Addressing Hallucinations – LLMs You Can Trust with Scott Cohen

Welcome back to Disciplined Trouble Makers! In this episode, we’ll dive into the AI world with Scott Cohen from Jackson AI, known for groundbreaking work with the Department of Defense. Scott explores AI’s ROI, scalability, and the buzz around AI hallucination. Uncover Jackson AI’s unique strategies and the importance of human creativity in the age of advanced AI. Discover their commercial shift and funding goals, bringing AI capabilities to the forefront. Tune in to learn how Jackson AI combines innovation with confidence for the future of AI.

Quotes

“If we solely relied on AI, the spark of creativity would be lost. Instead, AI’s true power lies in empowering us to reach new heights.” – Scott Cohen

“Maintaining a human element in AI operations ensures guidance towards tasks and broader objectives.” – Scott Cohen

Featured in this Episode

Scott Cohen
LinkedIn: ​​https://www.linkedin.com/in/scottncohen
Website: http://jaxon.ai/

Chapters

00:00 – Introduction
04:04 – Focus on Specific Industries for AI
09:46 – Verification for Token Prediction Models
11:35 – Hallucinations vs Inaccuracies in Transformer Models
15:39 – Efficient AI Building with DETAIL Tool Chain
19:39 – Cooking Skills: Learning and Adapting
20:19 – Enterprise Costs and Impact Inquiry
24:27 – Human-Driven Decision-Making
28:12 – Scott Chigauri Interviews Scott Cohen on Fundraising
32:04 – Commercial vs. Government: Value Propositions
34:46 – Balancing Human Uniqueness with AI in Business
37:03 – Human-AI Communication in the Workplace
41:53 – Collaboration Issues in AI Project Failures
45:10 – Navigating Organizational Hierarchy Challenges
47:07 – Importance of Educational Outreach in Data Science Teams
52:50 – Conclusion

Transcript

Scott Shagory:

Welcome back to the Disciplined Troublemakers podcast. I’m your host, Scott Shagory. Here on DT, we talk about scaling tech companies and helping CEOs do this with a greater sense of ease and flow. Today, we’ll be talking with Scott Cohen. Scott is the founder and CEO of Jackson AI, a Boston based company started in 2017 that focuses on solving hard real world real world problems, excuse me, in domain specific areas where proprietary data is critical. Prior to Jackson, Scott founded several companies focused on machine learning, big data, and high availability systems. Finally, Scott comes from a family of entrepreneurs, a previous generation immigrating from Russia, started an ice cream cone business after the 1907 World Fair, popularized the frozen concoction, the company which ultimately became the Sweetheart Cup Co, a name that I think is fantastic. Grew from 1 cone machine into a Fortune 500 company that employed over 10,000 people internationally.

 

Scott Shagory:

Scott, thank you for carving out some time to talk with us today, and welcome to the Disciplined Troublemakers podcast.

 

Scott Cohen:

Absolutely. Thanks for having me, and thanks for digging up that background on me.

 

Scott Shagory:

You’re welcome. You know, as you know, I’m a fellow New Englander born in Boston, grew up mostly in New Hampshire. And even in the depths of January, we all love ice cream. And for me, that’s a cone dipped in dark chocolate with probably coffee ice cream on top. So, it’s always a nice way to start things.

 

Scott Cohen:

Once you bring it up, I’ll brag about my family inventing the banana boat. Did you really? The plastic blue shaped canoe, if you recall.

 

Scott Shagory:

Yep. Yep.

 

Scott Cohen:

And the flexi straw. Oh, really? Yeah. Oh, wow. Yeah. Wow. So that’s a lot of fun things credited to my family.

 

Scott Shagory:

Nice. Well, congratulations.

 

Scott Cohen:

Good legacy to try to build on.

 

Scott Shagory:

I was gonna say, those are fantastic stories too, and everyone has either had a sundae or used those straws and so forth. So an easy connection, which is fantastic. Cool. Well, let’s go ahead and shift gears a little bit to, obviously, AI and yourselves at Jackson. We met about 4 years ago. At the time when we were originally talking with each other, you guys were and I’m sure are still focused on basically AI for AI as you described it to me, where you were focused on unstructured data, specifically text data for the purpose of labeling? Obviously, you know, you’ve evolved a great deal over the last, you know, 4 years and whatnot. In terms of as a starting point for us, what is Jackson’s core or what are Jackson’s core focus areas today, you know, by industry or industries and then solutions. We’ll get into detail, but I’m guessing there might be some other things there just to bring us all up to date.

 

Scott Cohen:

Sure. Sure. And it kinda coincides with how crazy, and hyped up AI AI has become over the past few years. Yeah. When we spoke, data labeling was the big rub. That was the big bottleneck in the whole process to get custom machine learning models to a point where you can use them in production settings. Then the advent of things like GPT 4 and transformer models and these pretrained machine learning models that can do a lot out of the box. It’s more about forcing them to do a particular task.

 

Scott Cohen:

And what Jackson’s now focused on, trusting the output coming from them.

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

And we’re not using labeled data to fine tune those models. We’re instructing those models as to what the domain of discourses that we’re working within Mhmm. And the task at hand. And it’s just kinda shuffling around within that big neural network as to how to apply its knowledge to that given, specific data.

 

Scott Shagory:

Yeah. Are there specific industries that you guys have chosen to focus on either explicitly looking at the market opportunities or has it been, you know, given the rush of events, you know, I’m the VP of sales at health care company ABC, and I say, hey, Scott. You know, we have these series of issues, you know, we think AI could help in these areas. Can you help us with that? I mean, you guys have done stuff for years with DOD or the US Department of Defense, but I’m just kinda curious in terms of layout around industries, if it’s all enterprise, if it’s a combination of things, and then what you’re seeing, basically.

