data, there's lots of GDPR, right? You you don't you wouldn't want to just give away raw data into a language model and not know what would happen with it. You don't want other people analysing that. You want to trust the company that you've given that data to to look after it securely.

Andrew Jones 1:16:21

So it's that is something, is it the one that music speaks to mind? Is it Scarlett Johansson was asked to be the voice for an AI, he refused, and now she's taken them to court because basically they've gone and copied their voice using it.

Ian Garner 1:16:30

So it's the same with Matthew McConaughey, right? He's patented his voice now or trademarked his voice, so he can't um you know, if someone does use it, at least he earns money from it. Um maybe I should be doing that a bit quick. But yeah, I'll I'll let Chris uh answer.

Chris Knight 1:16:45

No, I mean I I'm actually quite liberal. I've got uh a little tin hat on. So uh, you know, I I I'm my formative years as a software developer were in defense, so occasionally that jumps out at me a little bit. But in the main, I don't really I I don't want to go as far as saying I don't care, but I I come from a cybersecurity background as well, for instance. Like if people can get overprotective and not protective enough to be a Daicomi again. Like so face if I've got not that I've used that comment, but I've got a Facebook account. If someone hacks into my Facebook account, it's annoying, but so what? Well done. What are you gonna do with that? Not a lot. I was rather you didn't, but whatever. Uh and it's the same with data. Okay, I give data the chat GPTs and all that kind of stuff, maybe more than I should, but ultimately at the end of the day, it's unfortunate if they use it against me, it's not problematic. Um, so I tend to think about what if the how could they use this against me? And if it's annoying, I get more benefit than than not to take that risk. Um, but would I upload anything? Not that I do anything, if I if I were to do anything, it could possibly be considered illegal. Would I give that to ChatGPT if it could be used as evidence? No, I wouldn't. But um would I give it some data that, you know, whatever? Yeah, I don't really care. Because so many people use it, they're not going to use it against you. I mean, uh there thousand millions of people use it. They don't really care about you, to be honest with you.

Andrew Jones 1:18:12

So therefore, not as the individual. They just want the the block data maybe to work out averages or the Yeah, and again, I'd rather not.

Chris Knight 1:18:20

But I get but what's the benefit I get from doing it? That's the thing. That's the calculation I do. So there's a risk doing it, but my benefit outweighs that risk. So whatever I'm gonna do.

Andrew Jones 1:18:30

Is it no different than the fact we all carry phones around with us? And I suspect ultimately they can track us wherever we go because we've got a phone in our pocket.

Chris Knight 1:18:36

I think this is a different podcast as well. Yeah, fair enough. Happy to get my tin how, and I love about that, but uh I'm not sure that's the same.

Andrew Jones 1:18:44

Just really use an example we're all used to doing now anyway. We don't think about the fact that we have a phone in our pockets that ultimately can track where we are.

Chris Knight 1:18:51

Uh I yeah, your phone's a problem. I text messages are more susceptible than anyone would ever want to realize. So there's during authentication, don't use text messages. For God's sake, never trust the text messages or if you want to turn how on. Um, but yeah, there's but there's bigger, yeah, uh to take the more rendered part, there's bigger risks in your life than what you give Chat GPT. But you know, don't give us something you could use as evidence against you.

What AI Should Never Do

Andy Lessey 1:19:19

Hi guys, uh Andy Lessey. Uh so you kind of inclinated that it's not actually intelligence and it's more artificial capacity that you're essentially getting. You're getting something that gives you massive output but is very naive, needs guardrails. So what jobs shouldn't it do? And is there ever a time where Andrew, Sarah, and Mike are going to be replaced by a digital interface?

Andrew Jones 1:19:42

Well, if you want to pay me for my patented rights, I think as was just mentioned, and I can retire on it, then you know, crack on.

Sarah Bolt 1:19:48

Just imagine the intelligence that could come from this voice.

Andrew Jones 1:19:52

It can only go downhill from here, can it? Sorry, anyway, to to to who wants to to reply to Andy's Well, what jobs won't be replaced?

Ian Garner 1:20:02

I think my wife's a teacher. I don't think AI will ever replace teaching.

