March 16, 2023

Game On! The Future of AI and Machine Learning with Xyonix | Episode #61

"Game On! Everything is going to change in the next three years!" - Deep Dhillon - Xyonix Co-Found and CEO In this episode I talk with Deep Dhillon , CEO and co-founder at Xyonix . Xyonix is an #AI and machine learning compan...

"Game On! Everything is going to change in the next three years!" - Deep Dhillon - Xyonix Co-Found and CEO  

In this episode I talk with Deep Dhillon, CEO and co-founder at Xyonix. Xyonix is an #AI and machine learning company that specializes in providing solutions for data-driven industries such as healthcare, life sciences, and financial services. They leverage cutting-edge #machinelearning and AI techniques to extract insights from data and apply those insights to improve their clients' products and services. Deep and I talk about his career as a data scientist before that was a thing leading up to the founding of Xyonix. We also get down to the 101 fundamentals of Deep Learning, models, LLMs and AI/ML in general... the good, the bad and the amazing potential it has to improve humanity.  

The company was founded in 2016 and is headquartered in Seattle, Washington  

☑️  Support the Channel by buying a coffee? - https://ko-fi.com/gtwgt  

☑️  Technology and Technology Partners Mentioned: ChatGPT, Large Language Models, Artificial Intelligence, Machine Learning, Data  

☑️  Raw Talking Points: 

  • Explainable AI? 
  • What is AI? 
  • What does a data scientist do 
  • What is the difference between good AI and bad AI 
  • How does ML fit in? 
  • What is a model? 
  • ChatGPT GPT-4 released 
  • What sets ChatGPT apart LLMs 
  • Being able to manipulate 
  • ChatGPT 
  • AI Injection Pod 
  • Improving Mental Health Using AI and ChatGPT 
  • Accelerating Development with ChatGPT 
  • AI-Generated Code Plagiarism 2.0
  • ChatGPT AI and Generative Content Concerns 
  • How do you differentiate in this AI market 
  • What does it mean to unlock the true value of data Applications of tapping into data at scale 
  • The short term, medium term and long term future of AI

☑️  Web: https://www.xyonix.com/
☑️  Crunch Base Profile: https://www.crunchbase.com/organization/xyonix

☑️ Interested in being on #GTwGT? Contact via Twitter @GTwGTPodcast or go to https://www.gtwgt.com  
☑️ Subscribe to YouTube: https://www.youtube.com/@GTwGTPodcast?sub_confirmation=1

Web - https://gtwgt.com
Twitter - https://twitter.com/GTwGTPodcast
Spotify - https://open.spotify.com/show/5Y1Fgl4DgGpFd5Z4dHulVX
Apple Podcasts - https://podcasts.apple.com/us/podcast/great-things-with-great-tech-podcast/id1519439787,