 

Scott Cohen:

Sure. Sure. Absolutely. We have been intentionally focused on DOD out of the gate here, done a lot of collaborative r and d with them, and realized that the use cases that they have have to have high trust in the machine learning models. I don’t wanna go too deep technically here, but if you have a non deterministic system, meaning you can give a different answer to the same question. Yep. You can’t use it in operational settings, especially, you know, working with things like weapon systems. But you still wanna leverage that power of AI.

 

Scott Cohen:

And this is not just about the use cases we’ve been working on, but across DOD. They wanna harness the power of AI, but they have to trust the output.

 

Scott Shagory:

Of course.

 

Scott Cohen:

So that’s what got us into the current, you know, line of of the products, which is around validating and and looking at the output, verifying that it meets the required spec, and serving as a filter in between the big AI models, large language models, and the like, and the user. So it’s adding a layer of trust, I guess, I could position it as.

 

Scott Shagory:

Gotcha. Gotcha. Okay. And you have talked about that more recently with, again, some of the work I was preparing for today in terms of and we’ll mention it in a little bit here. Sometimes guardrails, again, that level of trust and so forth to add not just accuracy, but genuine confidence in what is being provided as a set of insights, maybe even we could describe it eventually as wisdom, etcetera, like that. So and that seems fair.

 

Scott Cohen:

And to finish, answering your question, the industries that correlate to that, you know, the commercial sector side of things, are highly regulated industries. Those that have big downsides and or upsides to leveraging this AI. So if they’re using it for things like SEC filings and financial services, you don’t wanna get that wrong. You’re using it for clinical decision support and health care or a drug discovery for life sciences. Those are things that you really wanna have rigor or what we refer to as formal methods in the process so you can have, you know, mathematical proof that what you are spitting out is sound.

 

Scott Shagory:

Yeah. Yeah. I mean, I think we all can appreciate, you know, exactly. You know, there’s a lot of excitement many people are using, whether it’s Chat, GPT or others. And it is, it’s just cool to have it in millions of people’s proverbial hands, so to speak. But like you say, when the rubber meets the road where, you know, as someone recently said to me, you know, how would you feel about going to the dentist and the dentist telling you that, you know, he or she has 80% accuracy, and then saying, Well, what do you mean by that, ma’am or sir? And you say, Well, I’m only gonna hurt 20% of the time. 80% of the time, no, man. It’s gonna be awesome.

 

Scott Shagory:

You’ll be just fine. Right. Surely, you find that that outcome or set of outcomes to be acceptable. And, of course, everyone laughs because that’s absolutely absurd. It’s the same thing where even if it’s 2%, do you wanna get hurt 2% or be in discomfort 2% of the time? And so it’s not hard to see scenarios that are really very exciting, but the risks are absolutely there and crossing t’s and dotting i’s is essential. Yeah. Yeah. In terms of, I guess, as a follow on a little bit because I’m sure, you know, as is true with many technology companies, you know, there’s a lot out there where people are not necessarily clear about what you do and how you do it.

 

Scott Shagory:

So, you know, what do you think about conferences and all the speaking that you do? What is most misunderstood about Jackson and what you guys do?

 

Scott Cohen:

Well, people are very much aware of the so-called hallucination problem. And it’s kind of missed by the name itself, it’s kind of a mystical concept of these AIs going off the rails. A lot of people are thinking that if you marry in representative data, it’s called the retrieval augmented generation technique.

 

Scott Shagory:

Is that right? Is that what you’re saying?

 

Scott Cohen:

The rag method is enough. That all you need to do is get your own curated data, marry it in alongside the the language models or whatever AI models you’re working with

 

Scott Shagory:

Yeah.

 

Scott Cohen:

And then you’re done. And it’s true that the the rag techniques add a lot of accuracy to the equation, but there’s still that room for error because the the way that the the architecture of the current AI models like ChatGPT

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

Is sequence based. They’re making a prediction as to what the next token is in the sequence. K. And if it has one little, you know, probability mismatch and gets it wrong, that starts to compound the issue. Mhmm. And so, yes, you can use rag techniques to try to improve that token prediction, but you can never have full trust in, at least, the current architectures of these models unless you do what the techniques that Jackson’s been doing and what is kinda forced upon us by DOD is once you get the output from these models, run it through the ringer, run it through a test gauntlet, and prove that the accuracy aligns to the data. You’re still looking at the knowledge base of data and any rules you wanna impart upon that process. Any constraints, assertions, assumptions, you can play with that in what’s called symbolic reasoning.

 

Scott Shagory:

Mhmm. Yeah. Yeah. Yeah. I know. And that’s a core part of what desale is, which I wish we’ll get to and to talk about here in a little bit. Because I’ve had, you don’t know, no surprise and all of us have been exposed to that word hallucination. We all have a historic representation of what we think that is, versus again, how it’s used, you know, locally within an AI context and different meanings and so forth.

