Summary:

This episode brings together Canada's sharpest minds at the intersection of AI and agriculture to make a concrete case: the future of Canadian farm competitiveness will be decided not by whether producers adopt AI, but by how intelligently they manage, govern, and deploy the data they already have.

Mohsen Yoosefzadeh Najafabadi introduces the concept of 'dark data'—the vast archive of unused research and field observations sitting dormant in breeding programs and farm records—arguing it represents an untapped foundation for training custom agricultural AI models. Donald Killorn reveals that choosing between cloud-scale processing and on-farm edge inference could mean the difference between 200 gigawatt hours and just 1.5 gigawatt hours of electricity consumption across Canadian agriculture, making edge computing not just a technical preference but a national energy and sovereignty question. Mohamad Yaghi makes the case that the battle for sovereignty is happening faster than we realize as global data governance threatens to outpace local agency.

Show notes:

This episode of The Future Herd gathers Jesse Hirsh, Mohsen Yoosefzadeh Najafabadi, Mohamad Yaghi, Jennifer MacTavish, and Donald Killorn for a rigorous, advanced-level conversation about what it will actually take to make artificial intelligence work for Canadian farmers. The central argument is not that AI is coming to agriculture—it is already here—but that the decisions Canadian producers, researchers, and policymakers make right now about data governance, compute architecture, and model sovereignty will determine whether the productivity gains from AI accrue to Canadian farms or flow to the proprietary platforms of multinational equipment manufacturers and cloud providers. This episode is less a primer on AI and more a strategic briefing on the infrastructure choices that will define the next agricultural policy framework.

One of the most striking contributions comes from Mohsen Yoosefzadeh Najafabadi, who introduces the concept of 'dark data' to describe the enormous volume of agricultural research observations that are collected, archived, and never used again. Drawing on his own plant breeding programme—where he evaluates 200,000 individual plants but ultimately selects a handful—he points out that the discarded data represents precisely the kind of rich, domain-specific training material needed to build trustworthy, custom AI models for agriculture. His argument is that researchers and producers often do not mind sharing this kind of archival data, but lack the governance frameworks and institutional strategies to do so responsibly. His proposal: train models internally on pooled dark data, then share the model outputs rather than the raw datasets, preserving privacy while unlocking collective intelligence.

Donald Killorn grounds the conversation in the physical and political realities of deploying AI at farm scale, reporting directly from a week that included installing a rooftop WiFi gateway in eastern PEI and presenting to the House of Commons Standing Committee on Agriculture. His core insight is that the question of where computation happens—in the cloud versus at the field edge—is not merely a technical detail but a strategic decision with massive implications for energy consumption, data sovereignty, and farm autonomy. He estimates that moving every data point from Canadian farms to central models would require roughly 200 gigawatt hours of electricity, but that training edge inference models to process data locally and send only what is necessary upstream could reduce that figure to approximately 1.5 gigawatt hours. He also raises the unresolved tension with OEM equipment manufacturers like John Deere and Case, whose proprietary data platforms risk locking Canadian producers out of their own farm data, and describes the limits of industry-led interoperability initiatives like AEF's Agon connector.

Listeners will come away from this episode with a clearer understanding of why the ROI conversation around AI in agriculture has to move beyond productivity metrics and encompass labour shortage economics, energy infrastructure, data interoperability, and national policy. Mohamad Yaghi frames the stakes plainly: AI is most valuable not as an automation tool but as a decision-support layer that helps producers commit capital more wisely before it is spent. Taken together, the panel makes a compelling case that Canada's agri-food sector sits at a genuine inflection point—one where the choices made in the next two years around data governance and compute strategy will either entrench dependency on foreign platforms or establish a sovereign, resilient, and genuinely productive AI infrastructure for Canadian agriculture.

Topics: Dark Data, Edge Computing, Data Sovereignty, AI Model Training, Farm Decision Support, Equipment Interoperability, Agricultural Policy, Energy and Compute

Hi, I'm Jesse Hirsch.

Welcome to the Future Herd.

Today's episode is exceptional.

And, uh, granted, part of my job as host is to use this part of the

intro, to hype it up and tell you why listening to this is worth your time.

But in this case, I cannot overestimate the, the knowledge, the insights,

the dynamism, the brilliance of the people participating in this podcast.

Talking about a really contentious subject.

Artificial intelligence, which is a catchall, is a whole bunch of technologies

and tools fit under that spectrum.

It's something people have a lot of anxiety about, a lot of curiosity,

we hope, but also sometimes fear.

So today we really tried to not only break down what AI is, uh, how AI models are

trained, what edge computing is, but the business, uh, case, the, the, the larger

ecosystem and the way in which agriculture is particularly poised to be a beneficiary

of the AI revolution, especially the ag agentic revolution, IE the rise of agents.

Now, if all of this is a kind of buzzword bingo for you, which as a regular

listener, I would hope that it isn't, uh, this nonetheless is an advanced

class where we've taken the three, uh, top minds around AI and agriculture in

Canada and have tasked them with coming together to, to answer not just the basic

questions, but some of the more advanced when it comes to literacy, when it comes

to data governance, data sharing, and some of the challenges to sovereignty,

uh, not just around agriculture, but around the use of AI and society.

So this is, I think, one of those episodes that demonstrates why the future heard

exists as a cross-sectoral collaboration that brings together different actors,

different voices in, in a way that isn't happening anywhere else in

society or, or maybe around the world.

So again, I hope you enjoy it.

Certainly this is the kind of work that inspires me and it really gets into the

substance as to why we should be curious about AI rather than just close-minded.

Alright, let's enjoy the chat.

I.

Welcome all of you back to the Future Herd for what is on the technical level,

a panel episode, but in reality is an all-star episode as we've been able to

gather some of the greatest, uh, brightest minds in the agricultural sector, but

also in the AI sector in that, uh, the three of you, the five of us, are a

bunch of heavyweights when it comes to dealing with these particular issues.

And of course, I, uh, much prefer in the age of Google, in the age of

LLMs to forego any introductions.

Our listeners, as they listen, can do their due diligence

and find out who you are.

I really wanna take the most of the time we have together and, and really

try to get a sense of the advanced applications of the technology

in the broader public context.