Andrew Jones 1:20:08

It's gonna it's gonna be the physical, surely someone on farm, unless you suddenly got um what what's his name? Elon Musk trying to make his robots and whatever, somebody's still got to be on farm doing the the stock work, the day-to-day work. It it's it's the data handling by the sound. From what you guys are saying, it's the data handling is what it does rather than the the physical.

Ian Garner 1:20:29

Yes, it's also the trust as well, right? Like you've alluded to. Would you trust an AI robot to put out a fire or to capture a criminal or to teach your children? That's what I'm saying.

Andrew Jones 1:20:39

Well, I think I've seen a video of one trying to load a dishwasher or something, and it was hysterical. The plates were just ending up all over the place.

Ian Garner 1:20:47

But would I would I trust equally, would I trust uh you know AI to fully code data analysis? I I wouldn't right now. In the future, like we're talking again, I don't it's the guardrails. It's it's it's it's a tool to be used to empower certain people. I think it goes back to the question about will it make people smarter or it's exactly like Chris said, it's duality. It's some like we will lose critical thinking, other people will gain empowerment and be able to self-learn and learn so much more with it.

Andrew Jones 1:21:19

Um well, I I guess I'm thinking my my kids are doing stuff at school that we probably didn't do till secondary school, but they're doing it now. But on the flip side, there's probably things that they don't know that we did know, or or I suppose we talked about electricity. If electricity went tomorrow and we all had to do things again the old way, most people wouldn't have a clue where to start with. They've lost those practical skills because they're being done for them by electricity or whatever it happens to be.

Ian Garner 1:21:46

I guess it's it's how people how people use it. You could you could use it to help develop your own ideas, you could use it to take away all your ideas and but and you become numb almost. So it's it's how you as a person choose to use AI, right? If you use it in a sensible, ethical, smart way, it can empower you, it can make you very like it can help, right? Uh if you just use it and rely on it blindly, it will reduce your critical thinking ability. You won't be able to analyze that. There's a great um example in games, uh probably another podcast. Um, but back when I was a kid, games used to challenge you. They used to, you know, you didn't just endlessly mine a something and you'd have to, it was lots of jumping puzzles, like you'd have to put a lot of effort in to complete the level and you'd learn how to defeat the level. Now it's just like, well, you can pay your way through if you can't do it, and you don't develop those skills, so it's again, it's down for the person.

Andrew Jones 1:22:42

Peter Megris, funny enough, last night uh My Larf and I watched a new series with um Hannah Fry um talking about uh interaction of AI and people, and as you say, you're talking about critical thinking, used the example of the young man that tried to kill the queen with the crossbow, and he just got lost in I can't remember what the AI model was, but he just got lost in it and it just became an echo chamber and it just went round and round, and he he effectively lost his um critical thinking and he ended up with a psychosis. But it's that's like I say a whole nother podcast about the morality, maybe, but we're trying to show that you know this is this stuff's coming and there are positives from it, and it's like everything it can be used the wrong way.

Chris Knight 1:23:27

Well, AI models have great knowledge, but they have awful wisdom. Um because what they're they're trying to do is what they what literally what they're working is if you give it a question, say that's 10 words, they're using those 10 words to statistically guess what the next word is.

Andrew Jones 1:23:44

Well, yeah, you you you mentioned was it large large language models, such as Chat GPT. Because she she explained it yesterday or on what I was watching last night.

Chris Knight 1:23:53

So it's just using its knowledge to guess the next word, and then once it guesses the next word, it guesses the next word, and so on and so forth. So that's the example of knowledge, but of course it's not an example of wisdom because it it knows everything, but it has zero wisdom because it's just guessing your next word. So it's you can almost categorize it as a con a confidence con, and that's where hallucinations come from. I always equate LLMs to like your good mate down the pub. Like nine facts out of ten, he'll say confidently and all right, but there's one he'll say he has no idea about, but he's delivered in such a confident way, you're like, okay, yeah.

Andrew Jones 1:24:27

Especially if he's had one or two in the crossing. You know what I mean?

Chris Knight 1:24:29

And it's all what is. But but that's where it's come from because it's just guess it, it's confidently guessing. I said um, I'm 100% sure that neck this is the next word, whether or not that next word is right or not. Um one job they should never do it in this in the short term, it should never be a decision maker. They're amazing advisors. Never let it be a decision maker, right? For any in any in even in the next five, ten years.