 

Scott Shagory:

You’ve partially answered it already, but what does hallucination mean to Jackson and yourself? I know there’s the 4 kinds of bucket areas of how it’s described and output and things like that. But I’m, you know, for me, there’s always just a lot of context and meaning behind that, I think, particularly in what we’ve talked about earlier in health care, finance, etcetera. What does it mean to you guys, you know, contextually there,

 

Scott Cohen:

please? So I think hallucinations is a fine buzzword of the day, but it’s really about inaccuracies. Does the model get it right or wrong? And before these large pretrained transformer models as they’re called, whether it’s for language or computer vision came along, we would train models and use a carve out test set to see how well the trained model did on that, again, hold out test set that it hadn’t seen before. How accurately would it make the prediction for a given example. Yep. And it was never 100%. It was always, you know, eighties or sometimes worse. Nineties is where you were aiming for, but you still have these inaccuracies. It got the prediction wrong.

 

Scott Cohen:

Whether it’s a factual inaccuracy, it’s semantically missed the context of the question or the prompt. It sounds very real, but it could be completely fabricated facts that it’s spitting out. And sometimes it’s also gibberish. You know, syntactically, it gets it wrong.

 

Scott Shagory:

Yeah. So I’ve I think we’ve all seen some some real fun, in in air quotes, fun, meme like examples that, again might be from a grammatical standpoint, absolutely spot on, but just again, be either gibberish or so just, you know, contort it that it’s just, it’s fun, but until it’s serious, and then it’s not quite so much fun after that. Yeah.

 

Scott Cohen:

Yeah. And there are lots of classic examples. Like, I don’t know if you saw the car dealership that used a chatbot that

 

Scott Shagory:

I did. That was for a pickup truck, I think, or something. And it was ultimately for a dollar, and they ended up actually selling what it was at? A 50 k truck for a buck or something

 

Scott Cohen:

like that. I’d classify that as a hallucination also. It got it

 

Scott Shagory:

wrong. Mhmm. Yeah. Absolutely. Absolutely. It’s just, those are fun ones, but they’re not so much fun if you’re the team that implemented that for the dealership. You probably Or

 

Scott Cohen:

the case files that were fabricated. Michael Cohen actually got caught doing that. Yeah. Other lawyers have been caught doing that. It sounds right. It looks like great case law, but fictitious.

 

Scott Shagory:

Yeah. Absolutely. Fictitious requires people who you know, have very clear domain expertise and yet can obviously quickly validate that for now, you know, until it moves more quickly. Yeah. I mean, when I think of when I was doing, again, background research on, you know, desal and and what you guys have been doing recently, you know, is, you know, Desale fundamentally a design system? Because you’ve talked about that for building AI models. You’ve also talked about it in terms of a set of guardrails and fact checking. Is it a bit of both? Is it like architecture? It’d just be nice to talk about it because I think there’s just a lot of meat to unpack there as to what you guys have really done, and the symbolic reasoning that you mentioned just a few minutes ago is, obviously, a very important piece of that.

 

Scott Cohen:

Very astute question. Honestly, that’s what you picked up on is that we have to convert natural language into a domain specific language. A domain specific language for whatever use case we’re focused on. For DOD, it’s weapon systems. For us, for our commercial product, the one that you can get through our SaaS edition, it’s a domain specific language for building custom applications. Software, AI, AI plus software, what have you. And it’s really about the proper architecture of those solutions and weighing the pros and cons of various design decisions. Various constraints there might be around scalability, might be around cost, might be around security.

 

Scott Cohen:

There are lots of different constraints that you need to consider when you’re building custom AI. And the DSAIL tool chain is designed to take that natural language, turn it into DETAIL, run it in through a compiler, and then we have about a dozen steps from there that are running through things like solvers and formally proving the output. Turns out, what we realized, this was just about 6 months ago, is that we could keep adjusting that DSL. And depending on the grammar, the vernacular of the particular domain, once we’ve codified that into a DSL domain specific language, we can use the rest of this tool chain. The compiler has to reflect the DSL, but those dozen steps I was alluding to, the solvers and whatnot, are all there and set up so that we can use the front end LLM to convert that natural language into our DSL and then the rest of DSL works. So we’re kinda back and forth with the LLMs or the large language models where they’re helping us create the DSL. We’re helping convert the natural language into that DSL, and then we are fact checking the output of the DSL given a new prompt or a new set of data that’s being processed.

 

Scott Shagory:

Interesting.

 

Scott Cohen:

You saw that.

 

Scott Shagory:

Yeah. So with the domain specific language, which, you know, you and others have described as thinking of it as a, a bit of a dictionary for a specific area. So law, I’ll just generalize law or health care or something like that. So you’re talking about not just a single input where we humans have validated or approved of a set of, I don’t wanna call it a corpus, but a set of information and of data just to keep the language simple.

 

Scott Cohen:

Yeah. Addiction. And then

 

Scott Shagory:

not just inject it once or have it be incorporated once, but have it be continually, I assume via compute as well as human added interaction and refinement if that’s a fair description to then go ahead and just keep, frankly, keep innovating, keep iterating till we end up finding a place where, accuracy is either acceptable or exceeds exceeds acceptability. Excuse me. Yep. Is that fair, or where have I gone wrong? Yeah.

 

Scott Cohen:

That’s fair. And we keep a human in the loop to ensure that accuracy as well. So we’re drastically reducing the amount of, say, analyst time, but we’re keeping an analyst in the loop to review the output of the system to make sure that for again, for these use cases that we’re dealing with, that there aren’t any, you know, anything that went wrong in the whole process.

 

Scott Shagory:

Yeah. Interesting.

 

Scott Cohen:

I have a quick anecdote.

 

Scott Shagory:

Sure.

 

Scott Cohen:

So describe a domain specific language. I don’t know if you know this about me, but after MIT, I still had the education bug, and I went to culinary school.