And I say that because I personally feel torn every day, that on the one

hand I am experiencing such incredible technology that is allowing my brain

to go places I never thought possible.

And yet I'm surrounded by a public that seems afraid, seems angry, seems

opposed to the tools that I'm using.

So I kind of wanted to start with a general question of, you know, I

find the three of you individually and Sectorally doing amazing

work with amazing technology.

How are you wrestling with the public debate, whether in the abstract or

whether in your family, whether in your community, whether people are like,

what are you doing with that AI stuff?

How do you translate it in a way that, uh, cuts through the

noise, cuts through the hype?

And I'm saying this as an intro question.

' cause unfortunately, I think even within agriculture there are the sceptics.

There are the doomers, there are the people who aren't really ready

to open their mind to this stuff.

And I think the three of you are exactly the right leaders

to, to help make that happen.

So, Mohamed, I'm gonna throw to you if only because I suspect that you

have experienced what I'm describing, uh, uh, given the travels that

you engage in, how do you frame this technology to the public?

How do you counter some of the disinformation and some

of the fear that is out there?

So first off, thanks for inviting me onto this Allstar panel.

I wouldn't say I'm an Allstar myself, but no, of course.

My two colleagues here definitely have, um, you know, a, a storied

history, uh, within the sector.

So great to be on with both of them.

Um, when it comes to the, the conversation about artificial intelligence and

within the ag sector, uh, I was joking to my wife last week that, um, I feel

like I have a house in Toronto, but I'm always outside across the country.

Um, so I've been able to see like on the ground what's happening and, you know,

the application of these tools, person.

Um, and know, fortunately a lot of many provinces, um, and I think when

it comes to the application of it, I think for a long time it's always

been about what could be possible.

And I think this, the conversation is shifting towards what needs to happen

in order to shift how we frame it into a return on investment conversation.

And the reason for that is, is that when we tend to ask whether artificial

intelligence can save money, uh, I think the more honest question is

whether it helps producers make better decisions before capital is committed.

I think that's gonna be the huge breakthrough for the sector and what

we need because when we just look at, you know, the sector as itself, it

contributes over $149 billion to GDP.

Um, and that's not a marginal gain and artificial intelligence

can compound it over every acre.

But the proof of profitability is not just productivity metrics, it's

also looking at the labour shortages we see in the sector as well.

Um, I think a recent study by the Canadian Agricultural Human Resources

Council, um, they have way too long of a name, but I love them.

Uh, con I think 28,000 unfilled positions signified a $3.5 billion economic loss.

So ultimately the conversation I'm having with producers, some who are

using Claude to build their own apps, all the way to producers who are starting

to, you know, tinker around with it, it's all about return on investment.

It doesn't matter what this technology is unless it can

bring that to them at the moment.

And return on investment isn't just about profitability, it's also about,

um, you know, how do they manage their farms and can we manage that data on

premise within Canada and keep our data sovereign as well, and not try

to take Donald's points away from him.

But, uh, yeah.

Donald Killorn | PEIFA: Good setup though,

I try my

Donald Killorn | PEIFA: and shout out Kark.

Shout out Former future heard guest.

Jennifer Wright, the executive director of Kark.

They are doing a great job.

Yeah.

I mean, build upon that Donald, he clearly set you up.

I mean, to what extent is that ROI self-evident.

Donald Killorn | PEIFA: yeah, last week, maybe two weeks ago now, uh, my team was

on a roof in Eastern PEI, uh, installing a gateway and, uh, a wifi gateway and,

and basically creating the mesh network to bring the realtime data, uh, into

our AI platform that we need to mix with the meteorological and soil data to,

to try and do what mo's talking about.

And simultaneous to that, I was at the House of Commons, uh, presenting

to the standing committee on our agriculture compute strategy.

Uh, it's become pretty clear to us pretty quickly that precision agriculture

is becoming about what data needs to hit the, the main models and what

data can be used inside the field, boundaries in an inference model,

and matching hardware to software and, and where we can use inference.

Feels like the cutting edge this weekend.

Um, and agentic memory was sort of the cutting edge last weekend.

And so we're, that's kind of where we're at, uh, in the curve.

And, um, so yeah, if we try to move every single data piece around Prince Edward

Island, or pardon me around Canada, uh, we will need about 200 gigawatt hours of

electricity to run AI on Canadian farms.

But if we allow the system to learn and, and decide for itself what

data is required, uh, we believe we can get that number down to

about a gigawatt hour and a half.

Uh, so, uh, significant, uh, an order of magnitude less.

And those are the types of decisions and the types of strategies that

we need to be thinking about.

Um, it's difficult to find megawatt hours.

Um, we have to, you know, distribute our compute need across the country,

uh, but to most point ensure that, uh, agricultural data is treated with the

utmost, uh, security and, and sovereignty.

And, uh, the interesting thing, the most interesting thing about

working on that bottleneck while my team was sort of filling in the

blanks behind me was that the AI strategy came out that day at lunch.

And, uh, last week the food security strategy came out.

And the place where that overlaps is a tremendous space for research

and development and implementation and productivity gain in Canada.

Well, and I, I definitely want to come back.

Donald Killorn | PEIFA: Please.

Let's do it.

jump in.

I wanna jump in.

I wanna jump in.

Sorry, Jesse.

Uh, Donald, I've, you know, I thought, I think like the, the way we're

positioning this discussion, I think is really important, and I appreciate

the efforts that you and your team are making, and I think when it looks, when

I look at the different federations of ag across the country, I think, you

know, you have a really interesting strategy, and I'm really curious when

it comes to that national sovereignty in terms of, you know, keeping data

within Canada or even the modelling capabilities for our producers.

How are you dealing with the OEMs, like the, um, you know, the equipment

manufacturers like Deer, case, new Holland, um, you know, like, are you, are

you speaking to them right now because they are some of the biggest providers of

that data, but I don't know how much of that information is within Canadian soil.

Um, and also when it comes to that utilisation of data, how is

it being used for our farmers?

I'm just maybe really curious, just because I was at a conference last

week speaking to other manufacturers, and that conversation was one of

the, I guess it was a, um, I'm trying to find the right word.

It was, um, it, it basically, it was something of like a elephant in the

room that nobody really was addressing.

