FYI - For Your Innovation: Coordinating Technology and Specialized Labor for Complex Needs with Francis Pedraza

Welcome to FYI, the four-year innovation podcast. This show offers an intellectual discussion

on technologically-enabled disruption, because investing in innovation starts with understanding

it. To learn more, visit arc-invest.com.

Arc Invest is a registered investment advisor focused on investing in disruptive innovation.

This podcast is for informational purposes only and should not be relied upon as a basis

for investment decisions. It does not constitute either explicitly or implicitly any provision

of services or products by arc. All statements may regarding companies or securities are

strictly beliefs and points of view held by arc or podcast guests and are not endorsements

or recommendations by arc to buy, sell, or hold any security. Clients of arc investment

management may maintain positions in the securities discussed in this podcast.

Hello, and welcome to FYI, the four-year innovation podcast. My name is Brett Winton.

I'm the chief futurist at Arc Invest. With me today, I have a returning guest, Francis

Petrazza of Invisible Technologies. Francis, hello.

Howdy.

How are you? I like your hat.

Thank you. I'm a Texan now. I love New York. I'm an inflation and tax refugee, and now

I'm in Austin, Texas. I'm a resident.

It's not a bad place to be. I hear it's a growth place. Lots of people coming in.

One of the fastest and not the fastest growing area in the country.

All right. Why don't you remind us for those that didn't listen to the original invisible

episode? What is invisible? What do you all do? Why is it being used?

Invisible is a new way to run your business. It's operations as a service. We provide

both AI training and AI enablement services to the fastest growing and most innovative

companies in the world, including OpenAI, Amazon, Microsoft, and AI21, Coheir, and others,

who demand extreme capabilities and efficiency from their digital operations.

Wait.

We're counterpositioning.

Pause, pause, pause. I think that answer has changed since we last spoke. I thought of

invisible as, hey, this is like mechanical turk, but done well. You have outsourced agents

doing things, but that did not naturally lead me to expect that you would be contributing

to AI training, for example. Can you expand on that?

Yeah. Understanding what we do requires a bit of an imagination. Suppose there's a black

box that could efficiently execute any complex process at any level of scale. What would

be the limits of that black box? There are use cases in every industry and function. It

just so happens that AI training turned out to be the biggest use case that we have right

now within the last two years that exploded as a need.

In terms of AI enablement, we were always an automation company, but AI has become the

primary driver of automation. We were previously selling automation to the enterprise. Now,

we're just selling AI-driven automation to the enterprise. We're counterpositioned against

20th century services companies like Accenture. They're not technology companies like us,

and they don't run all of their operations on their own proprietary technology platform.

We've built a digital assembly line that solves a complex coordination problem. How do you build,

run, and scale end-to-end digital processes with the right mix of specialized knowledge work,

automation, and AI? That's the problem. We could use some examples here like DoorDash

asked us during the pandemic, hey, can you digitize every restaurant menu in the United

States and help us onboard restaurants onto the platform? We were able to do that.

We help insurance companies process claims. We help hedge funds automate trades.

These are just examples. Every business has these complex end-to-end digital processes

that they either have to do in-house or outsource to these 20th century

non-tech services companies that bill by the hour and just have an incentive to be inefficient.

Wait, all of these things that you've just mentioned, if I were a business and I were

approaching that problem today, for instance, DoorDash, wouldn't I say, hey, this is an AI

solvable problem? I don't need invisible anymore. I'm just going to dump it into GPT-4 and it'll

solve my problem. I don't need energy in the loop. Isn't that the way to do it?

So we like to joke. There's an app for everything, so why isn't everything perfect?

That joke is the essence of our business and you just said it perfectly. Everyone who's actually

dealt with complex operations or actually tried to use all of these amazing technologies knows

that the soundbite that you hear on stage, the pitch, it's not a magic wand in practice.

In practice, you have to stitch together. We integrate 300 different tools into our platform,

so you have to break down your process into steps and each one of those steps

needs to be approached as a different problem. So this step is best solved by this tool and this

step is best solved by this AI and this step is best solved by this automation software.

And this step needs to be done by a PhD and this step needs to be done by just any college grad.

And so we coordinate 2,500 workers in over 90 countries around the world.

Believe it or not, if you had told me this earlier on, I would have been surprised.

The United States is 20% of our workforce and our biggest single country, but we have a large

presence in countries like India with 9% of our workforce, Europe over 8%, Kenya around 8%,

Philippines around 8%, Nepal, Ghana, Argentina with lower numbers. So we have people all around

the world that have very specific skills and inevitably any complex process can't be fully

automated. Otherwise, there'd just be no humans in the economy. So what the economy is actually

doing is using technology to continually push the frontier forward of what technology can do,

which sort of moves humans continually into what technology can't do. And that dialectic is sort of

never-ending. So to actually deliver futuristic capabilities, you need to have a synthetic

approach, not just an AI-only approach. We call it synthetic intelligence, something that combines

humans and machines into one end-to-end solution. So okay, I guess my prior would have been,

hey, AI gets more performant, that should actually hurt demand for invisible services

or invisible technology. You're kind of counter positioning that. Do you have evidence? Is this

true though? Is it true for your business? Correct. I think one way of putting it, Brett,

do you believe we're going to live in a world with one technology or many technologies? That's

sort of the step one on the thesis. So we're a bundler. As I said, we integrate 300 different

technologies into our end-to-end solution and we integrate 2,500 people. So do you believe in a

world where all of that is going to consolidate into just open AI or just barred? Or do you think

we're living in a world of technology proliferation and technology specialization? If you believe

we're living in an increasingly fragmented technology landscape where there's more and more

startups, more and more venture funding, more and more companies that are building tools that are

specialized AI, specialized automation tools that are solving very specific needs, then in that

incredibly fragmented technology landscape, invisible comes in and stitches it all together

into an end-to-end solution. And then also, same thing with human labor. Do you believe we're going

to live in a future where humans are irrelevant or where human specialization becomes increasingly

specialized and humans therefore needing complex coordination to actually deliver an end-to-end

solution for complex needs? So that's the thesis. Yeah. I mean, I guess what's interesting, again,

one, you're using PhDs to digitize door-dash menus. This seems inefficient. What are the PhDs being used for?