Mike Jones 1:24:54

Mike. Human interaction with each other is a priority. We would all be lost if we didn't weren't able to communicate and see and touch other humans. Artificial intelligence and you know, using copilot, it can be addictive. And that's why, you know, guardrails have you know need to be put in place. You know, there's a big dis debate at the minute whether smartphones should be in schools and kids and everything. Um how the hell are I gonna stop that with you know using copilot? Copilot or you know, using these um artificial intelligence learning machines will be taught in schools eventually, just so you can put boundaries.

Andrew Jones 1:25:42

Well, I I remember when we did the podcast down at Canningham College and they said it's changed so much now, everyone's got the phone in their pocket that kids will literally challenge them then and there in a way you didn't before because you'd have to go to the library, go through the books, whereas they can just literally pull it up then and there. It's like, okay, so they've had to change their teaching methods because the challenge can almost come back instantly, and they've got to go, okay, fine, let's go down that rabbit hole and find out what you're saying is or isn't, or whatever it is. It has changed things. And you're right with the you know, it's the interaction of just a classic example. Doing the podcasts in person are so much better than doing them virtually. We do do some virtually. I mean, we've done one with Chris, we've done one with you, and we've done one the second one with you, Mike, virtually. We can do it, but it's not the same. It's not, it's like here, I can see Sarah going holding her mic up, she wants to speak, whereas you just don't get that. It's seeing those little subtleties that make the difference.

Best Real Uses Farmers Can Try

Sarah Bolt 1:26:38

Definitely. It's that uh human behavior that uh that just isn't there in that AI model. I've got a question. Um, so what is the most impressive real-world example of AI you've seen? And can we make that applicable on farm at all?

Andrew Jones 1:26:57

That's a good question.

Chris Knight 1:26:59

I mean, I I would say Chat GPC and uh Copilot, like from a purist AI perspective, they're amazing. I've just hated on them for the last hour, rather than a little bit. But they are amazing. They truly are. And that's and where I think it came up earlier how how how farmers can use them. That I I'm gonna put ourselves out of business a little bit. One good example is just give it the data you've got. Like so some people can be overprotective with their yield data. If you've got yield data, weather data, give it to the Chat GPT and see if it can do correlations. You'll get a very rough, oh, you'll get a somewhat of a useful, again, all models are wrong, but some are useful. I think you'll get some use out of it. And it'll it would be able to do the correlations between your weather data, your yield data, or whatnot. They're good pattern recognizers. So if you can give it useful data that you think there's a pattern between, give it to the chat GPs and the LMs. They work, they're good enough to spot that kind of stuff. Um, so if you've got long, large historical data, um, you can't let me as a bit meta, but we are using AI to make AI now. I mean, that's a good example of it. You're giving AI model data to then make its own AI model to do your bidding.

Sarah Bolt 1:28:05

So in that situation, if we're a little bit scared, can we put it into like temporary non-modes of putting it out there into the big wide world on sort of things like ChatGPT and copilot?

Chris Knight 1:28:17

Uh that's a legal question. Uh I'm gonna I'm gonna go out and say check the terms of service, but I'm pretty sure that Chat GPT have a box you can fix somewhere. Yes. But again, I'll re I'll go back to my next point. Okay, yes, it's I'm not going to say it's not valuable data, but ultimately, how are they going to use it against you? And the potential upside of giving that data outweighs what risk it's for for you all to decide for yourselves. But I'd really ask, what's the benefit versus that risk? And make a decision from I suspect give it the data because there's a potentially good upside on it.

Ian Garner 1:28:49

I'd just add to that, I think it's it it yes, you could you can give it data. It's it it comes back to consistency as well, though, right? When those models update, you don't know exactly what's changed. So you might give that data one month, a new model might come out. You give the data the same thing and it analyzes a completely different way. So it's it's it's a great kind of screening tool. Um, but I wouldn't I wouldn't rely on it to be reproducible uh unless you set up again guardrails. You can set up agents that will run through exactly how you want to do it. But yeah, you you give data to different models, even you give it to ChatGTP, you give it to uh Gemini, it will give you slightly different interpretations. It won't it'll hopefully get the maths right every time, but it might give you slightly different answers to your questions.