Donald Killorn | PEIFA: Certainly one of, um, the best perks of getting on

this data governance train early a few years ago was being at the North

American European Union Conference in Cuomo, um, last summer, uh, in Italy.

And I was,

pretty, that's a nice

Donald Killorn | PEIFA: it was, it was great.

And, and, and thank you to the Canadian Federation of Agriculture and, and

the PEI Federation of Agriculture for making that happen for me.

Um, and so I got to present and, and had a a, a lunch and learn,

and, and I was grateful to be joined by the, um, technical, the, the

technology lead for the Italian Equipment Manufacturing Association.

And he gave the room, myself included, a full briefing on the European regulations

that, um, compel these companies to, uh, provide, uh, true interoperability,

uh, in, in their platforms.

And so that's been, um, tackling, they're, they're tackling that through

an industry association, the A EF, uh, agricultural equipment fe uh, federation,

and they're developing a software tool called AG in, and they're ostensibly

doing that in North America and Europe.

Uh, now we've gone through the whole, um, the whole process to

achieve API compatibility, basically receive the agon connector.

Um, we've done everything on our part and sent them the check, and

we've sort of gotten, it's sort of, we're a little bit concerned now,

it seems like, you know, the wizard behind the curtain is, you know, not

necessarily what we thought it would be.

Um, know, and, and so it's an API integration that's necessary

to do what we're trying to do.

We believe the marketplace will allow that.

We, we, it seems like those companies are adopting a, um, a, a strategy

that allows 'em to be compliant to the European regulations in

the North American market as well.

And, uh, we believe we're creating an option for farmers to work with those

companies, but not be, um, locked into their, um, know, proprietary

data, data platforms, and that we will have interoperability in

Canadian agriculture between equipment manufacturers and that, um, farmers

will have a great deal of control over how that data is collected and used.

Well, and this is a good opportunity.

is a

Let, let me just bring, let me bring Mosen into the conversation here.

'cause this is a good opportunity to both tie in the research angle in terms of,

I, I, I think Mosen, you're experiencing a, a different, uh, reaction from some

of the researchers that you're sharing your research with and you're sharing

your work from, to the extent that I think you're, you're, you're finding

more and more enthusiasm for, for the opportunities these tools provide.

But I'd love for you to also chime in on this data piece.

'cause obviously part of the process of research is data

collection and, and the, the kind of governance of the data collected.

So I'd, I'd love from the research perspective for you to wade

into this conversation before.

Jen.

We, we, we throw it to you.

That's good.

So basically, you know, just, I've actually enjoyed, you know, what,

uh, MOED and Donald talk about it.

But, uh, in terms of my experience, I have a kind of like day-to-day

experience with farmers, growers, industry partners and researchers, and

all of them have different ideas in terms of adapting new AI technologies.

Like for farmers, they are usually suspicious about is

it like the AI technologies?

Is it going to be a good, um, platform for decision making or how they can trust ai?

Is it really good for them on, even if they are going to, like, if AI in the near

future going to replace farmers or some of the jobs that they currently expert on it,

like these kind of things, uh, are like, are totally valid, are totally valid,

to be honest, and I really appreciate what farmers are bringing to me, but the

way that I'm usually addressing these kind of challenges is not by answering

directly the question that they ask, but providing them necessary information.

Like who said AI is going to be a good decision making?

Who said AI is going to replace your job?

I recently developed an app that is not going to make a decision for you.

It's going to help you to find relevant information quickly.

That's it.

It's your decision, your act, your action, whatever you wanted to do it, do it.

But if you wanted to find more information, instead of searching over the

internet for an hour or two hours or even a day to find something interesting, you

can use this app and find it in a second.

That's it.

So this is one of the, like, you know, one of the things that I'm

usually addressing, and I believe trust is not building in one night.

It's building gradually.

And especially, especially for farmers and growers, this is something

that they wanted to focus on.

And I, I've, we've talked about like in the last time that we discussed

that I'm so happy that I was invited by farmers to talk about being GPT,

like something totally related about ai instead of talking about variety

and some of the like, you know, new varieties that we are releasing.

Of course I'm going to talk about them like in a couple of like meetings with

them, but that I have in the future.

But that my eyes on that time was how farmers are enthusiastic to adopt

these new technologies once they are certain that these are not going to

interrupt their day-to-day activity.

So this is something that I founded for the farmers, for researchers there are

still uncertain about like, you know, how valid the results that they are

getting from it or how much they can replicate it, how much they can adapt it

in the actual like research or even is it really good to use AI for the research

or not, or how ethical is this, right?

So I'm right now a part of the committee for, uh, AI Strategy Committee University,

and I'm seeing this kind of concern.

And again, all those concerns are valid.

We need to come up with a good strategy, how to adopt ai.

And I don't want it to just totally and blindly advocate for the use of AI and

say that AI is perfect for everything.

No, it, maybe it's not right.

Let's adopt it.

Let's check it and see if it's good.

We can use it if not.

Find a new way.

Right?

So, and then, uh, last things I wanted to talk about is how much data we are able to

share and we should share in the AI basis.

And this is something that, in the conference that I had last week,

I discussed with a couple of my colleagues, you know, in Europe, and

they said we wanted to adopt this new technologies, but we are afraid to

what extent we can share our dray data, what we, to what extent we can share

our information over the internet.

And there are a couple of base, like at the university, we have our own

networking, our own domain that we can use to build like, you know, these kind of

ais and we can like, you know, develop and train our AI models based on whatever data

that we have and only share the results with everyone who wanted to use it.

So there is no need that everyone has an access to the fundamental

data set that we are having.

We can build data, we can build, sorry, ai, open all the data that we have and

share the final result with everyone so they can use these models, right?

this is one of the good strategies that we can do it.

Another one is type of the data that we are sharing.

It's most of the data that we have currently and we are not using, and I

call it dark and I call this data as dark data, is those data that we are not paying

attention to them and we don't mind if we wanted to share them publicly or not.

Like let's say in my breeding programme, I'm selecting, I'm working with

200,000 individual plants, right?

And I'm evaluating all of them, but at the end of the day, I'm selecting 7,000

of them and I'm cheering for one of them.

But how about the rest?

the rest of the data?

What, what I'm doing?

I'm just saving them, storing them as a archive.

archiving them.

