Full episode transcript
Rising up, I earned a repute because the man who was “good with computer systems.” I’d replace my household’s PC and maintain it virus-free. I’d assist my associates burn CDs and obtain MP3s. I’d use my supposedly arcane data to put in software program for my lecturers, and edit the movies I made with associates in school.
As an grownup, I type of assumed we’d have figured all these things out by now. I believed I’d be out of date! However the extra issues change, the extra they keep the identical. I nonetheless assist my in-laws with their TV and ensure my mother’s pc is all the time backed up (with Dropbox, in fact). At any time when my greatest buddy’s cellphone stops working, you’ll be able to in all probability guess who she calls.
Nothing feels as seamless or simple accurately. Particularly relating to the best way we use expertise for work. However I don’t suppose it’s our fault. If our apps and gadgets are so highly effective, why aren’t they higher at understanding—anticipating, even—the issues we wish to do in no matter means makes probably the most sense to us?
I’m your host Matthew Braga—and immediately I’ll be speaking to Babak Hodjat, the chief expertise officer for AI analysis at Cognizant, an IT companies and consulting agency.
Babak’s crew helps different firms combine AI instruments into their workflows. It’s work he’s been doing, in the event you can imagine it, for practically 4 many years. And if there’s one thread that runs by way of Babak’s profession, it’s the best way to make working with expertise really feel as pure or easy as working with a colleague—or the “good with computer systems” particular person in your life. Simpler stated than finished, proper? However with current advances in AI, we would even have a shot.
Oh, and belief me on this: even in the event you don’t know Babak, you’re virtually actually conversant in his work. That’s developing subsequent on this episode of Working Smarter.
~ ~ ~
Babak, thanks a lot for joing us.
Thanks for having me.
To start out off, who’re you and what you do?
I’m the CTO for AI for Cognizant, which suggests I lead an AI R&D crew right here in downtown San Francisco. My background is in AI. I’ve a PhD in AI and I received into AI within the late ‘80s. I began a number of firms, one in all them led to Siri, so I used to be the primary inventor of the pure language expertise behind Siri. Though I used to be not formally a part of Siri, my crew began Siri. Then, I began Sentient Applied sciences the place we labored on distributed AI, after which we joined forces with Cognizant. In a roundabout means, I really feel like I am again to pure language once more.
Properly, we’re going to dig deeper right into a bunch of that, however I believe the very first thing I wished to ask you—and I believe it’s a great place to begin—is, as you say, you’ve been described because the co-creator and the co-inventor of the expertise that ultimately turned Siri. I used to be questioning in the event you may clarify the story behind that.
I used to be engaged on agent-based AI within the late 90s and was searching for an utility for it as a result of I felt prefer it was very highly effective. A buddy misunderstood what I used to be doing after I stated “agent-based” and thought brokers are these representatives of people with whom you’ll be able to communicate pure language and who would perceive after which go do some stuff in your behalf. After I advised him that’s not what I meant, and that pure language may be very, very onerous, he challenged me. He stated, “Properly, in the event you suppose your AI is highly effective, why don’t you apply multi-agent techniques to pure language?”
I took that problem on, labored on it, and got here up with an strategy that was very totally different than how folks did pure language. They used grammar-based and language-based approaches earlier than. This one begins from the ontology of the area that you simply’re speaking about, which may be very totally different. That led to Dejima, which was my first startup in ‘97, ‘98. At Dejima, we labored on a mission that DARPA was working known as the CALO mission, which was finished by way of SRI.
The corporate Siri was born out of that CALO mission at SRI. Our VP of engineering, architects, and quite a lot of the oldsters that have been at Dejima ended up engaged on Siri, they usually did undertake the core pure language approach and expertise that we had give you.
I am curious, your buddy had one thought of brokers and what brokers meant. What have been you desirous about whenever you considered the phrase brokers? What did that imply in your context?
So, AI began off being this quest for constructing an all-intelligent system. And within the late ‘80s and ‘90s, AI scientists realized that that is too huge an issue to sort out. So, what if we really simplify the issue? In a simplified world, use an AI system to function in that world quite than the surface world with all its complexity. That’s what they seek advice from as an agent. A multi-agent system got here out of the truth that if I’ve an agent on the World Extensive Net, how wouldn’t it work together with one other agent? How would they impart? Would they be collaborative or competing? This multi-agent idea began there.