Andrew Jones 1:29:34

So it's a lot like I suppose asking two people, they might have the same data. Because they're all trained on different data, right?

Ian Garner 1:29:39

They're they're all doing very similar things, but their data sets that they're trained on are different. So it's their slant on yeah, exactly. It's why some some are really good at coding, like Cloud or Claude is really good at coding. ChatGP is really good at coding, but you know, Gemini and perhaps co-pilot better at different things.

Mike Jones 1:29:57

Um Mike, do you want to add? But I'm just gonna say, from a low-level point of view, actual practical um applications that I'm currently using, um, I've just updated a lot of risk assessments and I learned to ask the right questions or put the right questions into Copilot Chat GPT. And as long as you gave it those guardrails and you didn't just uh print it off and put it in a file, you read what it was t churning out, because you then have to edit it and everything. It has really helped in putting standard operating procedures written down in a format that is a lot of people can buy into.

Andrew Jones 1:30:46

So what you're saying is a bit like searching something on Google in the sense of you have to know how to ask those questions to get what you're wanting out of it. Because I mean, let's be honest, you could search something on Google, and if you don't ask it quite the right way, you won't get quite the data you're looking for.

Mike Jones 1:31:04

Yeah, exactly. And and learning what questions to ask it, they're the guardrails that you have to put in place.

Andrew Jones 1:31:11

Question from myself, I think in some ways you've kind of already asked it, but how do you think AI will affect farm management moving forward? Writing risk assessments was a lot easier.

Mike Jones 1:31:23

Um it is going to give greater focus on time efficiency and being able to get get, you know, uh standard jobs written out in standard procedures so that a whole team know what they're doing with it.

Andrew Jones 1:31:42

And it sounds like also tracking data, maybe tracking markets as to when might be the right time to buy or sell or Yeah, you know, like you know, Agribot's um program, they will be able to hopefully predict what countries are gonna have fantastic yields of maybe wheat, and other countries might be a bit short.

Mike Jones 1:32:04

Um so you know, w when to sell a farm's wheat, whether it's straight off the combine or whether they go and put it in storage for six months, um hopefully that uh gives a better chance of maximising the return of profit.

Andrew Jones 1:32:22

So it sounds like it's it's like everything, it's just speeding up maybe some of that decision making from what we've already said. It's it's sifting through that data quicker, presenting you with the options to then go and make the decision. Exactly. Either of you two want to add on to that one at all?

Chris Knight 1:32:38

Just just look, I I I think like we talked about there, the one thing I try and keep in mind with ours is that you know farmers are commodity traders as well, of course. So, of course, well, we produce commodity traders are interested, but farmers are they're asked the same questions as farmers, uh just in a very different race. So I think what AI would do on that point is it's a very good level up. Like, for instance, uh before you'd have to pay for expertise, knowledge, right? Like if you're doing particular weird commodity trading stuff. That was the preserve for the professionals to spend years in that. Whereas with ChatGPT, you can get at least a junior level understanding of that.

Andrew Jones 1:33:12

Well, I suppose you've just made me think a client of mine. He said he says, My he said to me one time, my life's easy, because he said I just do what you tell me to do. Or you and others, you know, you're the experts in whatever you do. So you tell me I need to do, I don't know, buy fertilizer or whatever it happens to be, then that's what I do because you tell me it's a good time to do it. So what you're saying is it's basically doing that that for you.

Chris Knight 1:33:34

That's good marriage advice as well as being good advice on how to how to how to use AI because that's it. You've got, I mean, you're all amazing professionals at your job, but chat GPT gives you entry level to any other profession as well. And I would use it for such. Like uh, you know, I'm sure if farmers are much better than me at commodity trading, but there's there's little tricks that commodity traders use as well, like for arbitrage and all that kind of stuff that ChatGPT can give you knowledge on as well.

Ian Garner 1:34:02

So I'd use it from that. Just to add to that, I think like there's something like right now you could do, right, as a farmer with no uh experience of AI, you could literally take a picture of, say, of your barn right now, and AI could analyze that and it could say, right, the lighting's low, there's not enough space for the cows, the barn's dirty, like clear, like things you might know, but it can like see that already. Um, it goes back to the neural networks, and and it it will have pictures in its data model of here's what a good big nice barn that's like here's not such a nice bar. And you'll be able to pick out things like that. So there's something like you could do straight away to use AI, maybe to help, maybe to give you some insights.