☑️  Music: https://www.bensound.co

Transcript
RAW TRANSLATION
 
it's a big deal you know like I mean I still remember the first time I saw the you know the web I was like Wow Everything's Gonna Change and really the first time I said Chad gbt I was like okay now it's now game on like everything's gonna change in the next three years and it's going to be a radical transformation of everything we know hello and welcome to episode 61 of great things with great Tech the podcast highlighting companies doing great things with great technology my name's Anthony spiteri and in this episode
we're talking to an AI and machine Learning Company that's helping organizations unlock the value of their data leveraging Cutting Edge technology to extract insights and data to apply these to improve clients products and services helping organizations once again access their data and making better use of that that company is zionics and I'm talking to co-founder and CEO Dave Dillon welcome to the show Dave hey thanks so much for having me excellent so just before we get into this amazingly relevant world of AI and
ML and data science just want to give a shout out for the show if you love great things with great Tech and would like to feature in future episodes please click on the link on the show notes or go to gtwgt.com and register your interest just as a reminder all episodes of gtwgt are available on all good podcasting platforms Google Apple and Spotify all hosted and distributed now not by Anker FM but by Spotify podcasts that's used to be anchor so that will keep you up to date with all the episodes and please go
to YouTube like And subscribe so with that out of the way deep let's um let's dig into zionics there's so much to go through today but I really want to start off by you know just introducing yourself um you know you're you're a pure data scientist I don't often have you know people are your field and and your credibility on the show with regards to this field it's some something that I think people you know still aren't totally up to date with but please just give us a bit of a background in
yourself and then talk about zionics and how you basically came to found this company which was I believe started in about 2016. yeah well thanks so much for your kind words and thanks so much for having me on the show um I got into um data science before data science was a word or before people use terms like AI um without being really shy about it um so I I you know my graduate work was um in a lot of adaptive signal processing so kind of algorithms that aren't quite like machine learning but um you know we're mostly in the audio
domain but trying to basically get models to like sort of automatically like learn from an environment and figure out what to do from it and I've been pretty active in the startup Community here in Seattle for the last two or three decades or so um I've started a few companies one of my companies was um you know back in the late 90s was a music identification and classification engine that I used a lot of machine learning back in the day that was um that stuff's actually done pretty well we sold it to a company called
Grace nodes which got bought by Sony and it's used in a few hundred home audio appliances so a lot of um yeah quite a few people if you if you put one of those apps like Shazam you put a mic in the air identifies a song we built the kind of first versions of that um that's awesome still out there um and so then after that I spent a number of years um kind of when sort of deeper natural language processing was kind of a lot more in its infancy than what it is today so we built a large-scale distributed kind of deep natural
language processing engine and this would be starting in the early 2000s up to about 20 2010 um you can the way I used to describe that system is it's like an army of seventh grade grammar students armed with a really large dictionary uh so we could do like really exacting uh queries so you know from unstructured text show me a list of all the people you know who flew helicopters near DC from just news for example This was um interesting just like some of our DOD clients and in biofarm we desk stuff like hey show me a list of all the genes
that inhibit the expression of a particular Gene like er BB2 in the mouse model or something by the way I understood every single word that you said just there but it's awesome to like hear that sort of stuff right and I think we're going to dive into you know the why I think this is so important to have these sort of conversations today because this is such a this is such a different world to what I'm used to and having you on is just already I'm just soaking up the info so yeah I just wanted to sort of
highlight that that this is where this is why we've got you on and why we've got zionics on because yeah I want to learn and everyone else wants to learn about this yeah I apologize if I nerd out too much no no please please I'll put those acronyms in the show notes yeah so getting back to zionics you know so I've been kind of like a attack Pig Zack in various startups over the years and what I found was in between companies um that I would sort of go through this mode where I'd kind of have to unplug
for a while because it's pretty exhausting I'd usually travel then I get home and I'd have no idea what to do with my life and I would contemplate being a painter or something random and then a friend of mine would be like hey can you work on this problem I'm trying to build a Purser to automatically understand you know what patients are saying about their disease conditions or it could be anything yeah and so then I would kind of like settle into you know a little side project and then I eventually you know take on a new kind
of techie kind of a more formal role and I found after my last company that those were the most like fascinating and interesting parts of my career were those in between moments where I was really just solving a particular machine learning problem for a particular domain in an area I didn't really necessarily know much about but with people who are really trying to make a difference and and and and and get something out and so that's kind of the Genesis of zionics which was can we um like if you think about a lot or a