Correct. The economist who inspired our thinking, I believe I mentioned this in the first podcast,

is Ronald Koze, who 100 years ago was asking questions like, why doesn't one company do

everything? Why are there many companies? And his insight was that between supply and

demand, there are frictions. You have to transact. You have to coordinate. You have to use things.

For us, it's finding the right technology, the right person for any given step in a process.

So we use PhDs for things like, when you ask GPT a question, tell me about the chemical

properties of helium. And you just keep asking it complicated questions about the periodic table of

elements. How do you know that it's not hallucinating? How do you know that it's giving you the right answer?

You've had a chemistry PhD or a physics PhD who's answered, trained the model to do something specific.

Or if you ask it a complicated question about World War II, about the battle of the bulge,

how do you know that it's not making it up? How do you know that it has the right answer?

So there are a number of processes, adversarial training is one of them, that we co-design and

co-develop with researchers at AI companies. And depending on what their goals are for the model,

we help train the model so that it's more and more performant. It hallucinates less and less

and is capable of doing more and more things. So your ability to trust an AI's answer. I was

recently using an AI to build a financial model. And I realized that it's very bad at calculating

IRR. Compounding is very unintuitive, apparently, not just for humans, but even for computers.

So any of your listeners who've actually used AI know that there's a feedback button where you

could say, this is wrong. And so imagine that feedback button was a whole business. That's

what AI training is. Okay, but then that brings up a question because clearly they insert that

feedback button because they're trying to get that data without paying for it effectively.

How do you think about like, this is a big growth area for your business. Clearly,

it's like driving revenue for you. How do you think about that risk of either them using

kind of like their user interface to basically supply the data that you're providing or

them like asking the model to self vet in some way and kind of, why do we need humans anyway

to provide training on these things? From a historical point of view, this question has been

asked since the Industrial Revolution. Now that we've created the cotton gin, now that we've

built an assembly line, why do we need humans? And so there's this continual fear that you're

going to live in a world of mass unemployment. And this would be this version for invisible,

that invisible is going to wake up someday and all of its clients no longer need complex

coordination, no longer need an integrated solution. Our thesis is that we're going to live

in a world not of 90% mass unemployment, but a world where the wizard and the wand are co-evolving.

So if humans are the wizard and the wand is technology, the wand makes the wizard more

powerful. The wizard gives the wand more things to do. And actually, there's just an incredible

essay by Martin Heidegger in the middle of the 20th century. We're almost in a prophetic way.

He was thinking about these things. The essay is called The Question Concerning Technology.

And his idea was that technology is always revealing human potential and human demand

is always driving technological evolution. So that's the sort of our general philosophy.

Now, how does that turn into business practice for us? We really value agility in our strategy.

So if a company comes to us with any operations need, they're like, we have this

really complex operations need or problem. We're almost like McKinsey. We're consulting with the

client. We're helping them design the process. And then we'll invent the process and run the

process and scale the process as quickly as they need us to. So almost like there have been weeks

where we've hired over 100 people in a week. We have the ability to, in theory, just scale up

and scale down really, really fluidly. So that means that as needs evolve across our client

base or in the economy, we can just rapidly respond. And that's a key point of difference

between the typical approach of a tech company. So a tech company usually builds software to do

one specific thing. And if you think about a tech salesperson, they're selling their

SaaS app over and over and over again. Here's what our app does. You're the

thousandth company using it. And you're going to use it just like everyone else does. It's just

a tool. But for Invisible, we've built something that can run any process. So the limits are

sort of the limits of the imagination. That horizontal technology platform and that architecture

that we've built means that we can spin up or spin down any process and apply labor in any

direction as the economy changes. So you're seeing demand for AI training right now,

and then your belief it. So you don't know if this demand is going to disappear at some point.

You don't have a point of view. You must have a direction, though. Clearly.

I have a point of view. So I have a point of view around AI training specifically.

I'll frame it as a bull and a bear. The bull case is that AI training is going to go from

maybe a billion or two of annual spend to 10, 20, some unknown number of many billions of

dollars of annual spend. And it's going to be one of the biggest industries in the 21st century

economy. It's not going away. Wait, let me clarify. You're talking about AI training

generally or specifically the human feedback portion of the AI training? Oh, no. The human

feedback portion. Okay. Yes. Okay. Yeah. The services portion. And that AI enablement,

which is using AI and automation throughout the enterprise, is going to remain a complex

coordination problem that's going to require many different technologies and many different

types of specializations. I'm way more confident in the AI enablement future than I am in the AI

training future. And that's because AI training, it feels a little bit like the internet in 1995,

maybe 1998, where everyone knows the internet's going to be a big deal. And everyone's right.

And everyone knows that they're right, which creates a speculative bubble, which is 1999.

And so the key thing for us is we're a profitable company,

and we've maintained financial discipline. So if there was a downturn, we would just

maintain that financial discipline. We would come out the other end. And so Amazon was able to survive

the dot-com crash. How many companies from the 90s are still around today? Very, very few.

And so we want to be that company where it's like, if there is a speculative bubble

and it eventually crashes, how do we build the company to survive that and capture the opportunity

on the other end, which we believe in? So we believe that 10, 20 years from now,

there's definitely an AI training market. There's definitely an AI enablement market.

But the cyclicality is the hard part to predict. Yeah. I mean, I guess from our perspective,

yes, there's a literally measured in the tens of trillion-dollar enterprise value opportunity

in being the AI foundation model that everything else gets built on. And it's very clear that

the way in which their performance differentiating is very much driven or at least catalyzed by

reinforcement learning with human feedback. So using this human labor to augment the models

is a way to differentiate. And there are, XAI is coming into the market. There's open AI.

There's Anthropic. There's Cohere. There's like the open source guys, Meta and Falcon.