James Yeatman 1:34:40

But again, putting me out of a job with the chaos signals, but fair enough. James. Um so yeah, James Yeatman. So I I I just think perhaps we're we're sat in a room with uh by perhaps two exceptions of really a bunch of Neander tools that are trying to grapple and um even even as presenters, you know, you're you're you're the wrong side of 40, aren't you? Do you know what I mean? So I think the the solution really is to employ amazing young people, intelligent young people that already understand how to best use these tools, and then use our experience and knowledge to make sure they're moving in the correct direction. So just making best use of the tools that we use it on the farm, but that's only because I've got four bright kids that help me use it well. Do you know what I mean? Because I'm I'm I would not even know how to turn it on. Do you know what I mean?

Andrew Jones 1:35:27

So you you you're the guardrail, effectively. What you're saying is you're letting them do the the searching and then you're making the decision.

James Yeatman 1:35:33

Yeah, and it is it's a bit like all young people, you know, you empower them to go for it, you know. So so uh and I think that's where we've got to go with it. And in in people mention the um caricatures on Chat GP that are created by Chat, but they're okay, aren't they? But they're a bit samey. There's nothing fantastic about any of them or anything very unique. They're they're just is a it's it's just repeating a model all the time, isn't it? So I I I I'm not an intellectual man, but I I I I'm a practical chat, but but thing I've got a crazy mind that brings up all sorts of stupid ideas, and then I need somebody to um just uh sense check some of those ideas. But chat GPT is not gonna come up with those crazy out-of-the-box ideas, in my opinion. Uh that it's just playing safe all the time using data that's already there. So it it isn't gonna solve all our problems. You still need the crazy bit that humans come up with. And your knowledge as well, that you've inherited as well, most likely, right? It doesn't chat GPT observes the stuff that's in there. It doesn't observe stuff that's on the periphery, and that's what humans do. They they use observations, don't they?

Ian Garner 1:36:37

So yeah, yeah. And you also get something that's called algorithmic bias in in AI. Uh, Chris will would be able to talk far more about it than I would, but with the data that they're trained on, say it's character characters, yeah, images of stuff. Um, like they will be trained on a specific data set. So you often get like a lot of bias in those. Uh if you ask it to produce like a I don't know, a Disney princess, it'll come back as uh probably a white young female. It won't give you that kind of range because it's been trained on on those kind of models. So it's uh yeah, but going back to your point, yes, I think it's um you know, the younger generation up, data's all around them, right? Technologies evolve, they evolve, they grow up with it, so they don't know anything other than that. You know, my my kids are what four and six now, and they wouldn't believe like you can swipe on a screen. Like they go to the well, when they were younger, they went to the telly and tried to swipe it. It's like it's it's almost embarrassing as a parent because they've grown up with that. They just they that's what they think is yeah, it's not it's normal to them. So they they grow up in that environment. So, you know, it it would be it would be the opposite, right? If if they didn't have data or machines, they wouldn't they think how the hell do you do bombing without this? Whereas you grew up without that.

Data Ownership And Getting Value

Andrew Jones 1:37:52

Well, I know my other half teaches first aid and she's done some at school, and like you're talking about the screen, she deliberately went and brought some old rotary phones and that off eBay, just so that and so like how do you work this kids? And like you say, they just got no idea, and yet for us it's just what's what we did, it didn't know any different. And it even the push button, I think, for them was a bit like uh because they're so used to like the the smartphones now where it's all touch screen and whatever. It is, it's well, like I guess point we made earlier, we the dichotomy, while we've gained in some ways, we lose in others.

Sarah Bolt 1:38:28

So we've been talking a lot about data, and we've been talking a lot about how data is valuable. Um who owns the data? And then probably a second part of that question is really um with that in mind, what data could farmers be using that they're they're not perhaps um doing much with now, but AI could help them do something with.