 

Scott Shagory:

Did you really? And then

 

Scott Cohen:

I got to play in a real kitchen. And in a real kitchen, the, say, sous chef can call out to a line cook and say, go saute me some onions. Mhmm. And that’s very clear to them. There’s no ambiguity there. Mhmm. You don’t have to go through all the extra crust that might leave room for error. Like, go get an onion, dice it.

 

Scott Cohen:

And what does dice really mean? You know, you might Correct.

 

Scott Shagory:

Yeah. Versus chiran or whatever. Yeah.

 

Scott Cohen:

Put it into a pan, heat it up. Don’t burn it, but make it translucent. All that extra cruft is where you can have room for error. And that’s basically what a domain specific language is doing is making it very clear for machines to understand what the task at hand is, what the communication is trying to say.

 

Scott Shagory:

Yeah. Interesting. Yeah. I, years ago, actually took some semi pro culinary classes with a cooking school up in Bethesda. Oh, sorry. Up at, yeah, Bethesda and and, Gaithersburg, Maryland, Cold Lake Academy do cuisine. It was very much, front classic French training with

 

Scott Cohen:

Sounds like that.

 

Scott Shagory:

Not recipe based, but skill set based. So you like, you know, saute versus sweat versus others and then just think dynamically where the the example is great even if you’ve only cooked a few things at home and it’s just eggs where we assume so much, and yet there is in a in a particular sequence, particularly in baking as well where, you know, cooking is a bit more free form, whereas essentially baking is is chemistry in action. So, it actually makes a very big difference as to what the outcome is as anyone knows if they flopped a cake, which I have certainly done to many patients in the past. The key is to iterate as always. Key is to, you know, not make the same mistake a second time.

 

Scott Cohen:

That’s true. Very true.

 

Scott Shagory:

Yeah. In terms of, I guess, along these along the same lines here, we’ve got what I’m gonna guess here, and you can correct me where I’m I’m not right, is that, you know, the I would imagine the queries that you’re doing in conjunction with a DSL are obviously either quite large or potentially very, very large. And that is not cheap. That is a very expensive enterprise. And so inevitably, even with the DOD’s budget, amongst other large enterprise players, you know, cost is a huge deal. And so I’m wondering, you know, are you also focused or are there any desires around cost impact or control or inefficiency? Or is it really about it at its core being able to do something with a level of accuracy that’s acceptable that you one was never able to do before with some cost consideration in there in terms of what the what the set of drivers

 

Scott Cohen:

are? Sure. There’s definitely some of that. With DOD, we’re working with open source models that are housed on their infrastructure, air gapped, and they’re not paying depending on how big the prompt is. It’s just for the back end computer. So it’s not as costly. For industry, we run up against this a lot where everyone wants to use the latest and greatest.

 

Scott Shagory:

Yeah.

 

Scott Cohen:

But you don’t always need to. You can get away with, even for industry, open source models that you’re housing yourselves Yeah. Or use cases that it’s good enough. And it depends on the type of data that you’re processing, but sometimes it’s actually overkill to use the big models, especially for things like tabular data.

 

Scott Shagory:

And is that just because of the time, or is it a case where it can be, if you will, computed just much more quickly with simpler models or both or what?

 

Scott Cohen:

Yeah. Both. It’s lighter weight.

 

Scott Shagory:

Interesting. Yeah. That’s And

 

Scott Cohen:

That’s part of the architecture too is, in my opinion, the future is multiple models working with each other, multiple agents, which are really a way of saying you have a model and you’ve instruction tuned it to work on a specific task.

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

I envision multiple agents working together. And so each agent can be powered by a different foundation model depending on their role in the whole process.

 

Scott Shagory:

I mean, it makes perfect sense that you would specialize in a variety of different ways, either compute, yeah, depending on what the needs are to get to that end result, or more quickly, more efficiently, accuracy, you know, all of the above. Yeah. Interesting. Do you think, you know, for you guys is you know, given what we’ve covered so far here, is Jackson’s fundamental differentiators, do you guys think of it as being at its core around technical competencies or are you guys fundamentally from a philosophical point of view and from an expertise point of view in terms of your context and so forth as you sell into the marketplace is or you guys about really building trust and confidence?

 

Scott Cohen:

I mean, I’ve been going through some internal exercises to arrive at what you just said. I I think it’s really about building trust and confidence in the models. Because even as they continue to progress to be that much more powerful, I mean, GPT 5, I hear is coming soon,

 

Scott Shagory:

and

 

Scott Cohen:

that’s gonna be even better than GPT 4. You still wanna have these mechanisms by which you can impart policies and rules, constraints, assertions, all that other stuff. Yeah. And you can have some type of verification system. Even if it’s just a flag, those anomalies that happen once in a while. You wanna have something that balances that. So to answer your question, I think it’s all about trust.

 

Scott Shagory:

Yeah. I mean, at the end of the day, you know, humans are still making decisions. All of us are analog, and that’s not going to change. And we are at the end of the day, and we’re both old enough to have gone through the dotcom boom, you know, global financial crisis, lots of other things. And in the end, all of those crises were precipitated by a lack of confidence, obviously rooted in some measure of data or assessment or, if you will, a collective belief that something was very much out of whack. But in the end, you know, humans made those decisions and and, you know, in a rational and fair manner, manners went ahead and made, you know, shifted focus, shifted direction, and we had a series of outcomes as a result of that. And so, yeah, I I can’t really, you know, when I was thinking about you guys, you know, as well, I I I kind of came to the same conclusion, just generally speaking without having done all the all the hard work that you have done, that it is very much about that because it’ll be humans that’ll make decision in DOD enterprise, whatever that happens to be. That just makes sense.