ton of the machine learning Talent is really like the deeper Talent is locked up in some of the big five big six tech companies and then but meanwhile there's like a ton of other companies that you know don't have access to that kind of talent and don't really understand in depth what's going on they can probably take some apis and you know shovel some stuff around but they they're not necessarily able to like really dig in and understand things so I I thought can we maybe follow some of the rudiments
SAS company where you know where we can build models and have a heavy customization period on behalf of a customer and really get into their headspace but talk like exact to exec sort of a thing yeah so that you can speak this high level business language and then like really rapidly come out with very with relatively little you know explicit Direction but come out with a transformative use of AI in their products so that they can have basically an engine of innovation over many years and that's kind of the the
idea yeah so it's almost like it's it's um like you almost alluded to there it's AI or data science as a service effectively um yeah but but not necessarily at the API level it's like you know think you know a higher level you know CEO even or product leader you know or CTO who maybe doesn't have a lot of deep ml experience um they say uh you know hey like this is we have this intuition we have some data right it could be a body of documents could be a bunch of imagery could be a bunch of
recordings whatever we have some kind of intuition about something high value that can be done with it and and we don't have the resources to like realize it or we don't quite know what the next steps are or we do but we're not sure we want a third party assessment and so that's kind of the bulk of what we do and we do have like a kind of a core Mission too you know like the other thing is you know I'd taken off some time and wandered around and while sitting you know in a hut a couple of
miles from the nearest road I was like what am I going to do with my life I feel like I've spent you know 30 years building morally agnostic hammers you know you can use it to build a house for some poor family in Guatemala or something or you can bludgeon somebody to death and it feels to me like 99 of the tech companies out there are morally agnostic Hammer Builders and I didn't want to do that and so I thought we'll make a simple test if a five-year-old girl can answer a simple question is the
world better off if we do this if she says yes we take the project if she says no we might take project because we might need the money or might find it intellectually interesting and fascinating still but we often times don't so like that's kind of the so hence we have a lot of projects in healthcare yeah education I saw that yeah that's really commendable right you're not only looking to to do something with awesome Innovative technology but you're looking to actually you know make a difference as
well I guess is is what you're trying to say there and and we'll talk we actually will talk about good and bad AI a little bit later on but yeah I think so what I'm getting you know zionics is all about you know tapping into some sort of data that exists you know that's being just generated by by just attritions by the natural course of a business doing what it does or an organization doing what it does and then trying to basically lock into that and seeing if you can derive some sort of value out of it with with the skill set
that you've got as a data scientist and your teams that you've got as well and then you're basically building the models and then effectively you know running with that for those companies and effectively looking after that whole side of it so they don't need to worry about you know hiring that Talent which you said was effectively somewhat locked up and that's quite an interesting when you say that is that just because it's such a new sort of part of our of our world that there isn't that many data
scientists to go around or is it just a case of that the good ones are all taken effectively I think it's mostly the latter yeah but I mean if you look at the salaries that like you know Google and Microsoft and Facebook are giving like you know just a bunch of friends of mine that I've worked with I mean they're making well over 750k a year and so for a startup to kind of compete with that kind of income or even just a regular industry job where there isn't necessarily the DNA to stomach such a salary for people which I
you know I don't blame Folks at all I mean it's just a completely insane amount of money yeah um and you know part of the reason they're you know the big tech companies are paying so much is because there's just a potentially massively disproportionate amount of impact that you know that the top tier Talent can have on a problem and it's yeah I mean it's it is just kind of what it is it's a combination of limitations on Supply and and really deep pocketed uh companies who really get it yeah it's um kind of people
listening a lot now thinking should I get into that with those sorts of salaries that's that's top in Australian dollars that's well topping a million dollars a year right yeah there you go no interesting space but I think for you guys you know obviously what you're doing you've got what sort of service at a very sort of you know top level like what sort of services do you offer like products or Services obviously I've looked how to look through your um your website but Predictive Analytics data visualization
data management tooling um is that kind of your bread and butter or is it or is it actually all mostly based on the model and we'll talk about what a model is a bit later on worry more like a if you think of a design agency you know like we have a core competency where we're good at machine learning and AI particularly formulating problems formulating strategies bringing them working with product um with product managers and Engineers to like integrate the AI capabilities into products so that's kind of our our