Ian Garner 1:38:54

Well, uh, if I if I start then, um answering who owns that data, I think, is is a it's a it's a difficult question. Um I think you know, you if you sit on data and you're not not doing anything with it, then I I think that you and you could be doing something with I think sharing comes back to sharing data, right? Is companies like say Antler Bio, we we're only trying to help farmers with data. Like we need data to be able to give back recommendations and try and then it's a constant feedback loop. You essentially you would make changes, we collect that data, we see what works, what doesn't, and then we feed that back into our algorithm. So it's it without that data, we don't exist. But if we don't exist, then farmers might not get better at certain things. So it's it's like that kind of two-way trust almost. Um, what data you're sitting on, I mean, with LeLai robots, there's a huge amount of data there. Um, they will be using part of it, right, for predicting things like mysticous from the conductivity or the um uh somatic cell counts, for example, they'll have their own algorithms that'll be developing on that. Um, but there could be other data that you, you know, that's very easy to collect. I think potentially the it the image data that I suggested earlier of barns, the state of the barn, the lighting, like is quite there, there are caveats with that. But you could take pictures of your barn, um, data that belongs to you. AI could go through that rapidly through through neural networks and and tell you, like, is your lighting too low? Is it too high? Because lighting will affect, again, back to I don't want to keep from saying out in the bar, but like we we look at epigenetics, right? And everything can pretty much affect epigenetics. So lighting can affect gene expression. That can go on to affect how you respond to nutrients. Um that kind of data that you just have all around you, if it's not structured, if it's not collected, it's kind of it's it's wasted data. Um but it's it's a difficult question to answer of who owns what.

Sarah Bolt 1:40:51

Uh I guess Perhaps we shouldn't worry about it.

Ian Garner 1:40:55

Perhaps. I mean, we're we're here I get I uh this is where I agree with Chris. Like if if everyone shared their data, if everyone in in the entire world just shared all the data, we'd probably advance a lot quicker. But then there will always be people companies that will be using it for not nice reasons, and there will be these companies that are trying to use it for to better humanity and and animal welfare or whatever. So it's it's it's a balance. I I don't know how to exactly answer that problem.

Mike Jones 1:41:23

We're always told that we own our own data in the farming industry. How much of that is true? I just don't know at the minute. And I suppose the big thing is how does a farmer put a financial return on that data on if he can uh use it as another income stream. You know, it's other companies are are gonna be wanting that data to help them develop other things, you know, uh technology. And if if that farmer could put a financial value on giving that data away, that's another income stream for the farming community to try and uh enhance their business. Um, but you know the the the big thing is is for me is bringing data in from different platforms and putting it on one platform.

Sarah Bolt 1:42:22

That integration is the big one is integration.

Mike Jones 1:42:24

If the the sooner we can get to that point, the better. I think we're a long way off for a lot of it.

Sarah Bolt 1:42:30

Again, it's that cooperation between companies, isn't it?

Chris Knight 1:42:34

Well, we we uh for instance, we had a project with Mike, so Mike uh we had Mike's data for his grass growth and whatnot. In my world, it's insane to think anyone but Mike would own that, or I could take whoever the legal is, but you know what I mean. Um and when I say whilst we sell uh whilst we help out commodity traders, we never ever sell data that's not ours. So I would never hand over or sell Mike's data to anyone else. We use it to train our models and to help his farm. It's not our place to then sell that or even aggregate it, then sell it to so we have a strict firewall between those kind of things. And that's fine because we use satellites, we can use satellites to measure it anyway. Um, and for that, uh, we owned all that data because we processed it and cleaned it and all that kind of stuff. And that's just you know, that's literally driving past someone's field and having a look at a field at scale, right? Using satellites. But if it's actually that kind of stuff, 100% farmers own it to give it a value, I think it's harder. Um, I think individual farms, not worth a lot, to be quite honest. Because the issue there is we could make a model for that, but it would only be useful for that farm, which is good because we do that, we do sell that as a well, we we do have that as a model. Like we all our AI models are per field. So we have a model for every single field in the UK individually, um, using using that. But if the farmer gives us their data, we make a completely separate model that's specific for their application. So we can make a more accurate model because of that data, but then that model is only useful to that field, but that's fine because that's what the farmer uses it for. So it it depends. Um, so I I would always I would always say the farmer's on it, but it's difficult. I think you're gonna monetize it, you're gonna have to create you're gonna have to create associations because a single farm's data isn't worth a lot to an AI model. But if you've got an association with, like, say, thousands, okay, now it does have a value. Um, what that is per farm, I don't know. I'd pay, but that's when you start to get the value when it's big data sets. Um and lastly, what I think I think your question was a great example to give ChatGPT. I would ask ChatGPT, I'd say, look, this is the data I have on my farm. How can you help me? Like, what can you what can you do to with this data to help me uh be a farm? I think that's a little meta trick that people forget with ChatGPT, is you can ask it these sorts of questions as well. Like, how can you help me if I have this data and do that kind of thing?