What I used to be engaged on was imagining many “fool savant” brokers which have very simplistic worlds they function in. Are you able to get them to resolve bigger issues simply by advantage of the emergent conduct that comes out of them making an attempt to outlive in an atmosphere? Whereas it was nonetheless known as brokers, in some methods it was extra about simplifying AI.
What my buddy misunderstood was, he was really at a bar, and this girl, this previous mama-san on the bar in Japan, was making an attempt to get the tennis recreation. And he or she tried with the distant and every little thing and could not get it. So she turned to her son and stated, “Get me the tennis.” The complexity of “get me the tennis” concerned discovering the fitting channel, working the distant, and turning on the TV. The agent, on this case, would deal with all these duties primarily based on her intent. It is worlds other than what I used to be engaged on, and I can inform you, pure language is difficult. It is nonetheless onerous.
All the things you are describing—this concept of splitting the issue into all these totally different brokers, they deal with totally different duties, they do various things, and it’s a must to determine the best way to wrangle all of them and make them discuss to at least one one other—does it really feel like we’ve come full circle once more into this current second, given the issues that persons are constructing? As a result of it looks like persons are making an attempt to resolve an analogous downside now, constructing instruments that may work together with your whole totally different apps, your totally different gadgets, totally different contexts, totally different information sources.
We’ve come full circle however the the reason why we’re now working brokers may be very, very totally different. It is really fully the alternative. So again then it was as a result of our AI wasn’t highly effective sufficient. So we needed to simplify the atmosphere they operated in. That was the agent.
In the present day, our AI techniques are highly effective—the truth is, so highly effective—that now we have to comprise them. They’re so sturdy. This agent can take any totally different persona. You’ll be able to inform it that it is an skilled in discipline one or an skilled in discipline two or an odd consumer or a hacker or no matter. And relying on what persona you give it, it will have a unique conduct.
For us to have the ability to make the most of this AI, now we have to really restrict its operation to the workflow that we’re eager about. And that is introduced us full circle again to, “okay, now we have to consider these giant language fashions as brokers after we plug them right into a workflow.”
So yeah, totally different causes, however the finish end result is similar. And in reality, I believe we’re transferring in the direction of a multi-agent world as a result of now, as we arrange these workflows, we wish totally different brokers accountable and being consultants in several components of the workflow to work collectively. However the world is way less complicated. Like we will program these brokers utilizing pure language. We are able to really inform them what we count on from them. They will discuss to one another in pure language. And that’s worlds other than the multi-agent techniques that we had again then. We needed to outline inter-agent communication languages that have been very elaborate and have been hiding sure info and revealing sure different info and so forth. And you do not want to try this.
Properly, and to your level, we have clearly gotten so much higher at doing this type of factor immediately, however there are nonetheless challenges with making our expertise perceive what we wish it to do. It nonetheless looks like there is a little bit of a solution to go. Why is it so onerous for our machines to know the intent behind what we wish them to do in a pure means, quite than us having to contort ourselves to what our machines count on from us?
There’s a number of causes, one being that at the moment the state-of-the-art is a really question-answering sort of state-of-the-art, and the context that these giant language fashions have is proscribed to the way you categorical your self in pure language. Pure language by nature is ambiguous and may very well be terse and takes so much without any consideration so far as the context is anxious. And these giant language fashions don’t have anything to go by apart from the precise phrases that you simply inform them.
You simply nodded. That nod tells me that you simply sort of agree with me otherwise you recognize the purpose that I am making. They do not see that. So in fact, we’re transferring past that. We’re making an attempt to make them multimodal. We’re making an attempt to make them extra reactive to different cues. So that can assist. However that is one limitation that now we have. The opposite limitation that now we have is the truth that these techniques are pre-trained.
The PT in GPT is pre-trained, which signifies that their world mannequin is mounted. And since they’re very, very sturdy, they are often pushed to view the enter and behave otherwise, relying on what we ask them and the way we set them up. However in contrast to people or different greater order clever animals, they do not study as they go. They don’t seem to be embodied, and that embodiment actually helps us situate ourselves in a context and perceive what is going on on.
Despite all of this, I need to nonetheless say that, as people, we misunderstand one another on a regular basis. I imply, that is only a byproduct of how our languages have advanced. So you’ll be able to’t blame machines for misunderstanding. We’re not programming them in a non-ambiguous programming language. We’re actually utilizing the identical language, the identical working system we use ourselves in our group. So some degree of bewilderment is simply par for the course.