Like there's a lot of people going after what is a very big opportunity. And they have some

amount they're spending on compute and they know to deliver performance. They have some

amount they need to spend on human feedback. And number two cost is training. Number one

cost is compute. Number two cost is training. And you just mentioned a bunch of our clients

and hopefully future clients, near future clients. And that training cost actually

is what makes the product good. It's like a quality thing. Right. And if you look at

what's interesting also, it's kind of like these systems are so complex, they're very hard to

benchmark. It's like, how would I benchmark to know that, you know, how many specialized

benchmarks do I have to have to know that it's good at the periodic table? Right. Like you don't

know unless you're the person that's trying to use it for that thing. And you're like, oh,

this is wrong or this is right. And so kind of like the degree to which they win is in some ways

hard to even know until users are interfacing with it. So I'm going to turn the question on you

and be the interviewer. Like how do you think that the needs here are going to evolve over the

next few years? You know, as these, so I perceive that these companies are very R&D driven right now

at some point. And that's to my point around it being a speculative market right now.

At some point they become very commercially driven, where they're actually responding

to real needs in the market, commercial needs, real clients, real users.

When do you think that shift will occur? And then what do you think is going to happen to

the economic models that are the business models that are supporting this R&D?

Well, I think there's a few ways to approach this question. But one like very tangible way is

if you look at what they're spending, forget training just on NVIDIA chips, NVIDIA, for NVIDIA

to be like a successful investment, just to meet expectations, it has to sell $110 billion

into data centers over this year and next year, 110. And if you think about typical

software economics, that has to be like a cost line into 80% gross margin business. So

there needs to be something like you can do the math in different ways, but on the order of $700

billion in AI software revenue that spills off the back end of those NVIDIA chip investments,

or else there's a problem. And so in some ways, it's like they need to begin monetizing almost

like in relatively short order, just to justify kind of like the capex that they're laying out.

That's $700 billion, depending on how you do the math. That could be like three to five trillion

in AI software market cap. And that excludes Google, which is on its own chip stack. So there is

conditional on like these being good investments for them, both spending on you and spending on

these chips. There needs to be like the creation of multiple trillions of dollars in AI software

market cap. And then here's the key question, because remember, I'm not asking Brett questions

about my business. I'm asking him questions about something that he's an expert in.

What makes the foundational models defensible? How does this not just get commoditized?

Well, I mean, I think that's a subject of like active debate, because you know,

meta is open sourcing them. Like how do you compete against free? And an argument could be you end up

with like kind of a virtuous cycle of data where you're getting data off the user behavior that then

allows them to cut you and invisible technologies out of the process. And they're kind of like

they have enough user inputs to continue getting user feedback on their models and they perform.

Or you get baked into so many processes on the bottom end that it's hard to extract you out. So

you know, open AI is clearly getting commercial traction right now. You can argue they're under

charging for the product. Like if you give a BCG consultant a GPT-4, they're 1.75x more productive

than the average BCG consultant. Isn't that incredible? It's like for $20 a month, that's

way too cheap. Correct. It's incredible. If they wanted to turn on the monetization engine,

they could, but they all see the competitive game of like we need largest possible strategic footprint.

And see, I think that's the argument for continued demand on the training side,

both human feedback and kind of dumping money into computers. Like the opportunity that's out

there is so big that it justifies the kind of like push of chips onto the table. And capital

markets will recognize it and continue to kind of, you know, I mean, ostensibly open AI's valuation

is tripled and anthropics valuation is something like 5x. So you don't take that valuation and not

spend some of those chips. You take that and say, hey, we spent $100 million on GPT-4. What happens

if we spend $1 billion? How good is it? How big is our strategic footprint? So on your side,

from what you're saying from clients and prospective clients, do you see any indication

that they're like, oh, this is a lot of money? Or is it kind of like all gung-ho from here?

Yeah, these are clients that are sensitive to the capabilities and the performance,

not to the price. And that's because there's no other firm selling these capabilities.

We're not doing low-level data training for these AI models or data labeling. We're doing very

advanced, you know, training, working with their research teams, co-evolving processes,

and training the model to do very specific things. And that's very different than like,

you know, commoditized labor. It's very, very specialized work.

If it's so specialized, why don't they just internalize the capability? Like, what is it?

Why do they need you as opposed to being like, hey, we can do this ourselves?

Yeah, so most of our clients realize they have like two alternatives to using Invisible. One is

building an advanced operations team in their own company, which means operations management,

and then actually having all the labor. And the labor would be worldwide. And so you have to

coordinate the labor and build your own proprietary technology platform, build your own digital

assembly line, build your own Invisible, which has taken us eight years to build,

to pay those people, coordinate shifts, train them to do work, hire them efficiently,

align their incentives, all of that. And then you would need to build your own process builder.

And so that when you automate a step, you can use other tools instead of your own tools.

And you need to build your own automation team and your own AI team to build automations

or build software when there's no third-party tool available. So you would need three things.

You would need your own workforce, your own technology platform to coordinate that workforce

and structure the processes, and then your own automation and AI team to build automations where

on the steps that third-party tools can't integrate. And so that's a huge distraction.

If you're an AI company, you just want to focus on building your proprietary models.

You want to focus your extremely expensive, I mean, actually your whole pitch to investors

and the capital markets is that you're investing in a software company that

pays expensive engineers to do miraculous R&D. You're not managing global labor

and integrating many, many, many third-party technologies.

So if you think about an Apple iPhone, it's one of the most valuable products in the world,

but none of the components in the iPhone are necessarily unique.

Samsung can create many of the same components. They've vertically integrated over time, but

where the actual alpha is or the value was is actually creating the integration of those components

and the creation of a user experience that was just so. It was exactly correct.

And Invisible's business model is very much the same. We're an integrator of so many components

that exist in the world, but the technology platform is the thing that integrates them all,

and you wouldn't want to build your own advanced AI and operations team in-house to do that or

your own operations platform to do that. What was the in-house option?

The second option is to go to a 20th century services company like an Accenture, Deloitte,

an Ernst & Young, a PWC, a KPMG, and these companies are huge, like $50 billion a year

of annual revenue, roughly scale, but they don't run their operations on a technology platform.

There's no digital assembly line-powering Accenture, and so they're not necessarily

at the cutting edge of integrating third-party automation tools, building their own automations,

and as a result, they're just slower, and their incentives are more misaligned.

Their incentive is to bill as much as possible, not necessarily be as efficient as possible.