 

Scott Shagory:

Interesting.

 

Scott Cohen:

Yeah. I was on a panel recently. This was at an army university, and they’re asking, are they gonna AI is gonna replace many jobs?

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

And, you know, some stats were thrown out, like, 30% of jobs are just gonna go away.

 

Scott Shagory:

Yeah.

 

Scott Cohen:

My opinion is that AI is giving those with those jobs superpowers so they can do their job x times faster. Mhmm. As even for a developer, you know, AI is gonna be able to write all the code. You still wanna have the expert in place to verify that that code code is right. You still wanna have someone that’s been educated in that subject matter, whatever the subject matter is, to be the one looking over the shoulders of the AI, directing it to do what it wants it or what they want it to do. So I don’t think it’s coming for everyone’s job.

 

Scott Shagory:

Mhmm. I

 

Scott Cohen:

I think it’s going to give those that embrace it, again, superpowers.

 

Scott Shagory:

Yeah. I can’t, I can’t argue with that as well. There’s always an adverse impact, but the reality of the matter is that I think it will either turbocharge, supercharge, you use the the term of choice that, yeah, it’ll just give those who have either genuine expertise or those who wish to acquire it, the opportunity to really just excel in ways that would be have been more much more challenging or just taken a lot longer to achieve than would be otherwise the otherwise the case. Yeah. Yeah.

 

Scott Cohen:

I’ll throw in one caveat. Of course, there’s some exceptions. Like Yep. I think call centers are gonna become mostly automated in the future.

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

Other tasks, like real analyst work, are not going away. It’s just gonna be empowered by AI.

 

Scott Shagory:

Yeah. And that sounds fair, we can certainly agree on agreeing on that. Cool. Well, let’s go ahead and just take a quick commercial break, and we will be right back. Purple Finch Group helps technology CEOs innovating at the edge of chaos create operational fitness across their technology stack and team of talent. Achieving significant and measurable scale, selling your story and finding time to think and plan can feel like a bridge too far. So if you’re a CEO whose company has found product market fit but is struggling or you’re a CEO whose company is looking to accelerate existing momentum, we can help with proven blueprints for optimal scale. Shifting from incremental addition to exponential multiplication can be absolutely achieved with less effort and stress.

 

Scott Shagory:

To get in touch, I can be reached at Scott with 2 t’s scott@purplefinch.com

 Now back to the interview. Welcome back to the Discipline Troublemakers podcast. I’m your host, Scott Chigauri, and we’re talking with Scott Cohen of Jackson AI. So to pivot a little bit here, Jackson and your team, you have chosen, at least not yet, to raise money during this AI boom. And I’m wondering, you know, why you’ve chosen to effectively bootstrap. I know you took on a little bit of, you know, did something with a Swiss team a number of years ago, but I’m just kind of curious how you think about raising outside capital, the at least to date current choice, you know, not to take on a huge chunk of money. There are obviously trade offs.

 

Scott Shagory:

But with what you’re comfortable sharing, you know, what has that process been for you guys?

 

Scott Cohen:

Sure. Absolutely. For the record, anyone watching this, I did not take Swiss money. We’re 100% US backed. Very important for the DOD work that we’re doing. The agreement is completely owned by US citizens. The Swiss actually had a lot of fun. I did an accelerator there, and I got to play with chocolate companies, which was fun.

 

Scott Shagory:

Oh, cool.

 

Scott Cohen:

And some interesting use cases outside of the core DOD focus. So it was a nice breather. We have raised funding. It didn’t make it to PitchBook or Crunchbase. We’ve raised, just shy of 3,000,000 so far. Okay. And we are currently raising another round because as of yesterday, we got an indication of an award that requires us to bring matching funding in. We were out on the hunt to find private investors to match the DOD, R and D funding that we’re about to get.

 

Scott Shagory:

K.

 

Scott Cohen:

So I’m back on the fundraising trail. In fact, all this morning, I was focused on my list, dusting it off, finding those that appreciate what we’ve done with DOD

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

Which for the VC world is not everyone.

 

Scott Shagory:

Mhmm. Yep.

 

Scott Cohen:

And, yeah, we did bootstrap, which was actually very refreshing. And I did it before with Big Rio, my previous company.

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

It’s just with this space, we need to have that growth capital. We need to get out there on the sales and marketing front, and the funding will help precipitate that.

 

Scott Shagory:

Yep. And is that in terms of pragmatically again with what you can, what you can share is a good chunk of the capital about, GTM or go to market strategy and what you guys need to do in the sales and marketing and sales space? Or is it also on the upside and being able to, you know, you know, build out the technical team, etcetera, like that.

 

Scott Cohen:

Yeah. No. It’s 100% on the go to market side. So it’s all about sales and marketing with this new funding round. I will add that we did go for multiple years just working off the DOD r and d funding. Yeah. And now we’re at a pivot point where we have a product. We just went live with IBM a few months ago.

 

Scott Cohen:

We’re trying to go after the commercial space in a bigger way. So it’s really good timing also to bring in this cash infusion to help, again, with sales and marketing.

 

Scott Shagory:

Yeah. No. And that makes sense. And I have spoken over the years with quite a few, you know, venture capital teams and stuff. It’s always interesting when they look at a firm that has been primarily either fed based or DOD based and the degree or lack thereof that they feel that the market has indeed validated.

 

Scott Cohen:

Sure.