Earlier you talked about the CALO mission that you simply and a few colleagues labored on that helped to put the groundwork for Siri. And I used to be wanting again at some articles that have been written across the time about a few of that work—particularly in regards to the Perceptive Assistant that Learns or PAL that your organization was engaged on for DARPA. On the time, in these articles, PAL was described as an workplace assistant, one thing that might arrange conferences, reply emails. It is quite a lot of the identical stuff that individuals appear to be speaking about utilizing AI and particularly LLMs for immediately. And I am questioning what you make of that. I imply, is that notable that the trail ahead virtually 20 years in the past was type of much like what persons are figuring out now?
Yeah, I believe it is the case that that’s nonetheless a tough downside and has not been solved. I believe it was 12 months two or 12 months three of the CALO mission after we really arrange a system the place you’ll decide up the cellphone and you’ll inform CALO, principally, that you simply wished to arrange a gathering with these folks. After which CALO would really contact them. It could search for your contact e book. If there was a cellphone quantity, it could name them—if it was an e-mail, ship them an e-mail or no matter—and coordinate and discover the time. If it had entry to your calendar, then it could use that. It could really discover the time, block the calendars, invite everybody. After which on the day and time when the assembly was going to occur, all you needed to do is stroll into the convention room and it could really dial all people out.
After I describe that, even immediately, we do not have a system that does that that simply for us even proper now. So yeah, I believe it’s prescient solely as a result of it is an issue we nonetheless have. The unique Dejima folks wished to program their VCR. I do not know in the event you bear in mind again then, however that was a factor—”I want there was an AI system that might program my VCR.” In order that was the unique use case that we labored on was how do you program your VCR utilizing AI. You continue to do not fairly have that. Hopefully with generative AI techniques and interfaces, we’ll lastly get it, but it surely’s been like 20, 30 years now.
Properly, and also you stated that is nonetheless as tough an issue immediately because it was then. Establishing conferences, answering emails, writing studies—why is that also so tough to drag off immediately?
I imply, now we have the elements. I can inform you, we set that up as a part of the CALO mission. So it was really working. The query of why there isn’t any product that truly does that comes right down to the viability of a enterprise case for it extra so than the precise expertise. I believe it’s doable to do it. How typically do you employ Siri? We discuss Siri. I personally do not use it that usually.
Why do not I? I had this traumatic expertise, the primary time I really arrange the precursor to Siri. There was this microphone arrange in entrance of a TV set, a DVD participant, and a satellite tv for pc with 500 channels. It turned on the lights and had all this performance. One in every of our advisors, the previous chairman of Borland, was there. I simply sat him down and gave him the mic, and stated, “You’ll be able to say something, go forward.” And he simply appeared on the mic and stated nothing. It was an ungainly minute or two of him not saying something. I requested, “Why do not you simply say one thing?” He stated, “I am considering.” And he lastly turned to me and stated, “Properly, I do not usually discuss to my TV set, so I do not know what to say.”
So there’s that aspect of it as properly. Even when we put that performance in a system, will you discuss to it? Would you anthropomorphize your system to the purpose the place you’ll belief it to do what it may do, and discuss to it in pure language? How typically will we discuss to our Alexa system within the nook? It is only a cylinder sitting within the nook of the home, and I am certain it does tons of issues, however do we actually discuss to it that usually? No, it is only a cylinder. We’re not used to speaking to a cylinder.
Properly then, I ponder, the place do you suppose AI may have probably the most affect on how we work and the way we get stuff finished immediately? I imply, is it group? Is it search? Is it collaboration? Is it one thing else? The place is your thoughts at proper now?
In the present day, most of not less than the oldsters that we see are scratching the floor of functions and use circumstances. It is primarily round, “Oh, I desire a ChatGPT for X,” or “I would like to have the ability to do a doc search on my proprietary doc repository,” or “I wish to make my builders extra productive.”
However I believe we’ll transfer to utilizing this agent-based idea and begin augmenting and enhancing our enterprise workflows utilizing these agent techniques—which signifies that we’re really utilizing them for the reasoning, for his or her means to make choices sort of out within the wild and make calls to the instruments that we offer to them, just like the API that we would have, or what have you ever.