Okay, so imagine that I believe in the AI training store. Well, actually, I guess,

so you differentiated between the kind of like, this is kind of bespoke, like solving a problem

for these companies on the training side, but these companies also rely on a lot of

like data labeling as well, I think. How are they choosing you versus just like kind of a more mass

data labeling type company to do a particular job? Yeah, a good example of a mass data labeling

company is Scale, Scale AI, and obviously, I'm not privy to insider information about that company,

but they've raised and burnt something like $600 million, you know, building this company,

and they have, so they're, you know, historically anyways, have been a loss-making company,

and they've built a lot of their technology or sort of very narrow and deep use cases

around data labeling, and this means that when a client comes to them with some sort of a need

that's outside of that very specific use case, they can't necessarily provide that need,

and they don't vertically integrate their labor into the solution. They just sort of have a

marketplace and a bunch of vendors and people who sign up, freelancers, who are then logging

into a very constrained interface and doing work, whereas we actually have a managed service,

coordinate our workforce, train our workforce, and are fully vertically integrated with our

workforce, so our technology platform is horizontal. We can support any use case, any

industry, any function, and it's not oriented just specifically at data labeling,

although that's something we can do, and our workforce is integrated. Wait, actually,

can you, Claire, Claire, how are you distinguishing between data labeling and other stuff that you

do? What do you mean? So let's think about the use case for autonomous vehicles. If you're like

tagging, this is a human, this is a pedestrian, this is a biker, these sort of basic tagging

operations, if you think, I guess, about what you get questioned on when you're doing a CAPTCHA,

when you're logging in to say your bank account, and it's asking you identify all the

images here with street lights, that's very basic data labeling. When you have an advanced

workflow with many, many steps, and those steps require very high-level skills,

and you have workers that are in that workflow, and helping you upgrade the process, that's a much

more advanced workflow than just labeling some piece of data as this or that, a cat, a cat, or a

mouse. You just distinguish between data labeling and what you all do, whereas I would have thought,

well, don't you just do data labeling? How are you distinguishing between these categories,

and how do you, is what you do different or broader than just data labeling?

Advanced AI training is what we do, and so data labeling, imagine you're logging into your bank

account, and it gives you one of those CAPTCHAs, or one of those image recognition puzzles, where

it's like, list all the bicycles that you see in this array. That's an example of very basic

data labeling, and there's more advanced data labeling. A lot of scale AI's technology is

around supporting specific data labeling use cases, for example, autonomous vehicles. How do

you train all these autonomous vehicles to recognize if that's a human that they're about to hit?

And so we do something quite different, which is that we'll build a full and complex workflow

with a client that has some steps that can be automated, some that can only be done by

like a PhD or a master's in a specific field. And we are training the model to do specific

things, and then sometimes taking the model's responses, and then showing, no, this is the

correct response. These workflows can be any number of steps. They can have any level of complexity,

and they have specific business goals in mind, and we are designing the processes and continually

building new processes and upgrading the processes based on the performance. That is,

not just basic data labeling, and it is human input into the model, for sure,

but it's not just basic data labeling, and that's the distinction.

But it's kind of like, it's a more complex, it's almost like, imagine I'm trying to teach a kid

how to do math, and is the human input I need for teaching a kid how to do math just like

answering what three times 10 is, and three times 11, and three times 12, or just like a very like,

hey, I'm just going to give lots of math problems, and I'm going to provide the human answer to all

those things. Or is it kind of like walking the kid through like, hey, these are steps that are

interesting in a complex problem. And so in some ways, like from the concept of what is this human

feedback providing to the models, it's, as the models get better and more performative, maybe

you need kind of actually a more complex like training operation in order to deliver performance

advantage. Correct. That's right. That's right. That's on the AI training side of our business.

On the AI enablement side of our business, it's more of like normal companies and the rest of the

economy, whatever industry they're in, financial services, healthcare, marketplaces, that have

normal business operations, commercial business operations. And they're thinking about how do

I make these operations more efficient? How do I streamline my business? And how do I use AI?

I'm hearing about AI. I'm listening to the FYI podcast. I know that AI is going to change the

world, but how do I actually use it in my business in this operation on this process?

And that's where we are going to plug in as many software tools as much automation and AI as possible

and whatever steps remain, we're going to run it with our humans in the loop at whatever level of

complexity the human steps require. That's the AI enablement side of our business.

And so like, why should that be you? As in, again, back to the, shouldn't I just be able to plug in

the menu into GPT-4 and say, tell me what the items are? Like, why do I need kind of invisible

for this when I can just sign up for open AI? It's the same dilemma. We know that no single piece

of software is going to solve all the problems. We live in a world where there are many, many

different technology companies offering many, many different solutions. And so you need invisible to

use 300 different third-party tools and have our own in-house team to bridge whatever gaps exist.

And then it's the same thing with on the labor side, which is, you know that there's not one

single type of person who is necessary to complete all the steps in an advanced workflow. You mean

you need different people at different levels of skill with different levels of training. So

your alternatives to invisible, if you're a client, are alternative number one, build invisible in

house. Good luck. It took us nine years. You're going to have to build your own technology platform,

your own operations team, your own AI team, and integrate 300 different tools into your platform

and basically become a competitor to us. Good luck. Or the other option is higher

eccentric. Okay, fine. Yes, higher eccentric. No, but I'd offer like the alternative that people

are probably unintentionally selecting for is actually just status quo ante. Like I have a person

that's kind of doing this thing, you know, and they understand it's like not super productive,

but I don't know if the person who's doing the thing really is incentivized to be like,

hey, this thing's not productive because they're getting paid for doing the thing.

Right. And so it's kind of like, it seems like the alternative number one or the thing that most

people are biased into is not do any of that. They like say, hey, well, let's give this employee

like a general AI tool, maybe they can figure out how to make it more efficient, but that person

doesn't really know how to use it. And maybe they kind of managed to cobble together some things,

but maybe it's just kind of like stick with how we were doing before, but now the person has a great

chatbot they can ask for recipes with whatever's in their fridge. You effectively gave the person

a pay raise. Yeah, they're able to get able to get more done in less time and get paid the same

amount. You're not actually making your business better. Yeah. But that's, that's assuming that

one general AI tool is going to, going to solve the majority of the use cases in your business.