So, okay.

If I wanted to build an AI or in what, like, you know, decision, if I wanted

to include them in decision making process, do I really mind about it?

No, I don't mind it.

I wanted to help on that case.

So there I, I'm sure that there are so many researchers like me, they wanted to

share such this kind of information, but at some point they need to have, like, we

need to have a good discussion around it.

We need to come up with a good strategy and also we need to have

a good transfer learning in terms of the knowledge that we have.

Good on Jen.

Donald Killorn | PEIFA: is funny.

You, you guys are funny.

Yeah.

You guys are funny.

Keep it civil,

Donald Killorn | PEIFA: anyway, no, the, no, I mean the, you know, academia,

there's a lot of like, you know, caution I understand that, you know, there's

also a lot of fervour behind the scenes about, about trying to position Canadian

agriculture and, and acade academic data management as part of the next policy

framework in agriculture in Canada.

And, um, you know, caution is important and that's sort of

always been our thing, is that data governance is gonna be, become the

most important thing to profitability and business risk management.

And, uh,

And, uh,

Donald Killorn | PEIFA: and here we are.

And so, you know, we've got two years to figure out the next policy framework.

Uh, nobody can,

can can

Donald Killorn | PEIFA: up with

up

Donald Killorn | PEIFA: effective business risk management, um,

um,

Donald Killorn | PEIFA: strategies, you know, both sides are

unhappy with the current state.

And,

And,

Donald Killorn | PEIFA: you know, the on the key is, is in better data

processing and analysis and, uh,

and, uh,

Donald Killorn | PEIFA: of,

the use of, um,

Donald Killorn | PEIFA: these new software tools to, to

deliver that on a massive scale.

And, and a great example is

is

Donald Killorn | PEIFA: credits.

I mean,

I

Donald Killorn | PEIFA: long as I've been in Canadian agriculture for

five years, I've been waiting for environment in climate change Canada

to release the protocol for making regulated soil carbon credits in Canada.

They seem to be completely unable to.

And our agentic AI system has developed a, a very clear pathway that not only

satisfies Vera, but also keeps us from having to deal with the licencing

of third party models and, um,

and,

Donald Killorn | PEIFA: and are able to deal with the standards and, and

parse them and, and deliver real,

real,

Donald Killorn | PEIFA: data requirements to actually deliver

environmental goods and services credits to, to the marketplace.

So if I could, I just, I'm gonna interrupt here because the initial

question that Jesse asked was, are we at risk of leaving some farmers behind?

He, he said it much more eloquently, uh, when it comes to ai and then

we dove into an entire discussion about data and decision making.

And what I find super curious is when I talk to farmers about ai, they get

nervous If I talk to farmers about data and decision making, they know

exactly what they need, what they're nervous about, and the kinds of

data they need to make decisions.

And so I'm wondering if we're just having parallel conversations

it's just the words we're using that are making people nervous.

And I'm curious to think if you think if we changed our dialogue or how we

are defining some of the words, if not really leaving farmers behind when it

comes to ai, and in fact, they could be further ahead than most people.

That's, this is something, uh, sorry.

Sorry, mom.

I just wanted to, uh, yeah, sorry for that.

So this is Jen.

No.

Jen.

This is exactly something that I'm really advocating for that, and I just wanted

to make sure that farmers and growers are very well, like, you know, understanding

what's the use of ai, how they can use the AI for the, for their betterments of their

lives, not just replacing their job or even like threatening any of their lives.

And this is something that I, I I, I, I've mentioned at the beginning of

my talk that like, you know, instead of make them how AI can be a good

decision maker, how AI can be a good, like automating something, or that we

can tell them how AI can help them.

instead of just as I said, you know, instead of finding information for days,

they can find any information on a second.

Instead of just calling me every time or calling any of the persons every time

I, maybe I'm not as a person, I'm not, I may not be available on that time.

They can ask an AI and getting help like, you know, from ai.

These are the come of some of some of the things that they can use it.

And once they are started to use AI in this way, they are getting adapted.

And they are, I'm sure, and this is exactly the same thing that happened

with the Ontario bean growers.

They are very well, like, you know, I'm so happy and I'm so proud that

I'm working with this community.

They are very well, they are awesome Ontario bean growers.

So they adapted this new technologies and gradually right

now they are advocating for ai.

So this is something that we built based on not a very good trust at the

beginning, but we started to use it and test it to see how it goes, right?

And it goes perfectly right now.

And we are both are in a agreement that we should advocate for AI for

the next, like, you know, future, uh, whatever, like projects or subjects.

So this is something that I'm trying to, uh, do it, uh, when, when it comes

to the farmers and growers, not only for decision on how I wanted to use

the data or how I wanted to keep the data or managing data, but also try

to make AI simple and understandable and try to assure everyone that AI is

here to be a good tool for all of you.

So I, I'll jump in.

And, Jen, I think to, to answer your question, I think the, the term

artificial intelligence is a cliche.

And the reason I say that is it has been, been mentioned so many times

in so many capacities that it's lost.

Its like real definition.

And if I were to put a definition for artificial intelligence, I

would say it is advanced statistical pattern recognition, at scale.

And there are multiple ways of achieving this through tools, like large language

models, through neural networks, through, uh, a plethora of different tools.

So when it comes to the conversation of ai, because at a really high level, it's

been talked about in really broad terms, it seems as if, you know, farmers have to

walk through this new paradigm when they are trying to operationalize this, these

tools on their farms, for the most part, in my opinion, they are using elements

of artificial intelligence, albeit, okay, they're not using, they're not

maybe building their own neural networks and processing billions of images at a

time that, but they are using, again, introductory tools at the very least in

order to evaluate different decisions.

Now, whether artificial intelligence or the data they have on their farm

will enable them to make real time decisions, that's a, that's a journey

every farmer is at a different stage on.

So when it comes to using artificial intelligence at scale, it really

depends on the tool that they have.

And the thing I would, I recommend to producers is don't

buy something off the shelf.

If you think that you're going, because you think you're missing a gap on

your farm, use your, the strength that you have on your operations

to fully utilise these tools.

So for instance, if you, as a producer, again, let's take Jesse for instance.