 

Scott Shagory:

It’s been my view for a while, and not that it’s fair because it’s certainly not, where if you start in the commercial space, say enterprise, and you, in air quotes, validate the model according to, you know, VC team A or B, and then the federal stuff and or DOD black box work is additive. That’s a very different conversation than starting the other way around and then trying to utilize it the other way. It’s just tough. But some things, they just don’t lend themselves. They are naturally, you know, a very natural place to start is with the DOD or the intel community or some area of government. We could argue about what that is and then kind of move from there. So the story is, I think, really critical in terms of how you really frame your value proposition and then just move forward, which, yeah, I can definitely sympathize with that having talked with quite a few, say, cyber teams here in the Washington DC area. Yeah.

 

Scott Shagory:

It definitely makes a difference.

 

Scott Cohen:

I will add that the fact we did bootstrap for the past few years or or I I shouldn’t say bootstrap. We’ve survived on our own without taking on additional capital. Yep. As it is set up, we set up properly where we have a stable base of revenue now. And this new round of funding is purely for growth where most companies are still and it’s a very big challenge, navigating product market fit to get that first revenue, and we’re showing up with real revenue in hand.

 

Scott Shagory:

Yeah. I think, you know, what I think is great for, you know, companies like yourselves and you’ve, you know, that you’re this is your, what, 3rd or 4th 4th company. So you’ve been to multiple rodeos, you know which game that you’re playing, you’ve been primarily bootstrapped and could continue to do that. It’s a constraint, but again, it just, I think, gives you flexibility to ultimately find the right financial partners for you, for your growth and for the companies that you feel that you’re gonna serve while protecting your DOD work. And again, everybody’s in alignment with the work that you do and what it brings you guys all the most meaning. I just think that very quickly can get lost when a large money gets thrown around and control shifts very quickly as do priorities, and you guys have just side step that in particularly in this environment where AI is amazing in terms of funding, but everybody else are often are, you know, if you’re not getting it could be cash flow positive, you’re just, you know, there are a lot of firms that have just simply shut down, and, there’s there’s just no going back. Yeah. No.

 

Scott Shagory:

It’s a tough spot. Cool. So again, shifting gears a little bit more here because it’s a question that has been banging around in my head, and I really I’d I’d like to get your input on it because I tend to go a little bit back and forth on it, which is that, you know, as AI proliferates very, very quickly, even if it isn’t, you know, with with where we are, You know, what do you think needs to be retained that’s uniquely human in a company, yours or, you know, a client’s in an increasingly AI world. So the company always maintains its individual position in the marketplace. I just think that there is this line sometimes where we’re heavily focused on efficiency, speed, accuracy, etcetera, which AI, we’ve talked about, does a fantastic job of that. But I think that there are also things within a company that are very uniquely human and that should remain so. You know, whether it is the company’s philosophy, some of its expertise, we’ve talked about expertise and the importance of that. I’m just curious, you know, what you know, how either you think about that as it relates to Jackson or just thinking about it more generally as we all have a conversation about, you know, what that I’ll call it that AI line is, where we think about or how we think about that.

 

Scott Cohen:

Yeah. It’s an interesting question. Sam Altman went on record as saying he thinks the next $1,000,000,000 business

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

With just one human and the rest AI, I think that’s certainly possible. Mhmm. But I agree with you too that there needs to remain a human element. And picking off of what I said earlier that the humans need to oversee the AI and direct them to the task at hand and have that human oversight that’s not just making sure they’re accurate, but going toward the goal. What’s the vision of the company, the mission of the company, and how do you stay on track with that? Certainly, you can instruct an AI to do so, but the human element, I think, will remain important. And I’m just reflecting. I I hadn’t thought about this question. Just today, I’m a Slack junkie.

 

Scott Cohen:

So I’m always communicating to employees in Slack, and sometimes it can come off wrong. Like, my intention was not cold, but it comes off cold. And if it’s a whole bunch of AI and some humans, you’re gonna have that coldness, and you’re gonna miss out on the culture of the company, the, you know, with the chutzpah, I like to say, for Mhmm. Your company really drives it forward. So I think it’s gonna be a very delicate, not difficult, delicate balance, to make sure you keep that human element in. I think it’s also so new for us that we don’t really know how to treat those waters to really definitively say at this point in time. But, yeah, I definitely think it’s gonna be a human machine collaboration, and I think the human element needs to remain not just for the direction of the company, but for the human to human interactions that will persist.

 

Scott Shagory:

Yeah. Yeah. No. That seems fair. I am with some of the clients that I work with. One of the visuals I use a lot of visual models to, you know, kind of tease out stuff and one of the ones that I use is called an origins of Genius Model where we talk about the company’s original philosophy. You know, wins and losses, the scar tissue, combined with their expertise and one of the things we spend quite a bit of time on is context because oftentimes when, you know, the founder like yourselves or others who’ve been there since the beginning, either shift roles or there’s something else going on. The company basically scales. Context is often lost in the marketing and sales side of things.

 

Scott Shagory:

And so when that happens, we all have done this ourselves when we’re on the when we’re a buyer of something, basically anything when a company hasn’t shared the context and we don’t feel connected to that company’s philosophy. Regardless of what it is, we always substitute our own. And oftentimes, that’s price, that’s speed, that’s okay, let me compare specs. It just all gets lost very quickly. And I just think that holding on to that and that’s, you know, essentially what some of what you shared is essential without it. There won’t be a need or a real driver behind it. That is, I think, sustainable in the long term, whether it’s about sustainability around profit or serving customers or, you know, all of the above. It’s just, I think, very early days, and we don’t always have the answers for that.

 

Scott Shagory:

Yeah.