That's just not actually the world we live in today or the world we're necessarily going towards.

We're probably going to a world of many different tools that need, need integration.

Right. And then there's the last mile. We talk about last mile incentives. I mean,

this is also a good, a good time to talk about that. You know, in our, in our business, we have

this philosophy of alignment, alignment of incentives all the way through our business model.

So we try to align our incentives with clients around unit prices instead of hourly rates.

Because once a unit price is established, everyone's incentive on our side, our incentive is to,

you know, do it better, faster, cheaper, because our margins will grow and the client gets,

you know, clarity on how their unit prices drop over time as, as we scale the process.

And then on within our business, you know, the incentives are, you know, for the equity shareholders

to obviously grow the equity value, but even down to the agent doing work, we have results

based agent pay. So the agents get paid more, the faster they get the work done, the higher the

quality is and the more complex the work is. And we were continually trying to make sure that

that last mile incentive does not break down. And I just got an email yesterday from an agent

who is proudly sharing like, I've used this script to create the following 10

shortcuts. And he listed out the shortcuts and I immediately CC'd our automation team and was

like, great, let's, let's use this throughout the, you know, the company. And we're trying to be in

that continual upgrade mode where everyone's pushing for more and more efficiency. It's very

difficult to create that in most businesses. Because most businesses are focused on, you know,

delivering whatever their core good or services, they're not necessarily focused on upgrading

their operation, their, their, the way that they operate. The classic test of this is like,

are the operations in your business on a technology platform, and are all the incentives

of all the, all the workers aligned with, with efficiency. And most businesses would fail that

test. I think it's actually somewhat hard to measure operational efficiency within, particularly

within like a knowledge work type business, where there's like a lot of, you know, it's kind of like,

do you know deeply like how many emails had to go back and forth for this particular step to happen

and, and what the right way to kind of actually do it is. And I think for a lot of businesses,

they don't. And actually from an AI enablement perspective, or what I, one of the things I

find interesting about kind of you all's approach to the problem is that by forcing the business

to processize things more, they actually create more opportunity for automating the entire

operations flow. As in right now, I think there's very little visibility into operations across a

lot of businesses. And so kind of like getting, you know, you, it's not only just, it's like

making the entire operations chain of the business like able to be measured and more efficient if

they go on to your platform. Yeah, by moving on to the platform, you get all this visibility and

reporting. And you can see you could sort of almost like a factory that is also a glass box,

you can look inside the factory and see all the stuff moving around. And so you're, I think

businesses are going to require, you know, small tightly coordinated teams of, you know,

equitized people who are, you know, trying to innovate. But then once you create a process,

and that process needs to be run a million times, you're going to, you're going to want to hand it

off to invisible. I read a book when I was still in college that said, you don't want to work in

your business, you want to work on your business. This is exactly what we want to empower our clients

to do. You don't want to be turning the crank. Let invisible, invisible is a new way to run your

business. Let invisible do all the crank turning. You want to be upgrading the crank, upgrading your

business model. And, and so let us, let us upgrade the operations around the operations and you can

just focus on innovating. So I understand how like the urge to do AI or the, or these new tools,

like as kind of gives a tailwind to enablement for invisible is there, is there like being good

at the training side? Does that enable you to be good at the enablement side? Or do you see those

as separate capabilities? They're interrelated and I think synergistic. So when you are at the

cutting, when you're literally building models and you're at the cutting edge of what the

capabilities are, you also are going to be the best at figuring out how to apply them in the real,

in the wild, in the real world, in the economy. And then, then the sort of the feedback loop goes

back around when you see the limitations, you're better at training them. So, so it's sort of the

yin and the yang of our business. AI training, how many, how many companies are there in the world

that are spending millions of dollars a year on training their own proprietary AI models?

Actually, I'm curious what your answer is.

I don't know. I mean, I think that so Databricks and Mosaic definitely facilitate companies doing

this. And I know companies are beginning to invest in it. And I don't know how many there are. I think

broadly the sense from the from call it the investor, you know, particularly public and private side

is that the boards of directors have all turned around to their CEOs over the course of the last

six months and then like, what is our AI strategy? And CEOs like, what's AI strategy?

But my sense is that that's AI enablement, not AI training. Like, I don't think those boards that

are pressuring, like if you're, if you're any random Fortune 500 company, I don't know, if you're

the CEO of Coca Cola, your board is telling you use AI, but they're not saying build an AI

from scratch. And so, so I currently, let's call it, there's a couple dozen companies in the

market for AI training. And we work with a lot of them. And then there's AI enablement,

that's the whole rest of the economy that is going to be using all these AIs and trying to

figure out what the gaps are, what they can do, what they can't do for my business here in,

you know, oil and gas, or for my business here in solar panels and, you know, California or

whatever it is. And so that's, that's the AI enablement opportunity, which is much broader.

And I think much, much more resilient to cyclical forces in the economy. If there's a recession,

guess what? People still need efficiency. People still need productivity gains. They

actually need it even more. And, and, and all that volatility means that processes are changing.

So AI enablement part of our business is, is, is the true future of our business almost no

matter how big and no matter how attractive AI training is, because AI enablement will always

be bigger unless you believe we're going to live in some future where literally every business is

building its own proprietary model, but that just doesn't make sense to me. I think there will probably

be a lot of growth in AI training, but not, not universality. I mean, if you think about

the, what reinforcement learning with human feedback provides to the foundation models,

or people talk about it as alignment, like, can I make this model helpful and harmless?