sweet spot excellent particularly uh we do pro that said we do projects and in lots of different types of data so we have a number of text projects that you know use large language models um you know like the stuff that chat GPT is built on and we also have uh done some projects with audio so for example we built a smart stethoscope um for a company that had a an iPhone case based stethoscope we've done a lot with video so for a number of years we built like in in surgery room in body video analysis capability to just sort of see like hey
is the physician suturing now are they cauterizing or placing mesh and helping ultimately power a you know a surgeon a surgeon Improvement feedback system so you know and then like more recently you know we have um we've got a project where we're pairing up it's like match.com for doctors and patients so that's kind of how I would describe it yeah um that that's that's one we have another one with a you know a company that uh works with college students and uh sorry with colleges and it's it's
like a chat bot for students from the time they accept till the time they graduate kind of addressing all the Practical question answering like when does you know second semester start when spring break to you know hey I'm not feeling good and yeah you know I've been sad or whatever so it's like it's an array of projects yeah yeah at the end of the day what binds them all together is there's um data there's some ins some intuition that something high value can be done with it there's some patterns naturally
inherent in that data and unlock talking them is very valuable yeah and hey I don't swear often on the show but that's some really cool like you know you know what I mean so yeah but on that let's let's move into this because what I what I wanted to sort of grasp and for people out there who obviously ai's been thrown into the Forefront of our world um recently you know with chat JDP just absolutely dominating the airwaves and obviously like you said I mean it's this this science has been around for decades
right um you know because it is obviously just a thing when there's data you can see inside so you can get Telemetry out of it you can grab you know tangible information out of that so what is AI so this is my first question to you like how would you define artificial intelligence in this field I don't know how to answer that question so I'll try to answer from a few different angles stop the show as a machine learning you know practitioner I think of AI and I'm a little bit old school but I think it was
this goofy term that people who aren't into machine learning use to describe what we do so um I would say that's kind of part of it but that was sort of pretty deep learning you know where we really didn't think that the place that we're at today would happen in our lifetimes wow um I think I think at its core you know when we talk about this generation of machine learning and AI systems we're talking about systems that ultimately um have the ability to recognize patterns inherent in the data and make
projections based on it so one kind of key category you know within there we call supervised machine learning so you can think of it as like example based learning so the example I often give is you know imagine you have a toddler who is coming up to speed on language and doesn't yet know the word Furniture so by example you might show the toddler a chair and you might say furniture show them a table they say you save furniture and then you know maybe you show them the sofa same thing you go outside and
you point to a wicker chair and you say what's this and they say not furniture and you're like well it turns out that those things can also be outside and so like so through example you're giving the system um examples that Define the thing and maybe Define the not thing and so that's an example of classification based super revised machine learning which is a lot of what we see and then there's a whole other category within AI that we call unsupervised machine learning which is like hey you've got this bunch of data
um tell me stuff about it so like let's say you have a population of a bunch of customers who and all their shopping habits their buying habits and uh and then you kind of dig in there and you find out some natural patterns in the data so you might find well okay well this chunk of folks are like quick buyers they jump on your website they buy really quickly and this chunk of customers are you know really hesitant and they wander around they browse a bunch that kind of thing so and I think more in the last you know
since deep learning came out sort of one of the big breakthroughs if you will was to sort of on some level kind of marry the two so to leverage the lack of structure so let's take text as an example if we think about chat GPT and large language models one of the big breakthroughs was to take like a big body of unstructured stuff and instead of giving it little itty bitty examples of furniture and or how to read particular sentences in a particular way or how rating you know sentiment or something on IMDb reviews
or something like that instead we came up with this idea like well let's just um teach the thing to predict future sequences of text so given some text it predicts the next word so um uh he went to the blank okay maybe store maybe um you know like there's a few words that fit into that sentence but there's a lot of words that don't like you don't go to the orangutang or you know yeah okay so so it turns out when you when you do this and then you combine it with a neural network that has sort of a
certain properties and topology that it gets incredibly bright like it it turns out that you in order to do this task well and predict future sequences of text you've in essence learned language not just English but all the other languages that we're training it on from all this content on the web and not just human languages you're also learning programming languages and you're learning how to you know Common all those common patterns and that's kind of the like base layer if you will of like what you know chat