 

Scott Cohen:

I was watching a clip recently from Steve Jobs talking about how great Nike has done it insofar as setting a company vision, mission, their messaging, their marketing. Mhmm. If we’re done by AI, they wouldn’t have had that creativity. And, you know, what they do is they empower the athlete. They’re not talking about their shoes. They’re not talking about, you know, a better soul. They’re talking about how great this athlete is, and they they, you know, do things push things like just do it. It’s a different stance than a company would take if you didn’t have that human element.

 

Scott Shagory:

Yeah. No. And Nike is a great example for that, you know, as a way of not just positioning themselves from a marketing standpoint with the just do it, but it’s a reflection of the philosophy behind it, the fundamental core of who they are, that it is about athletes for athletes. And we, you know, if we both had Nike shoes, we get to share in that. It’s about building that community. And like you said, would you be able to create that as a strategy just by, you know, putting something into chat, you know, chat or, GTP 4 or whatever? And the answer is no. You wouldn’t, you wouldn’t get that outcome at all. Yeah.

 

Scott Shagory:

So no. Very very, always very informative.

 

Scott Cohen:

Just riffing on that. I wonder if big leaps of innovation would happen without humans in the loop. Mhmm. For example, I was also watching something recently where someone was talking about when Henry Ford was out doing focus groups Yeah. And would ask what people are looking for, they would say faster horses.

 

Scott Shagory:

Exactly.

 

Scott Cohen:

They wouldn’t envision a combustion engine and wheels and all that, you know, what we know as cars today. Yeah. I wonder if that’s an important aspect of this as well is to keep having humans dream and be creative as we do.

 

Scott Shagory:

Yeah. No. I think so too. I mean, Steve Jobs with the, you know, the Ipod amongst other things would a focus group have actually created that and, you know, we could argue about those sorts of things too. Yeah. It’s kind of fascinating where most original or originating ideas are often from either one person or from a very small passionate group that just simply will not let it go and will continue to drive. Yeah. No.

 

Scott Shagory:

Interesting. Speaking of people and teams and so forth, there are and and you guys have talked about this where, you know, we’ve all seen AI projects and other tech projects not succeed. They’ve, you know, failed. And for AI, one of the big reasons that you guys have talked about has been collaboration failures between the biz team or, you know, or biz side of the house. And you guys focus a lot on data science teams for understandable reasons. For the CEOs who are listening here and are obviously gonna be in some cases, you know, layers above the DS team, and also above some of the analysts and biz team side of the house as well. What do you think are some basic steps for their, you know, biz team and for everybody to be dialed in? Or if it’s easier for the CEOs listening, what questions not necessarily at the analytical level or where that team is, not necessarily at the analytical level or where that team is to really kind of be able to bridge that communication gap where stuff, you know, comes apart with any technology team. I’ve certainly been part of that in one way or another in my past lives.

 

Scott Shagory:

Just a little bit of advice here for CEOs, I think, would be fantastic if you could share that, please.

 

Scott Cohen:

Sure. Let me just share or explain the problem a little bit more. What I see is that a business unit has an idea for AI. You know, they’ve been watching movies and seeing the news, and they dream up something.

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

I’m not saying it’s, bad idea, but sometimes they’re not aligned perfectly. So engineering has to actually build it.

 

Scott Shagory:

Indeed.

 

Scott Cohen:

And can translate between the business unit request and where the rubber hits the road around the developers, the data scientists working together to build the thing. Yeah. I have seen people communicate via email, Slack, Skype, and things like that. I can’t believe I just said Skype. Is that still a thing?

 

Scott Shagory:

I haven’t used it in a while, but I do know of some folks who actually use Skype.

 

Scott Cohen:

So, for Teams,

 

Scott Shagory:

We can say we can substitute Teams quickly. Yes.

 

Scott Cohen:

And maybe they’ll have a meeting and they’ll articulate, hey. I’m looking for AI to do this for me. Mhmm. And then oftentimes, what I see is the data scientists then interpret what that means and go off and build it. And a few months later, they come back and they miss the mark. Or they come back and something’s changed. New requirements came into the picture. They designed it toward a spec that was actually the wrong line of thinking.

 

Scott Cohen:

They weren’t the domain experts, and they inferred things based on what they knew about the company and what the use case exudes. But the devil is in the details, truly.

 

Scott Shagory:

Oh,

 

Scott Cohen:

yeah. And so collaboration is my thing. I advocate doing it not, you know, once the model has been returned to you, but, ideally, every week, have a check-in to make sure it’s going down the right path, and don’t just trust that the engineers had to translate business unit speak into real world applications.

 

Scott Shagory:

Yeah. No. I think that’s excellent advice. I’ve been on the technical side early in my career, been on the strategy side and then just the opportunities for just genuine honest miscommunication based on where people are in the food chain or the org chart and them all trying to do very much the right thing where, you know, a CEO is obviously focused on the vision, he’s focused Rashid on the financials and a couple of other areas, but then delving 4 layers deep if it’s a bigger team. You know, you as the data scientist or me as the biz analyst, again, I’m obviously gonna focus on what it is that I know best and think and how I’m connecting my work or not to the company’s vision. And I just think, you know, the opportunities for for standard, tech related activities, networking, dev, all the sort, let alone then talk about, you know, the work that you guys do, you know, on the AI side or all of the work that data scientists do and just the the grind of just cleaning data, which is, you know, productivity.

 

Scott Cohen:

Yeah. I I wanna add 2 more things. 1 is a lot of business unit executives are a little bit hand wavy over the data. Just go clean it up and use it. It’s right there. Just go grab it. It’s not that easy.