Jesse is an operator and he has a really keen ability to maybe forecast

yields, and he has a data set that helps him every year build on that

expertise as opposed to introducing a new tool to enter your operation.

And I think that's the, the, the evolution, a lot of producers at

a are at a different stage yet.

And I think that's part of the conversation as to why AI is a bit

convoluted these days is because we just put any tool that we're

seeing, uh, you know, mention all the news into this bucket.

And it's, that's why it's become a bit of a cliche and it's sometimes

why some producers are a bit apprehensive to have that conversation.

But when I talk to a lot of producers, even using your mobile

device is a form of using ai.

If we're gonna be nitty gritty about those details.

Donald Killorn | PEIFA: mobile devices are great Edge computers.

That's, that's a big, uh, like, that's a big, um, rallying statement in our

shop these days is that every farmer's already holding an edge computer.

Um, you know, the exciting thing is agriculture has a, so much to

make up or, or sort of has missed kind of Web 2.0 and, and a lot of

cloud-based computing opportunities.

And so managing a trillion dollars worth of capital.

Um, and they're doing it with, systems that haven't been fully modernised,

uh, compared to other serious capital buckets and, and supply chains.

And so if you boil it down to it, I still think that this moment is about

to Muhammad's point about advanced statistical probabilities and at

scale, agriculture's a perfect example.

And bringing the ability to collect, uh, analyse and manage data effectively

and scale, um, scale up our ability to learn from that full dataset through

a, a sovereign compute strategy that, that keeps, you know, effective gates.

That's how we fix business risk management.

It's business risk management is slow backwards looking.

It doesn't incentivize resilience.

And in order to do those things with a national, uh, business risk management

programme for Canadian farmers, we have to use cutting edge technology to

collect, manage and, and analyse data and, uh, deliver insights both to the

farmer and uh, to the lenders and the insurers and, um, and the marketplace.

Well, and, and Donald, let, let me get you to ho

in super quick.

ahead.

Go.

Go ahead, Jesse.

Go ahead.

Well, I, I wanted Donald, you, you evoked something really powerful

earlier when you talked about, you know, the way in which your agents were

outrunning, uh, environment Canada's carbon credit system, to, to what

extent is the agentic era accelerating everything we're talking about?

Like, on the one hand, I think Mohamed, you know, quite correctly was

pointing out the win, which a range, an ecosystem of technologies are

being collapsed into a single phrase.

But the agents now are a, a, a, a driving part of that ecosystem.

It's a part you're playing with, I, I think, uh, in,

in a really interesting way.

So take what you just described and throw agents into the mix and

how do we ground that back into the AI for all kind of rallying cry.

Donald Killorn | PEIFA: Well, the models, you know, the, these models

and, and I agree with Mohammed that artificial intelligence is cliche.

I think that's very well put and that it encompasses many

different types of technology.

But the, um, these, a, these mo these age agentic approaches to

utilising the models is really about giving them managing context.

So you're, you're trying to create boundaries in which they're

responsible for in an effort to manage their context window and

produce results more efficiently.

And efficiency, you know, is why we need a sovereign compute

strategy for agriculture.

We need to under, you know, research, conduct, research and development on how

to identify and move the data around and ensure that we can maximise the benefit

of, of all agricultural data while, um, operating within the bounds of good

taste when it comes to data governance.

Um, so that ag agentic moment, we saw a lot of, saw a very rapid movement

between, um, Claude code and, and the ag agentic approach to utilising

those frontier large language models.

We're now at a point where the frontier models already are becoming,

uh, prohibitively expensive for the type of day-to-day modelling that

we were using them for in February.

And so we need, and not only that, but they're a huge risk to data governance,

uh, and, and data sovereignty and need to be utilised appropriately

and, and, and effectively, uh, to manage cost and, and manage, um, risk.

And so now the, you know, we're, we're, we're working with to, to Mohammed's

point, we, we match these frontier models with the, the research loops.

And, and really what, what's happening now is trying to get these systems

running, um, fully air gapped, you know, we've got this mocked up air gap

system using a series of Mac minis and these days at, at the Federation, it's

all about bare metal, you know, and, and getting the models on bare metal.

And there's no question that there will be a hardware device designed

that combines a raspberry pie and a wifi gateway with, with a, an

agricultural, um, inference model.

And that that will be sold to farmers in the next, you know, two to three years.

That that hardware, software, um, to facilitate the field boundary management

of data, um, is the, the whole, that's where the, that's where LLMs

are headed across the whole economy.

And, uh, you know, this is as an exciting a space for this technology

as any, uh, in the Canadian economy.

So, uh, ma Matt, I, I just wanted to, uh, ask you another question.

So would you please again define what is artificial intelligence?

What did you say?

Sorry, for what?

How did you define it?

yeah.

The definition I've been using just as a, a, an approachable standard

is it is advanced statistical pattern recognition at scale,

Okay.

um, leveraging different tools like large language models

or neural networks, uh, in order to uncover insights, forecast or,

um, tech patterns within datasets not discernible to the naked eye.

Um, again,

So.

my, I'm just, I'm just riffing here.

mentioned, so you mentioned advances, statistical approach, but

my question is, do you think, uh, simple regression, linear regression

model is a part of AI or not?

You know what's funny Mo then, um, the Ontario Securities

Financial Institution, whatever it's called, office fee, would consider that ai

it's not, it's not only Ontario, it's the science.

It.

Yeah, and it

ai,

the

one of the, one of the aspects, one of the subjects

of the AI is machine learning.

And on the machine learning we have supervised learning.

And on the supervised learning we have the regression models.

And on the regression we have the simple linear regression.

So AI is not only pattern cognition and AI is not like classification.

We have so many areas in ai, you, when, when we are talking about ai, it's, it's,

it's exactly look like that You talk about agriculture, but what's inside agriculture

is plant breeding, plant physiology, geneticists, you know, all those areas.

I, I, I, to be honest, I cannot say AI is the advance advanced statistical approach.

I can, I can refer to what is AI is based on the IBM that provided as artificial

intelligence, the technology that enable computers and machines to simulate human

learning, comprehension, problem solving, decision making, creativity and autonomy.

This is something that I can refer what is ai, but, and it comes to the

advance, I'm, I'm, I'm against of that.

I don't, I don't think that, you know, AI is the all about advance.