So, I believe we’ll transfer to that time. It is a leap of religion, although. When you have this expertise that doubtlessly is a data employee for you, you are able to do all types of issues. However how do you constrain them? How do you make it possible for they function in a accountable means? Who’s accountable after they screw up? There are quite a lot of these sorts of questions that we have to reply. However however, they’re so helpful they usually can accomplish that a lot that I believe they’ll begin infiltrating our organizations and and enhancing all method of resolution making and enterprise workflows.
At Cognizant, you lead this R&D crew that helps deliver superior AI options to companies. What are a few of these workflows, or perhaps a number of the largest challenges, that you have been serving to firms tackle in your position thus far?
We begin with the KPI. We’re like, inform us the KPI you care about, and let’s work again from there. As a result of we wish our AI to be aligned with you so far as what it is maximizing and minimizing relating to the KPI. I will offer you an instance. For instance we’re serving to a retailer make choices with respect to its provide chain. The choices may very well be issues like which service to rent for a specific supply, whether or not to alter the supply route, what number of runs they need to take, and so forth. However main to all of that is the KPI. What are we making an attempt to resolve for? We would like the shipments to be well timed, we wish to reduce our prices, and we wish to maximize income or another top-line KPI that we care about.
So, now now we have the entire scoping of our use case. Listed here are the actions you’ll be able to take, listed below are your levels of freedom, listed below are the outcomes that you simply’re making an attempt to optimize for, and here is the data you might have—like, what service am I working proper now, or what am I really transferring, and stuff like that.
So then the following factor we do is we really line up the generative AI-based brokers. We give them the instruments, which is, “Oh, I’ve this predictor mannequin that may predict this KPI or that KPI. I’ve this optimization that I can do that offers me some sense of what actions to take.” After which the excellent news is these brokers can discuss to one another in pure language. In order that they’re very, very sturdy. As issues change, attributes change, new information is available in, they’ll account for that. They usually can discuss to us in pure language as properly. To allow them to give us their greatest estimation as to what to do for a selected route. They usually also can reply questions on it like, “What if this service is just not obtainable?” Or, “You understand what, I did not have an excellent expertise with this different service. Do you might have another ideas?”
In order that’s only one use case, however you should use it for procurement, for manufacturing, in your help desk, you identify it. So for a few of these, you’ll be able to really defer to the generative AI’s world mannequin as properly, which is one thing you’ll do with a data employee.
So within the examples that you simply simply gave, we’re speaking about KPIs, we’re speaking in regards to the enterprise targets which are doable. I am questioning, on the extent of a person data employee, what does success seem like for a deployment for workers? Is it time saved? Is it decreasing the quantity of toil that you simply spend on a specific job? Is it rising productiveness? What does success seem like in that context?
So, a human data employee is confronted with choices they make on a regular basis. Lots of these choices are knowledgeable by their experience and their expertise, however that might lower each methods. So they could really miss sure features of a nuanced resolution level. And so really having a generative AI-based data employee on their aspect that may assist them and develop their horizon of what they take a look at earlier than they decide, and permit them to contemplate alternate options, is definitely useful. It is helpful. So it isn’t simply the effectivity, it is the standard of the selections that you simply’re impacting as properly.
And I am curious as properly, if an agent like this or a system like this may save us time, could make us extra productive, what does that free us as much as do extra of as an alternative?
A fantastic query. I can inform you, I do not know the reply to that, however each time a expertise as disruptive and general-purpose as this has come round that has made us extra productive, it has made us busier. I suppose in some methods we’d wish to look ahead to a world through which now we have fewer issues to do. However I do not suppose that is going to occur. I really suppose that we, by advantage of being extra productive, are going to have extra issues to do.
We’re all the time going to be a step forward of this. I believe generative AI is sort of a calculator: all of us ought to be utilizing it and all of us can be utilizing it, and all of us can be considering of how to make use of it to make the world a greater place—hopefully, most of us. I believe that is our position. Staying forward and mastering this expertise for the nice of humanity is what we’re all the time going to be doing.
The place have AI brokers had the largest affect in your day-to-day life these days?
I exploit AI brokers for lots of issues in my private life, in addition to work life. It is my greatest coding buddy. I exploit them to jot down code. I exploit them to make choices and decisions relating to how I will strategy essential conferences and workshops, particularly if there is a consumer concerned and I actually wish to know what’s the greatest strategy to speaking to those people.