As in, can it, does it successfully answer the questions I'm asking? And does it not

lead me down a rabbit hole into like building my own nuclear bomb? Right? Those are like the two

things. And, and, and, and they present that as this is an ethical thing that we need to do to

release these things, but also practically helpful and harmless and being able to be helpful in a

persuasive way is how like sales happens. As in, if I'm training, if I, if I'm not just a general

purpose model, but say I'm the model specific to a software vendor that's supposed to be the sales

model that's making a sales pitch to a client, you know, maybe they need kind of a training,

basically human feedback to essentially train their AI model for sales persuasion specifically

to their specific product set. So it could be like a gloss that they're putting on

purpose models that's actually more human driven than you imagine. And, and that, and that kind

of deploying it against in customers does require like, you don't train your sales force, you train

your model once. And so because you're training it once, you don't like hire up a bunch of an

ongoing operation, you outsource it to invisible to then come back and kind of like, hey, give it a

new gloss. Let's see if we can increase sales efficiency here. Let's train it on the newest

step. So it's possible that could be a continued growth area. Yeah, I bet, I bet that we don't,

we don't have, we've not done any government sales yet. But I bet that the models that the CIA

are building are very different than the models that Disney needs to build, right? So that's,

that's, that's a way of visualizing it for me. It's like, yeah, you need these models to do

different things. And I sort of, I sort of doubt the CIA's models are harmless.

Well, I guess from whose perspective. Okay, but you said something earlier that I'd like said,

like you said, you're profitable, which seems like an insane and dumb thing to do. Capital

markets will fund AI driven businesses to, you know, like scale raised $500 million. And it's

clear, like, from what I know of your revenue and growth rates, you probably should be like

gearing up for an IPO. Why, why not? And how do you think about like, why be profitable when

you could raise more money and scale more quickly? So we're approaching 100 mil run rate by the end

of this year. And we've, our growth rates are high. We exited last year at a 25 mil run rate. So

we've nearly, nearly quadrupled this year, our top line revenue, but our profits are growing faster

than our revenues. And that's because our overhead is not scaling at the same rate as our,

are the rest of our business. You can think of it in terms of, again, let's take an iPhone as a

metaphor. When Apple builds these phones, every single time they build the phone, there's the

actual materials that make the phone. And that's the cost of goods sold. So they have to manufacture

this thing and ship this thing. And that's the cogs. So you take revenue minus cogs. And for

our business, you get 50% gross margins. So we're definitely trying to automate the and grow our

gross margins, but that's, that's what it is. And then you subtract the cost of the Apple Store,

which is the cost of your sales marketing support. So the cost of the billboard ads and the cost of

the Apple Store, and then you get your contribution margins. So for us, that's 40%. And then there's

GNA and R&D, which is like your CEO, your CFO, your CMO, and then, you know, all the engineers,

the, all the designers, all the product managers. And then you get your EBITDA margins at the bottom.

So ours are, we're, we're increasing profitability faster than we're increasing revenue.

And that trend should continue for the next four or five years. And this is my favorite

interview question, especially for finance candidates, which is, has revenue scales?

What do EBITDA margins approach asymptotically? And the correct answer is they approach

contribution margins because you just don't need to hire more CEOs, more CFOs and more CTOs.

And there's not a one to one correlation between the number of engineers, Apple hires,

and the number of iPhones that Apple sells. And it's the same, same thing with us. So we are able

to meet all of our investment requirements in every department. I'll give you an example.

Next year, if we shrink the total percentage of revenue going to GNA and R&D from 26% to 20%,

we will still be able to invest at least 7 million more, at least 7 million more in

GNA and R&D next year than this year, because the number of absolute dollars flowing through

is increasing. And so, so then this leads to the question of what's sort of a company culture

and what's your, what sort of a company strategy do you want? And so in, in a, in a mindset,

which is a grow at all costs mindset, you're just trying to jam dollars into the business

to eke out anything that sounds like a good idea we should do. And that actually has been

traditionally the, the operating culture of most venture-backed startups. And the strategy of most,

the strategy of most venture-backed startups, which I call the venture game, is you raise money,

you spend it as, you know, quickly as possible to create as much growth as possible, you raise more

money, you spend it, and you keep raising bigger round after bigger round, running bigger loss

after bigger loss, and then you either IPO or you sell. And it's a five to seven year long game,

that's the VC game, series A, B, C, D, E IPO or sell. This, this is an incredibly high-risk approach

that tends to destroy most companies. Sometimes it works really well. And for certain companies,

it's kind of the only, certain companies in certain market situations, the only way to go.

We're playing a game we call the sovereignty game, which is we got to profitability, we're in our

third profitable year now. And we actually have started buying back stock. We bought back 25%

of the company from investors in 2021 and 2022. And so now 70% of the company is owned by the team.

So if you had to ask yourself, if you were a shareholder, which company would you feel,

feel better about owning the company that is just focused on growing enterprise value,

or the company that's focused on growing shareholder value? That's a trick question.

What really is the difference between shareholder value and enterprise value? I'll pause there.

Okay. Well, so let me present the counter argument, which is that imagine that there's a,

there's a certain size you need to be in order to compete in the game. As in, like you need to

establish a, like imagine from invisible's perspective, you know, there's, there are certain

processes that are shared and common across different enterprises. And so if you can sell

in the same process to multiple enterprises, then you don't have to invest in the kind of like,

or decost of like spinning up the process, you basically get to sell in at the scale price

to the next enterprise. So it looks that much more efficient to them. And so kind of having like a

breadth of customers at a certain scale is important for the unit economics of the business.

And it could be that, hey, actually to, to get a large enough strategic footprint requires you to

essentially pay for early customers at net negative margin to like hit kind of like that

strategic footprint that ultimately becomes like the money printing machine. Is that?

Yeah. That's like a Uber, Uber versus Lyft deathmatch. And those Uber versus Lyft, Lyft deathmatches,

they tend to destroy shareholder value. Everyone's competing each other in a race at the bottom.

We are not in such a game. And part of the reason is that the services industry is so large. It's

not a winner take all market. And if it becomes a winner take all market, it's probably through a

platform strategy like ours. So just, you know, you know, these numbers, but just to, you know,

to, to bring the listeners into the, into the game with us. The global economy is roughly 100

trillion dollar annual economy. About 30 trillion of that is knowledge work. And 10 trillion of that

is knowledge work in China. So there's 20 trillion outside of China. And over the course of the decade,

say by 2030, Brett predicts this is going to grow by at least 10 trillion. So by 2030, it'll be like

30 trillion of knowledge work outside of China. And then how big is Accenture say is one of the

biggest services companies, they do about 63 billion of annual revenue. So I'm not good enough

to do that percent in my head, but it's just a tiny, tiny percentage of knowledge work. So if you

take all professional services, you know, what percentage of that knowledge work is done in

house, what percentage of that knowledge work is outsourced, there may be, say, three or four

trillion dollars of outsourced work and the rest is done in house. What do we think is going to

happen by say 2050 or 2040 or 2030? I think the percentage of work done out of house is going

to increase. And there's going to be tremendous consolidation in the services industry as the

services industry becomes a technology industry. And a lot of the biases that we had in the early

years as we were building the company was the belief in the technology industry, the belief

amongst technology investors venture capitalists was that there's no such thing as a true technology

company. That is a true services company. And we just flew in the face of that dogma. We said no,

we were a true technology company and a true services company. We're, you know, trying to build

the iconic 21st century services company. And we're trying to build the BPO, business process

outsourcing company to, you know, the one BPO to rule them all. And that has been our approach.