It's

So.

and ai, this is something AI established in 1915.

don't want to split hairs on like a definition.

I think my intention with the term advanced is not as simple.

It's not, it's not, it's not binary, right?

It's just, it's, it's more of like, you're using sophisticated ways to get to,

let's say, um, an insight for instance.

let me.

and it's like the language I'm using isn't

Donald Killorn | PEIFA: Co codifying, codifying language and, and,

and, and conducting modelling and being able to, um, I, I'm not,

I don't think it's particularly elegant, but, you know, neither

are, um, power lines, you know, so,

But let me, let me refocus.

Donald Killorn | PEIFA: but, but breaking language into code, the same way we

broke genetics into code, and we have economics, you know, analysis is, is

So.

Donald Killorn | PEIFA: you know, an advanced piece of technology.

There's no question about it.

Like it, it

We're, we're.

Donald Killorn | PEIFA: one of the main three languages, like

the three, three codes that humans use to create the associates.

So let me, let me bring this back.

Hold, hold, hold on.

You know, Mo Molsen, I think made a really important point, which is, if we

think of AI as a spectrum rather than a binary, then that spectrum on both

sides goes much farther than a lot of people really entertain or are aware of.

And, and one thing I I really want to bring the conversation back to,

because Donald evoked this a couple of times, you know, while really fleshing

out why sovereignty is such a crucial part, uh, of this larger conversation,

I know that the vast majority of people do not understand what, what's

involved in training a model, right?

They hear about these models, they hear about Mythos, they hear about Fable,

they hear about, you know, a, a, a deep seek, all these new models out there.

But we do have two people here on the call, maybe three, uh, who have

actually trained and built models.

So break that down again, uh, not just for Jen's benefit, but I think for a

lot of people's benefit when we start to imagine a, a sovereignty, a Canadian AI

ecosystem that runs in Canada, trains in Canada, infers in Canada, how difficult

or easy it is to train and build models.

Moen, I'll go to you first, but then Mohamed, obviously, I, I, I want you

to jump in when you, when you see fit.

Sure.

That that's a, that's a great question, Jess.

The point is, uh, it, it totally depends on the use of the AI models

and after, like, you know, in what area, in what like, you know, aspects,

do you wanted to use AI models?

And then I can tell you how difficult it is for training the models.

I'm, I'm thinking as a, like, as a person who wanted to do the research, and I

can't give you a very immediate answers, but to be honest, like, you know, as

long as we have a broad spectrum of the data that cover all the instances that

we are facing and we have faced before, I think that would be a good, uh, like, you

know, a database for training the models.

Let's say if I wanted to make a decision about like plant breedings, I'm, these

days I'm trying to move toward the digital twin and developing digital

twins for plant breeding, instead of just wasting so much of energy and like, you

know, plants in the field, I can just simulated what will happen in terms of

the future climate change scenarios.

So in this case, what I fed my models was the 50 years of the databases that all

was grown in different areas, locations in North America, and I collected all

this information and train, and it still is not enough because I can't predict

what will happen in the future, really.

It's so hard to predict, right?

But this is one instances.

Another instances is I, I'm usually, I, I wanted to use AI to predict

the pattern on the maturity of the pattern of my dry beans, right?

So the maturity pattern are usually, usually consistent.

So if I have a couple of years data and I can include it to the training,

the models, that would be enough.

So the major, the point, the question is and how you wanted to use this model.

And then training is based on like, you know, training the amount of the data

that you need to train is based on what, where you wanted to use this AI model.

So, Jesse, can you repeat your question more time?

I just wanna make sure that, uh,

So

all the

well, what's involved in building a model?

I, I think when people hear the phrase AI sovereignty or Canada needs to

control its AI infrastructure, they don't totally understand the pieces.

And obviously the model, whether the r Tom's points specifically the focus or the

capabilities of the model is important.

So again, from a big picture, from a leadership perspective, what's

involved in training a model?

And is that, is that the kind of capacity we can and should

be developing across the sector?

Okay, so I'm gonna cheat a bit, uh, and I'm

gonna take it further than that.

But again, to my two colleagues on the call include where

you think I'm messing up.

But essentially, I'm gonna keep it simple, but essentially, I think the first thing

when it comes to model development, and this is for you Jen, I mean, like

the way I think about it is like, let's define the question, um, what are we

building the model to predict or decide?

And that's where we would start things off from.

Um, for instance, like when we're thinking about agriculture in particular,

like will, like can we build a model to predict if the yield will drop

below a threshold, for instance, based on a lot of different elements.

That's maybe the question we want to answer.

After you have that ref question again to the best of your ability.

And again, that question can also change.

This is a dynamic process.

You would gather that data.

And I think when people say data is the, the new oil, like that's the

point that I think this is all about.

It's because when it comes to looking at forecasting or trying to uncover insights,

more data you have historical, or in terms of let's say when you're looking

at yields, do you have weather data?

Do you have market information?

Do you have, um, information by activities and the impact of those activities

like a farm activity, like cover crops, or let's say no-till for instance.

How is that going to impact the model ultimately so that you can answer

that initial question you have.

And then of course, once you have that foundation of data, and it can

come from different platforms and the reason why, and just to go little

sidebar for a second, the conversation about that ecosystem play where all

three of us or all five of us on this call can share information freely

amongst each other, which is what we call an API, it's basically a digital

handshake for lack of a better term.

If we're able to share that information amongst us all, that actually enables us

to answer the question more accurately.

In my opinion.

of course takes a lot of training.

It, it's not easy, it's gonna take time as well.

But once you have an ecosystem where you can share that data, you're actually able

to get more information that you can, you can incorporate within your model as well.

then when you, once you decide on the appropriate data set, once you clean

that data as well, because Jen, I can't, I can't emphasise enough how many

times I've seen glyphosate spelt in.

I've seen it maybe spelt in a hundred different ways, like

I've seen it with the U once.

So you have to go through that process and maybe make sure

that the data itself is clean.

then like, I guess like the last two steps is then you have to test it.

And then you have to measure the confidence you have in the results

you get from that information.

And once you test it, then you can, uh, deploy and monitor it as well just to,

again, when you use, when you deploy it, you you're never going to have

it percent effective from the get go.

You have to improve on it.