 

Scott Shagory:

Yep.

 

Scott Cohen:

The other issue I’ve seen is that they think that ChatGPT is the silver bullet that answers all questions and can do all tasks, and that’s not the case. Yeah. It needs to be architected properly alongside other rules and data repositories and perhaps other models, other preprocessing steps. It’s not one and done. So it is it is rather involved, and I I think that a lot of not the collaboration mismatch, but the understanding of the the scope of work

 

Scott Shagory:

Mhmm.

 

Scott Cohen:

As it relates to the return on investment is another mismatch that I’ve seen.

 

Scott Shagory:

Yeah. Do you find you need to, even within the data science and maybe biz teams, depending on who your economic buyer is, maybe that’s the CTO or someone else, do you find you need to do a lot of edge still a lot of education, both in terms of, again, what you do and offer, how to go about organizing that within the group or at least say, hey, here are these other three things that we see teams over and over again struggling with. Just make sure you address these. Do you find you’re doing a lot of education in addition to selling, or is it a case where people do understand the level of, if you will, complexity with who you deal with on a, if you will, a daily basis.

 

Scott Cohen:

Sure. I’d say it’s a split. Maybe not a perfect even split split, but a split between those that I can just set it and forget it. They get what Desale does. They get what Jackson’s all about, and I don’t need to go through that extra effort. But there are a lot of companies that are still kicking tires on this whole AI thing. Mhmm.

 

Scott Shagory:

And

 

Scott Cohen:

they don’t have the AI team set up. They don’t have the PhDs on staff. I mean, some companies I’ve worked with have over 200 data scientists. Quarter of them are PhDs. They don’t need that extra hand holding. Yeah. But there are still many large companies that haven’t gotten to that level of sophistication yet. They’re still figuring out what it means to them, finding the use cases that make sense to go after, aligning the business units with the data science teams.

 

Scott Cohen:

That’s still something they’re figuring out. And then I do have to jump in and help with that piece.

 

Scott Shagory:

Yeah. I could see where you could, not only on the front end, but throughout the entire engagement get pulled into conversations just like what you’ve described and just having that be a very well meaning, but time consuming activity, whether it’s yourself or your CTO or other teammates. And then as you scale and bring people up in sales and marketing, just have them get pulled into those conversations, which again can be helpful and embed you, but at the same time, there’s that double edged sword about scale and being able to do that efficiently. Yeah.

 

Scott Cohen:

Sure. I will say I don’t mind it. I actually like that piece. I feel like one of my superpowers is to be able to translate business speak into tech speak. Yeah. But, yeah, it certainly doesn’t scale, and it does take time.

 

Scott Shagory:

Yeah. No. That’s understandable. As we kind of get close to the top of the hour, and I wanna be respectful, how would you know what would be best for yourself? Do you prefer email or socials? What works best for yourself?

 

Scott Cohen:

Oh, sure. They can come to jackson.ai. Look around. Hit contact us if you wanna reach out to me and find me directly. Go to my dot jackson.ai if you wanna try our product out for yourselves without talking to humans. Although the human element still is important. Yep. Yep.

 

Scott Cohen:

And via email at jackson@jackson.ai.

 

Scott Shagory:

Okay. Nope. That sounds fair. And we will, everyone, include that in the show notes below as always so that you can, go ahead and make that easy. And then the last question I always like to ask is, you know, do you and and Jackson have maybe 2 challenges that you would like either some help solving in the next 12 months with my audience being a bunch of, predominantly tech CEOs. They might be able to help and then obviously reach out to you and say, hey, Scott, you mentioned this 6 months ago. I caught the video, you know, caught the video or audio now. I might have some help here if you’ve got 1 or 2.

 

Scott Cohen:

Yeah. Sure. My current quest is to find those use cases that actually require high confidence in the output. You’re talking about a marketing use case. You get it wrong. It’s not earth shattering. You talk about a call center use case, you know, call transcript analytics, you get it a little bit wrong, not earth shattering. Mhmm.

 

Scott Cohen:

So what are the use cases that are out there? And I mentioned some that I thought might be in the regulated industries. But what are the use cases that absolutely require that rigor, that extra step of confidence that the models are spitting out accurate answers. And I haven’t found them all yet. I I was researching some in life sciences over the weekend. I’m sure there are some in other industries that I haven’t thought through yet. So I would love it, it’s really just a fact finding mission for me at this point. Find those use cases that we fit with, that makes sense for us to align our technology too.

 

Scott Shagory:

Gotcha. Okay. Well, that’s a call out to all the CEOs out there who may listen to this podcast. If you’ve got something, reach out to Scott and his team and have a conversation. Well, Scott, I really appreciate you carving out some time today to chat with me over the last 60 minutes or so. Again, I’ll include all the show notes or details at the bottom so you can get in touch with Scott. And we will be back with another episode of the Disciplined Troublemakers! Thank you for tuning in to the Discipline Troublemakers podcast, and we hope you found the episode valuable. If you did, please subscribe to the Discipline Troublemakers podcast on your favorite platform and leave a review.

 

Scott Shagory:

Putting this podcast together is a group effort, so I’d like to provide a shout out to Heart Cast Media and their team for all of the assistance they provide with each episode. If you wanna get in touch with me directly, I can be reached at Scott with 2 t’s scott@purplefinchgroup.com Purple Finch Group is the sponsor of this podcast. We have another podcast episode out very shortly. Stay tuned.

 

Powered by heartcastmedia.com