It is superb how highly effective these items are. Like another device, after you mess around with it slightly bit, you get a knack for the best way to use it and the place it falls quick. If you begin utilizing it, you begin utilizing it increasingly and for extra issues. Now it is sort of slightly little bit of a cheat. If you run into an issue that it’s a must to remedy, your first inclination is, can generative AI assist me right here or not? So yeah, undoubtedly.
What’s one thing that you simply want AI—generative AI, an LLM, no matter it might be—may do or an issue it may remedy that it may’t but?
That is a troublesome one. As a result of these are basic techniques, there’s so much they’ll do, but it surely’s on a continuum. There are specific issues they can not do very properly. The want is extra about them getting higher at particular duties. Can they get higher at math? Can they get higher at writing code? Can they get so good, so far as their context window is anxious, that I can provide them a whole e book and have them learn it after which give me some recommendations on it or one thing?
There are some elementary weaknesses that generative AI techniques have, and that is a problem for us working in AI to beat. One in every of them is the truth that they don’t seem to be embodied, they do not study as they go. They’re very generalist, which has its benefits, however each time intelligence has advanced within the pure world, it has been a studying system that adapts and learns in its atmosphere. You’ll count on that from a machine studying system, however essentially, at its core, a generative AI mannequin is a deep studying neural community. It is very tough to have it intrinsically be studying because it goes. It’s important to play methods to make it mimic that. In order that’s one space I believe is necessary for us within the AI world to attempt to remedy. I actually do suppose that we’re a couple of main breakthroughs away from with the ability to crack that nut.
Properly, and we have additionally been speaking so much about pure language—issues like speaking and writing. However I additionally ponder whether there are different types of interplay which are going to be more and more necessary for us to contemplate as properly, particularly relating to individuals who perhaps have totally different skills, proper? Perhaps those that are visually impaired or deaf, the place the issues that we’re speaking about as “pure” do not come as naturally to them.
The structure on the core of huge language fashions is a transformer structure. So long as you’ll be able to pose something as a string of tokens, these architectures are literally fairly good, surprisingly good, even on visual-related duties. The most recent fashions which are popping out do have some multimodality. For instance, Gemini is an inherently multimodal system. So if video is taken into account a string of tokens and you may see it as that, then a video feed may very well be a part of the enter to a generative AI system. So you’ll be able to discuss, gesture, use signal language perhaps, and the system would nonetheless perceive that.
Now, the primary subject now we have proper now’s that to allow that we want very, very giant fashions. The bigger they’re, the slower they’re, and the costlier they’re to function and practice. Then again, like another expertise, initially you get the massive sluggish model, after which we as people maintain optimizing and make them sooner and smaller. That’s undoubtedly the trajectory we’re on. Hopefully, within the subsequent few years, we’ll get to some extent the place these different modalities are viable.
So all through this dialog, I’ve slipped into utilizing the phrase agent as a result of we have been speaking about brokers so much, however I do know that individuals additionally use different phrases to explain the AI instruments which are obtainable immediately. You hear co-pilot, you hear assistant, you hear helper, even companion. What appears like the fitting body to you?
I just like the phrase agent, and the rationale for that’s as a result of it really forces you to think about this method as a data employee versus a data retrieval system. I believe that distinction is essential. You do not wish to depend on its studying corpus to come back again to you and do stuff. It is pre-trained, that studying is outdated in a short time, and who is aware of what sort of biases or no matter else is in its world mannequin. You actually wish to give it some instruments, such as you would to an agent, after which give it a job and have it go do this job for you. So I do just like the phrase agent.
Co-pilot is nice if we wish to give folks the consolation that it is all the time going to be side-by-side with a human—however not all the time. Why? There are quite a lot of duties already that we delegate to expertise to do on our behalf. Why ought to each use of AI be a type of co-pilot utilization?
Properly, and you have used the phrase embody a few occasions as properly—this concept that these aren’t techniques which are totally embodied but. What’s going to it take to get there? What must occur earlier than we will have a system that does extra of what you are describing? That means to study, that does not simply depend on a pre-trained corpus of data?