Now, the part that I want to cheat a bit about is the concern I have

within the industry these days.

Um, and what we're seeing is, and not in a Canadian context.

I think agriculture for the most part has been forward looking

when it comes to sharing data.

And all five of us on this call have, are again, are building protocols to enable

ourselves to share that information.

And I think that is amazing.

The trend I'm starting to see, and this is coming from the equipment manufacturers

again, like tractors and whatnot.

They're starting to charge their users money in order to not only collect that

data, but then even to share that data.

And the way I would it in an analogy is it's like you're going to a pizza

parlour, you're paying for the pizza, and then you're getting charged to

share those slices to your friends.

And that's not really building an ecosystem then.

And that's the big concern I have.

So when it comes to that foundational model building,

that's the way one were to do it.

And the way we're able to maybe cultivate that ecosystem so that we

can share better and more pivotal information amongst one another.

So I try to keep that short.

No, that, that was fantastic.

And Donald, it, it really, uh, allows me to bring you in as someone who I

know is following news, uh, in the AI world, although I suspect all

present are, and with the removal of the, uh, uh, mythos class models of

Fable and, um, the, uh, larger one that was per, uh, made available.

There's a lot of people talking about open source.

There's a lot of people talking about sovereignty.

To your earlier point, the price of accessing the frontier

models continues to go up.

Is there a critical mass now, not just of collaboration as I think Mohamed was

inferring, but of active data sharing, of creating, uh, an ecosystem of models

that are more accessible, more affordable, or, or am I just dreaming here?

Right.

I I'm curious how, if you could connect some of these dots for us,

uh, especially grounded in, in the work that you're actively doing.

Donald Killorn | PEIFA: Yeah, I mean, ev it, this is a great panel with, with

an NGO rep and an, and an industry rep and, and an academic rep. And, uh, all

those sectors along with government have to work together to build more

sustainable systems, whether it's the food system or another complex system.

Um, and that's what, what we'll find here.

Um, I, I think, I personally think the, the whole mythos rollout is,

um, marketing, you know, up to and including this, um, is impressive

and, and sort of aligns with this kind of massive um, growth strategy

around, um, you know, picks and shovels and, um, what you can do with them.

Um, so yeah, gold Rush is tough to manage, and, and you never know who's working with

who to make sure that this thing continues to drive new records in the stock market.

but there's no question that, our future in, in agricultural data,

uh, means ESP at the farmer level.

And motion.

Motion has said a few times about how fast agriculture is and the different

implications for data in agriculture.

And so I do wanna just say that, you know, from the, from the grassroots at the farm

field level, some data will leave the firm, the field, I should say, the field.

Some data will leave the field only if it empowers the rest of the model, if it

empowers the model to grow more stronger.

For the most part, uh, decisions will be inferred at the field boundary

based on, uh, lightweight models that can be housed in edge computing.

And, uh, I suspect again, that there'll be hardware development

around that for farmers.

But, um, we will not be moving all of the data off farm.

Uh, we'll move hashes to the blockchain, and we may move the data into the cloud

if that's what the farmer wants, and, and they're willing to pay for that.

But, um, the effort to, um, collect and manage and analyse and build

decision making capacity for farmers at the farm field level, um, we cannot

move all of the data outta the field.

Um, it doesn't make sense for data security, and it doesn't make sense for,

um, a practical, efficient use of, um, of large language modelling and, and, uh,

building, uh, decision making capacity.

Artificial intelligence, it's, it's, um, know, we can outsource intelligence,

but we can't outsource understanding.

And so we will deliver intelligence to the farmers, uh, in a way that

they can, uh, uh, utilise and then they will interpret it and, uh, and

implement it, uh, on their land.

And do you wanna just take a moment to unpack edge computing and 'cause

again, I think it's one of those phrases that those of us in the know understand,

but you've made a really good argument today as to why it is an essential

element of any agricultural AI strategy.

Donald Killorn | PEIFA: Yeah.

Yeah.

We, we need $150 million, uh, to do a, a few years of

research and development here.

And, you know, in around 75 or 80 million of that has to be in equipping

our farmers with edge computing.

And it should have been part of off calf, it should have been

part of the last policy framework.

Um, but we have to build an intelligence layer, uh, right on top of the field.

Um, and in order to do that, we need edge computing, which would be any form of

compute that's found sort of in the, in the three dimensional world, like in the,

in the space and time as we understand it.

Um, so that edge can be sensors, it can be cameras, uh, it can be GPS,

um, you know, that's life on the edge.

It's, it's the, um, it's your, uh, real, your connectivity to the real world.

Uh, as much as we want to use our 50 year bio geochemical model, and, uh, as much

as we want to use our climate modelling, uh, to help build decision making to e

everyone's point today, uh, we need to mix that with what's happening on the farm

right now, uh, with those outlier events and, uh, make sure that, um, the models

are ground truth effectively, constantly being updated as the climate changes.

And, uh, that they are equipped to not just deal with the status

quo or the average, but with, um, what's actually happening on farm.

And that happens through the, through the edge layer.

Right on.

And we are, uh, just about out of time.

So I'll, we'll take a moment to do kind of forward looking or speculative

thoughts on where this is heading.

And, you know, Mosen, I'm gonna, I'm gonna throw to you first only 'cause I just

read about, uh, you, uh, uh, alluding to being GPT expanding beyond Beans and

expanding into other research areas.

So perhaps use that as a way to talk about where, where you see the kind of near

term in terms of this stuff evolving.

Yeah.

That's great.

So, uh, so before I talk about like, you know, what would be next, uh, after

being GBT, I just want to say like, you know, uh, I totally agree with Donald

and Mohammad, what they said, what they mentioned about like, you know, data

management and also how to get this data because I think AI is been, has been here

for like more than seven years, but the major problem right now that we are facing

is how we can obtain the data and how we can share to what extent we are able to

share the data and to what extent we can ask farmers to produce this and collect

this amount of data and make it accessible to us for developing different models.

But anyway, so in terms of the being GPTI initially, like, you know, we initially

created being GPT specifically for beans.

But right now, after I had this, uh, good like discussion after the presentation

that they had last week for in, uh, international Legum conference, uh, I, I,

I found a very great interest on the use of binge GPT and they said they wanted

to like, you know, uh, I, I received a good like, you know, uh, interest in

terms of, uh, advancing binge GPT beyond beans, like, you know, on the lentil, on

other legumes and also on other crops.