And we're not embarrassed about being a services company. We actually think that's what clients

want. And that's where most of the value is, where Silicon Valley has just tried to invest in

only product companies that are basically tools. So it sounds to me like, given the right value

for your equity, say a venture believed in you as a technology company and so it's giving you a

technology multiple, then would you accept kind of like equity and then try to see how fast you

could press the accelerator? Or is the way in which the business is structured or the way in

which you see the market, you couldn't really efficiently deploy capital to accelerate or try

to like take larger strategic share? It's the latter. I don't think we can efficiently deploy

capital to grow faster than we're growing, which is very fast to give you a sense of what this

looks like on the ground. We've just grown beyond 100 partners on the core team. So those are the

top team members with equity positions. And we have 2,500 people in the company. At last September,

we just hit 1,000 people at the company. So we've more like 2.5x the size of the total team.

And then we had something like about, I'd be surprised we had more than 50 partners. So we've

more than doubled headcount in a year. Now, if you think about what that's like as a manager of a

business, actually, probably a lot of our listeners can imagine what that's like at their own

workplaces. If you add too many people to the team too quickly, there's a bunch of newbies and no

veterans. And so you need, you actually can't grow the headcount of the organization responsibly

beyond a certain rate. You have a bunch of people who don't know what the company is about. I've

never, you know, trying to work with other people who don't know what the company is about.

And so if you want to absorb people into the culture, there's sort of a maximum rate. And then

there's this issue of scrappiness and resourcefulness, which is that the game should not be about

trying to maximize the amount of R&D dollars we are spending. The game should be trying to,

for every given R&D dollar, maximizing the leverage we're getting out of it, squeezing as much value

out of it as possible. And that comes through innovation, creativity, resourcefulness. When

are you most creative and resourceful? You have constraints. Scarcity actually creates

productivity. Scarcity creates efficiency. So by creating a company culture where everyone's an

equity shareholder and everyone's incentives are aligned with profitability, because they know the

company's going to stay private and eventually become a dividend machine, they know that that is

their path to decrease risk and increase their returns is all around getting more value out

of every dollar. And so if you suddenly become this spending culture that is spend at all costs,

grow at all costs, you just, you get really fat, really bloated really quickly.

But what about, okay, so then you have partners in the organization that it's nice to wait for

that healthy dividend when it comes, but they need to buy a house now. How do you think about

kind of like the trapped equity value they call it they have in, you know, like a great company,

but it's also kind of majority of net worth type stuff. How do you think about that?

Correct, correct. So I'm in this position. We're going into our ninth year and we're just now at

the end of this year for the first time, creating a market for people at the company to sell their

own stock. In 2021 and 2022, we bought back a bunch of early investors. But now at the end of this

year, we're doing a combination of buybacks and secondary. So the company is setting a floor

and saying at this price, we're going to buy at least $4 million off of balance sheet and we're

willing to repurchase and retire those shares. That's the buyback. But then we're going out to the

world and saying, Hey, is anyone interested in buying shares and invisible in a secondary

transaction? And for those who are listening who aren't familiar with this terminology,

a primary offering is when the company is issuing creating new shares out of thin air,

creating new stock, and then selling it in order for the company to get money on its balance sheet.

But a secondary transaction is a shareholder in the company selling to some new outside person

who's interested in selling. And that's not the company is not the broker is not not in between

that transaction. And so we think that we're going to get like we're currently raising 15 to 20

million of secondary to facilitate liquidity for the cap table. And we're probably going to do this

every six months. And eventually, you know, we think the market will be so liquid kind of like

SpaceX stock. So SpaceX is, you know, a company worth more than $10 billion. It's not public.

And shares are pretty liquid. So they basically have liquidity in their shares is enough demand

for SpaceX shares worldwide that they have not needed to to go public to create that liquidity

for shareholders. That's our path. And and right now it's an expensive ATM. But it's an ATM that

works. And it's an ATM that you can pull money out twice a year. Eventually, you know, you'll be

able to pull that money out of the ATM at any time. And the it's not going to be as expensive as an

ATM. There will be less of a delta between what we think the value is what the world thinks the value

is with the one difference between you and SpaceX. And one reason why you would consider the public

market is not just for, you know, like you could do it on a direct listing when you're not even like

kind of necessarily raising like a big primary chump at least immediately. But it does provide

effectively a billboard for the company, you know, CNBC gets to talk about you every day for all the

good and the bad that comes with that. SpaceX has like giant billboards that they send up into the

sky, you know, on a regular basis that they get broadcast of. So like there is a there is a like

advertising to clients that you are a company that exists and a company that is like very healthy

and stable, hopefully if the stock market is treating you well. How do you think about that of

like the revenue growth that could come if you were to kind of IPO and go through the the rigmarole

of the public markets? We're not currently growth constrained. We're growing so fast. And so there's

and there's so much low hanging fruit when it comes to sales and marketing and even expansion

within our existing client base, that it's not like we don't necessarily need, you know, to

shout from the rooftops and use the public markets as a marketing channel. While you're right,

there's a long list of pros and cons to going public. I went through a pretty extensive exercise,

and it was pretty clear that the cons outweighed the pros for me. You know, a few of the, I'd say

one of the biggest cons is the short termism. It gets everyone focused on what everyone else

thinks. So if you're trying to build a culture that is driven by first principles analysis and

driven by long term thinking, a great way to ruin that culture is by going public.

So you don't want to ruin your culture?