James Yeatman 1:45:06

And that's the sort of thing that my children do because they understand it. Do you see what I mean? Whereas I even have perhaps thought of that, but they they they understand how to best utilize the tool.

Chris Knight 1:45:16

The geeky term for that is prompt engineering. If you if you want to go to Google, Google prompt engineering courses or YouTube videos. There's lots of things to do that. I mean, I'm I'm not that good at that, actually, to be honest. I'm far better making AI. I'm a terrible prompt engineer, awful. Uh, but that's the term you want to search for.

James Yeatman 1:45:34

So perhaps our generation need educating on how to best use these tools.

Chris Knight 1:45:38

Uh it's just anything. Like I like I said, I'm a data background, so I attack agriculture because it's data. No AI, machine learning, computing software game is number in, number out, right? I'm from the countryside, I grew up in a countryside, but I'm not a farmer. So I but that's what I do. But of course, we've got people such as Emma who are from agricultural background to keep me in bounds. So that's our bounds and checks. We do that kind of thing. So I don't I think you're being hard on yourself there. I I think it's uh, you know, I'm from the AI sector, but I need to train myself on prompt engineering as well. I'm not inherently good at it, it's a different skill set. Um, so yeah, I think ever I think everyone can be better at it, to be honest with you.

Andrew Jones 1:46:15

I I wouldn't necessarily kind of today, really, is to make people aware that they might just think, oh, it's there's a little bit of chat GTP or there's a little bit of on our phones, but actually it's out there and it's now being aware that it's out there and then using um trying to use it in a more productive way.

Chris Knight 1:46:33

Yeah, like Chat GPT. I use I'd be surprised if I use more than five percent of its features, to be quite honest, but I get a lot done. And so on your point, you could levy that me. I should be watching these videos and being better at myself. So I I I wouldn't use that as a reason not to get you've definitely not missed the boat or you're behind the scene. I'd I would imagine 99% of people are basic users of ChatGP.

Andrew Jones 1:46:55

But it's a bit like I suppose I made the comment earlier about Google, it's knowing how to ask the questions and what you're saying, it's ChatGTP, the prompt would you say prompt engineering? It's about asking the right prompts, the right questions of it.

Chris Knight 1:47:05

Yeah, and yeah, I think the advantage I have is I understand how it works, so I can kind of second guess it sometimes maybe, but even that's a dangerous path to go on.

James Yeatman 1:47:16

And that was exactly how we got involved with the debate with my son as to whether it was a good tool or a bad tool. And and uh there was a debate on um uh LinkedIn, like I and his his comeback was the the more the tools use, the better it gets. And that's very true, isn't it?

Final Takeaways And Thanks

Chris Knight 1:47:33

And it's AI in general, the more data you give it. But again, it's you have to be careful because you can do what's called overtraining. And so, example, I'll give the example of one farm's data isn't that useful to us because that's 100% overtraining. That AI model can solve that problem and that problem only. Well, you want lots of data is because then it becomes more generalized. We'll talk about how well does the model generalize? So if I've built made a model of this, and you mentioned bias earlier, that's a bias. If say I have uh a data with nine fields, uh ten fields, nine is from your farm, one's from another farm, that AI is going to be more particular to your farm and kind of forget the other one. And that's why we talk about bias and overcorrection. So more data isn't always you want good, broad general data. That's a good picture of it. You don't want it being biased because then that bias comes into it. And I touched on the point earlier, like just just the chat GPT does a little of these hallucinations because it's learned from the internet. So if you've got a million people with the same wrong opinion, um the AI is going to repeat that opinion. So I think the standard one is if you were to ask ChatGPT 2,000 years ago, it would have been a flat earther because that was the prevailing knowledge then. It would have said the world was flat. So it's it's the majority opinion. AI does the majority opinion, not necessarily the correct one.