We’d like some main breakthroughs to have the ability to do this. Proper now, we use again propagation, which requires quite a lot of information. And it’s a must to undergo that information again and again to nudge these neurons to do what you count on them to do, so the coaching is essentially offline. After which post-training, we do some fine-tuning. You might be actually not altering the world mannequin as a lot as simply tuning it to choose one over the opposite. And none of that may be very passable.
There are approaches which are rather more environment friendly. Evolutionary computation is one. There are symbolic approaches that, inside the world of AI, folks have labored on. And I believe if we take a look at generative AI as a proof of existence—in different phrases, in the event you can scale a system the best way now we have, and job it with an issue like language modeling, the best way now we have—there can be emergent conduct that ranges from reasoning and a few math to poetry to language to all that sort of stuff. So, that proves that there’s not less than one path to get there.
So perhaps we will use components of that, however use a very totally different strategy that may be extra explainable, could be, really, conducive to on-line studying and to correcting and modifying the world mannequin of those techniques as they go. That’s when that, paired with rather more environment friendly techniques, would get us to an embodied model. There’s a proof of existence of that as properly, which is the human mind. So we all know that can also be doable. It does begin kind of from scratch and learns and absorbs. And it may change its world mannequin. With one occasion—you present it one instance, one counter instance—it modifies its world mannequin. It is very plastic in that sense. So, why not count on that plasticity in our AI techniques?
We’re speaking in regards to the future proper now in a way, and I am questioning, what are you wanting ahead to sooner or later, each professionally, but in addition with a number of the ongoing improvement in AI extra usually?
Within the shorter time period, I am searching for extra environment friendly, extra highly effective fashions, for certain. I’m additionally seeking to see folks adopting this past the apparent. Like most of us, our first expertise with generative AI was ChatGPT. It is a chat interface, and it may do sure issues that must do with language or writing code. So a lot of the use circumstances we see on the market are sort of offshoots of that. However as I discussed, there’s a lot extra that may be finished, so I am actually seeking to that flourishing of use circumstances and the pervasive use of those techniques.
In the long term, I wish to see breakthroughs. In reality, in some methods it isn’t a great factor that we’re myopically targeted on one structure, a technique of coaching a mannequin, and one sort of use for these generative techniques. The sector of AI is way, a lot, a lot wider than that. And everybody appears to be targeted on this one path that has resulted in some fascinating breakthroughs, but it surely’s only one. And sooner or later, I do suppose that we have to begin exploring. We have to be extra artistic, in order that we will overcome a number of the challenges of this specific strategy.
I believe that is a great place to go away it. Babak, thanks a lot for becoming a member of us immediately. It has been very nice having you right here.
It has been superior. Thanks. And actually, actually good questions. You saved me on my toes.
Thanks.
~ ~ ~
One other factor I bear in mind from after I was youthful was the primary time I attempted to speak to a cellphone. Not on the cellphone, however actually, to it. It was an previous Nokia, like T9 period, that had some very primary voice recognition on board. It didn’t matter how onerous I attempted, I may by no means get it to work. However we’ve come a great distance since then. Our assistants are fairly good now! Whether or not I’m asking my automobile to alter the music, or speaking to ChatGPT, it appears like we’ve received the popularity half down.
However as Babak says, the place issues get actual fascinating is after we transcend the essential instructions, the query and solutions, and throw reasoning and resolution making into the combo. Already, we’re beginning to see AI-powered instruments that anticipate what you want and whenever you want it—with out you even having to ask. The sorts of instruments that received’t simply do what we inform them, however really perceive our intent… whether or not you’re making an attempt to prep in your subsequent huge assembly—or, simply wish to watch tennis on the bar.
Working Smarter is dropped at you by Dropbox. We make AI-powered instruments that assist data employees get issues finished, irrespective of how or the place they work.
You’ll be able to hearken to extra episodes on Apple Podcasts, YouTube Music, Spotify, or wherever you get your podcasts. And it’s also possible to discover extra interviews on our web site, workingsmarter.ai
This present wouldn’t be doable with out the gifted crew at Cosmic Customary: Our producers Samiah Adams and Aja Simpson, our technical director Jacob Winik, and our government producer Eliza Smith.
At Dropbox, particular because of Benjy Baptiste for manufacturing help and our illustrators Fanny Luor and Justin Tran.
Our theme track was created by Doug Stuart.
And I’m your host, Matthew Braga. Thanks for listening.
~ ~ ~
This transcript has been flippantly edited for readability.