I don't really doubt. No, no. And there's also there's also control value. There's a famous,

you know, story of, you know, Napoleon, just defeating five armies with one army that was,

you know, smaller than any one of the five of them. And how did he do it? Well, those five armies

were so busy trying to coordinate and figure out what to do. You know, he was able to divide and

conquer. And there's an incredible value in having control of a company because you can pursue

contrarian strategies. In a public company, it's very, you tend to lose control. And therefore,

everything sort of defaults to increasingly towards a consensus view, a consensus strategy.

You get a bunch of what I call, you know, the professional managerial class

that is pursuing best practices. And all of this sort of tends to approach beta,

which is, you know, your IRRs will tend to become the S&P 500s IRRs, you're going to get 8 to 12%

on average over any 20 year period. Whereas we're shooting for legendary IRRs, like multi decade,

28% plus IRRs. There's, you know, if we have three business Bibles at the company, the Bibles are

number one outsiders by William Thorndike, who invested in the company this year, put in half

a million dollars. And that is a book about capital allocation that has informed a lot of our

strategy on this. And it's a very, you know, counterintuitive. You think the best CEOs are

people like Jack Welch, you know, great managers and sort of very charismatic, but some of the best

CEOs are really focused on financial strategy. And so the book details how to do it. And a lot of

the stuff I've even shared on this podcast about the way we're thinking really comes from that.

Two, by the way, a classic example of that is acquisitions. You know, if you're focused on

growing enterprise value, you want to acquire as many companies as possible and even use stock to

do it. And you just grow and grow your enterprise value. But in the process of buying companies

with stock, you're diluting your shareholders and destroying shareholder value and 75% of acquisitions

destroy shareholder value. So that's an example of something you might learn from outsiders.

Second book that's a Bible for us is Seven Powers by Hamilton Helmer, also a friend of the

company, there's a book about defensibility. How do you build a moat? How do you build barriers to

entry? Not just benefits for your clients, but how do you make sure that the benefits that you have

are benefits that can't be easily copied? And then the third is the innovators dilemma by Clayton

Christensen, which is actually something we're facing, which is as one of our businesses is in

hyperscale mode. It increasingly, you know, becomes very difficult to divert time, energy,

and resources into diversification. And you have to be incredibly strategic about diversifying.

Otherwise, you end up hyper concentrated in this very specific market, where you're always attacking

up market, attacking up market, and then you just end up, you know, exposed. You're talking about AI

training versus which is hyper scaling. Yeah. Yeah. Yes. That makes sense. That's right.

Well, Francis, it is always a joy to speak with you. You have such

interesting perspective on how things should work. I do wonder, I do wonder,

in fact, I suspect, I suspect you are going to revisit the IPO question.

Are you sitting with self interest here? No, no, no, no. I mean, I am like, you know,

we do private and public, but it's more like, at some point, you have to compete in the big leagues.

And it really is the pro game. And the pro game requires you to be able to have a culture that's

strong enough to surf across the short termism that the market implies. And there is no

future circumstance, I can imagine, where the largest AI enabled servicing company in the world,

which is what you're on trajectory to become, let's say, is not also something that is publicly

listed. And so if you're going to deliver legendary AI or R, actually, I think doing so will in some

way be conditional upon having at least an audited set of books that you're producing for people so

they can say, hey, this is the company. We're relying. We have audited financials. Well, yes,

but I'm saying that it's not just audited financials, it's also kind of like the

knowing that the business is being kind of like unpacked and poked at by by its strongest critics,

which is really what, you know, being public with your information invites and being in the

public market invites, it's like, can you withstand kind of like that criticism? I think that's

critical for companies to trust you in a robust and profound way with kind of like their operational

integrity, which is ultimately what you're going to be asking for. And so I think that you're going

you're going to need it. You're going to need it to be as big as you possibly can. And by big,

I mean, I mean, not big, just enterprise value big, big shareholder value big. And so this is

why Brett is one of my favorite people to get advice from because he is the courage to disagree.

And and then I'm not sure I have the courage to disagree with Brett, because he's one of the

smartest people in the world. But let me let me just, while you're probably right, and we will,

we will ultimately make the right decision for the company. We see an opportunity to shrink the

shrink the number of shares by 25 to 50% over the next seven years. And if you had a business that

was going to generate 20, at least 24 million or more of profit next year, 60 million or more

the following year, over 100 million the following year after that. And you were able to do that

while paying down debt, reinvesting in the business, growing the balance sheet

and financing M&A. Would you really go public? Why? And there are other ways of creating trust,

there are other ways of marketing. That's true. I'm not saying it's the right decision now.

I'm saying you will cross that threshold. I'm saying you cannot, you cannot say you've made

the decision not to do it and then perpetually say, okay, that decision is made. I think that

there's a deep in liquid capital market becomes like an incredibly valuable, particularly like,

okay, and then from a high level, like we think that kind of there's going to be right now,

honestly, being in the public markets is not that fun, low and behold. But over the course

of this business cycle, there's going to be a huge tailwind to innovation. And so just from a

having having access in a corporate sense to that, that checkbook, you can also, you know,

strategically use that in interesting and sometimes profound ways, in a way that you can't,

if they're privately listed, or at least not with the same agility as in like you can,

you can make a decision to raise equity, but you can't like do a secondary within a month that

raises a ton of equity or use it for M&A purposes. So we will return to that debate when we have you

back because I always love having you. Thank you. It's great speaking to you, Francis. This has been

FYI, the For Your Innovation podcast. And if anybody has is curious about how to use HI to

actually make their business work, I'd suggest talking to Invisible because they know what

they're doing. And Francis is a good hang. All right. Thanks much. Cheers.

ARC believes that the information presented is accurate and was obtained from sources that

ARC believes to be reliable. However, ARC does not guarantee the accuracy or completeness of any

information. And such information may be subject to change without notice from ARC. Historical

results are not indications of future results. Certain of the statements contained in this

podcast may be statements of future expectations and other forward looking statements that are

based on ARC's current views and assumptions and involve known and unknown risks and uncertainties

that could cause actual results, performance or events to differ materially from those expressed

or implied in such statements.

Machine-generated transcript that may contain inaccuracies.