Andrew Jones 1:48:55

I'm looking at the time now, thinking like usual, it's time we wrap this up. So I guess starting with you, Chris, any last words of wisdom?

Chris Knight 1:49:03

Uh I think we've stretched on my wisdom, to be honest with you. I think I've overperformed, it's just over I am. Um no, it's just it's exactly that. I think don't be afraid of AI. It's I'm I again I'll I'll emphasize the point, just because I make it doesn't mean I'm good at using it. Quite the opposite. Uh I've I've faced the same struggles on using ChatGPT anyone else does, so I've got to battle for that. I could definitely be a better user. I tried to make a video the other week there just to show training. So I had to do a video. I was terrible at it. But through for the wonders of AI, I managed to churn something out. Whether it replaces a professional, I highly doubt. Uh, but yeah, I think that the emphasis is use it as a life, a quality of life tool. It saves you time and makes you more efficient. It doesn't replace you. It doesn't, the only person who could argue it replaces is junior hands, maybe. But it's a quality of life tool and it makes you uh be more strategic. Than uh necessarily tedious jobs.

Ian Garner 1:50:03

Ian. Yeah, no, I'd I'd I'd echo exactly what Chris said there. I think don't be afraid of it. It's it's here, it's here to stay. It's just like the internet, it's never gonna go away now. It's it's mainstream. Um let it empower you in in and use it carefully. But yeah, in like this it's a good tool. They're good tools, and I agree with Mike about the whole you know, human interaction. Don't don't forget it. Um but yeah, use AI responsibly and learn.

Mike Jones 1:50:32

Mike Yes, machine learning, artificial intelligence is here to stay. Embrace it, but always be mindful that you need those guardrails and to just think about the questions that you ask it. Um and it will help you to be you know, just better strategic decision making. Sarah.

Sarah Bolt 1:51:01

I think the the takeaway that I've got is that uh somebody said a really good phrase earlier. Don't let um AI be the decision maker. And I think that's my takeaway message that help use it to help with some of those uh some of those points, but actually don't let it make the decision in the that final decision. That human element is still obviously the most important part. So I think that's my my learning for the day.

Andrew Jones 1:51:30

Well, I think it's been a very enjoyable or a couple of hours now. We'll see what it edits down to when we get there. Um I think there's a lot we'll take out, is the honest answer. But um, from my point of view, no. I thank you very much to our three guests. It's been absolutely fantastic. Um, I think there's definitely been some great learning um for all of us listening uh and uh for those taking part today. Um but otherwise, again, thank you to uh UK Agri Tech Centre for hosting us today and and um letting this happen uh and running with the idea when I approach Mike. I've got this idea. What do you think? Uh and came back very quickly and said yes. So thank you to Mike and to uh say UK Agri Tech Center. Um, but otherwise, um, no, I think it's been really good. It shows that there are many positives. All we hear about maybe is the horror stories uh in in in the news. Um, but there are a lot of positives from using AI, and maybe it's not going to quite take over the world just yet. I think you said 50 years, wasn't it, Chris, before the next big step? Not before 50 years. So, so other than that, um, I'd uh I'd just like to say thank the three guests, and uh otherwise I guess it's a goodbye from me.

Sarah Bolt 1:52:44

And a goodbye from me.

Legal Disclaimer

Andrew Jones 1:52:45

Thank you very much, everybody. Thank you for listening to the Chewinthe Cud Podcast, a podcast for the UK dare industry, brought to you from the southwest of England and listened to around the world. And now for the really boring bit, I'm afraid, the legal disclaimer. The information provided during this podcast has been prepared for general information purposes only and does not constitute advice. Information must not be relied upon for any purpose and no representation or warranty is given to its accuracy, completeness or otherwise. Any reference to other organizations, businesses or products during this podcast are not endorsements or recommendations of ChewintheCud Limited. The views of Andrew Jones are personal and may not be the views of ChewintheCud Limited, and the views of Sarah Bolt are personal and may not be the views of Kingshey Farming and Conservation Limited and any affiliated companies. For more information on the podcast and details of services offered by ChewintheCud Limited, visit www.chewinthecud.com. Thank you and goodbye.