Definitely possible, possible like, you know, to extend it to another crops.

So we wanted to do it.

And also another feature that we wanted to add is the powder recognition.

Definitely wanted to add it to the BGPT as the be vision that everyone

can just take the picture of the raw seed bin on ZI bin and then, you know,

post it on the BGPT and binge G PT can tell them that what with the quality.

But this is all about the BGPT.

But the future of AI I think's going to be more on, uh, um, quantitative ai.

So I think, you know, uh, I think it's going to be more on, oh, sorry.

I'm just saying, uh, it's going to be more on the quantum quantum approach and in the

next couple of years I think we are going to shift toward the quantum approaches.

And I'm seeing that a couple of interest around that they are sorting

to adopt it, the quantum approach.

And again, quantum approach is not something new.

It's, and it'll be new like in the couple of years everyone will

talk about the quantum approach.

But I think, you know, the future of AI is belongs to quantum approaches.

Yeah.

Mohamed.

I'll let Donald go.

Okay, Donald.

Donald Killorn | PEIFA: Yeah.

And, uh, yeah, the, the Bitcoin owners are definitely focused

on a quantum approach for sure.

The, uh, and farmers will be interested to see quantum is, you know, it's

very, it's a huge part of if this thing is gonna get, get somewhere.

But anyway, um, yeah, that's very interesting.

I, I think that, um, you know, I haven't made any, any big, um, secret about

where I think we're headed, I guess in the, in the near term, like, uh, we

need a tiered sovereign architecture.

Um, we need the edge layer to be a secure public gateway

back and forth from the field.

Uh, we need the compute layer, uh, that has that frontier AI capability for when

we need advanced reasoning and, um, AI assisted automation and decision support.

Um, high performance retrieval operation, scalability.

And then the key is the sentinel layer that's sovereign data control.

That sentinel node that sort of is the traffic cop, uh, that allows

us to unlock, uh, the full benefits of the edge and the compute layers.

Um, that third piece, um, helps us use Frontier Cloud ai, but also ensures

fully sovereign local ai, uh, for those sensitive or regulated workloads.

Um, I think, you know, mixing all those together, um, and, and, and ensuring

that we're focused on, sharpening our pencils when it comes to, how

much compute we need and how, and therefore how much electricity we need.

'cause there's no question there's a bottleneck in electricity.

And so being as efficient as possible will help us execute on our, on

our sovereign compute requirements.

Then I'll go, um, I guess wanna make you richer, Jesse.

That's, that's where I want to get to in the next five years.

And if we can use artificial intelligence to get you there, like the different

tools, uh, that's the way I wanna do it.

And I think, you know, I think again, the tools that are encapsulated within

artificial intelligence will move away from a tool that's an individual op

from individual operations to adopt.

Um, that, that, that are adopted in isolation actually to the connective

layer that connects equipment, agronomy, lending, uh, and markets

into a single intelligence system.

And it's through the use of agents as well.

We didn't get into too much of that today, but maybe we can assemble all the

Avengers again and, uh, talk more AI soon.

Yes, absolutely.

If, if in no small part because of the rapid rate of change that I think

has been one of the undercurrents, uh, to our conversation today.

And that is where I think the original challenge that, that Jen paraphrased

of making sure that no farmer is left behind is going to be an ongoing process.

Uh, and that's where your respective leadership is indispensable.

Jen, any final thoughts or comments from you before we conclude?

Yeah, so many.

Uh, I have a lot more questions, so I would love to assemble

the Avengers once more.

Um, you guys have taught me a lot.

I have some deep thinking to do,

Right on.

in touch for sure.

Right on.

Thank you very much.

Good.

And again, thank you.

Uh,

again,

will have to on Mohammed's call, uh, uh,

reassemble, uh, some point soon.

I will work, uh, have my scheduling agents talk to your scheduling agents and

we will be able to work something out.

Thanks again, everybody.

Donald Killorn | PEIFA: Thank you, Jesse.

Thanks everyone.

Talk to you soon.

you.

Thank you.

later.

Take care.

Bye bye.

Future heard assemble.

What a fantastic session that quite frankly exceeded the already

ambitious, uh, expectations I had when gathering together the

kind of bright minds that we did.

Now granted, this is the kind of episode that we could have gone on,

uh, for another hour, maybe even two.

I really do try to limit these episodes, uh, both for my own, uh, biological

needs, but also so that we can allow these conversations, uh, to, to uh,

cultivate with time to like fine wine, continue to ferment and allow

that knowledge, uh, to multiply.

' cause there were a lot of threads that came up today.

In particular, Donald kind of evoking how language is really at the core

of the current revolution in ai.

And it's our ability to use our vocabulary, in particular, an agricultural

vocabulary, a business vocabulary, a natural vocabulary, biological chemical.

All of this suggests that we're entering into a world where it's

not just about talent, it's about collaboration amongst the talented.

And granted, the agricultural sector has always valued collaboration.

But one of the very subtle through lines of today's conversation that Donald,

to his credit, kept emphasising is the value of cross-sectoral collaboration.

What happens when you get a researcher and a banker and an association rep in a

room with a hacker, and you start having conversations about what is possible, what

we should focus on, what is important, what is, uh, the source of curiosity

when it comes to not just surviving but thriving in the technological

revolution we find ourselves in?

I'm already starting to schedule, uh, uh, these three wise guys,

uh, into another session where we will go even further because the

AI revolution waits for nobody.

And the question is not whether we need to conform to the marketing

and hype of the hyperscalers.

Quite the opposite.

We can and will bend the technology to our needs, our culture, our values, and

that's where Muhammad Mosen and Donald are really, uh, exemplary of the kind

of leadership we have here in Canada and how that can be applied to technology

to make the technology work for us.

So thanks to them.

Thanks to Jen, thanks to you, uh, for listening, for watching, and for sharing

this with your friends who are wondering why AI should be taken seriously, why

we can be both critical of it, but also bending it to meet the desires and dreams

that we as a society, as a culture hold.

Uh, so thanks again.

Uh, we'll see you soon here on the Future Herd.