GPT and this new generation of really smart chat Bots is is kind of leveraging yeah so you talked about AI not being equal to ml and then you talked about deep learning as well so in terms of machine learning where does that fit because obviously Ai and ml get clamped together almost all the time so you know is that is that just because of something that's just become you know habitual in the industry or other do they are they obviously for my end go together but I'm trying to work out if machine learning is is actually what you
guys are really doing using that deep learning and AI is just the overall term that's given to it I honestly I just say like if I'm talking to practitioners we use machine learning and if we're talking to if we're trying to like if it's marketing like jazz it up everything's AI there you go yeah there is no I mean from a technology from a technologist's standpoint AI is like the marketing a term um or but some people can kind of use it for these much larger models that are more generalized and doing stuff that's
like wowie you know as opposed to kind of more narrow stuff but even these really large models people are now doing narrow things on top of them and does that make them not AI even though they're using the same quote you know mind or brain if you will so I don't know I mean it's largely it's largely like terminology that you know that kind of flies about but most technical people use the term machine learning to mean much the same thing yeah because I kind of and to you you've actually articulated it beautifully
right in terms of you know there's a there's an element of marketing there's an element of excitement there's an element of fear that gets brought in with AI like I think that's a very real sort of thing that we feel with it because we've been we've been trained ourselves if you want to you know kind of almost put it together we've been trained to fear artificial intelligence almost all of all of my life I know that I've been trying to fear what happens like all the robots rise AI is
effectively always going to end up in a really bad way I've I've read a couple of books um overseen and a few other ones that talk about the side rise of cyborgs and so I think Ali is used to Leverage The the emotion of what it is and then the marketing angle um but yeah that at the at the back end of that it's all the cool stuff that you guys do as data scientists so as a data scientist like how would you for anyone starting off how would you classify that like how would they go and I'm assuming
at universities now there's there's data science courses and whatnot but is that kind of the path to it is that how you got to it or did you start in some other way to become a data scientist and work where you work yeah I um as far as me personally I came up through a field electrical engineering right and within electrical engineering I mean I'm this is showing my age a bit but you know Digital Signal processing was really big so you know like so I was kind of looking at algorithms like adaptive noise cancellation you
know like how those algorithms work and so um so like the maths all really similar where it's not very difficult math it's generally like linear algebra and you know and some some probability and statistics are valuable but people people have kind of come to machine learning from a lot of different disciplines you know at least in my generation of folks whereas now I think a lot are coming from straight CS background but um I think you're almost like you know I mean I've we've had clients um I was working closely with a couple
of uh professors at Columbia University they have like one runs a history lab that where they they literally are all like Political Science and History grad students they're all doing machine learning yeah and you know another one that's that's in the literature Department you know like analyzing texts and using the same techniques on classical literature texts as uh as they as they use on you know analyzing social media uh posts around vaccine hesitancy for example yeah and so Ai and machine learning techniques are in probably
every Department in a graduate school in a University at this point and so it really depends kind of like what do you want to do right like if you want to be coming up with new network topologies that are going to be really transformative and you want to really get at the most core basic part of the models than probably computer science is your best discipline and um that's probably the place to you know to really focus and just take a lot of extra math and stats and machine learning classes yeah which is cool my
son's doing linear algebra right now I'm enjoying doing it I'm not saying he's gonna and he enjoys it as well but yeah it's good I might tell him it's uh it's this is this is something like you guys what I wanted to do this Dad I'm like well you know in the future this is kind of the stuff that's at the core of a lot of what what's what's going to drive our world basically right yeah it's impactful stuff you know I mean this is the the you know the Dirty Little Secret of grad school is most of
most folks are using math from you know 100 years ago or earlier I mean it takes a long time for today's mathematicians work to be harnessed in the real world so so yeah but um that's what that's what I'd say and then there's you know of course to be a good data scientist you know like being able to program is is key you know being able to you know know how to write software so I think that that angle makes sense uh to spend some to get really good at I think that's what really differentiates the kind of great
folks from the maybe more um okay folks is can you like take your ideas and get them from concept to reality in short order and that really often comes down to different types of programming skills you know like having some basic you know being able to be able to jockey some Python and transform your data is one thing and certainly valuable but you know being able to operate in a distributed computing environment and spin something off onto a few hundred machines or a few thousand nodes and get something done is another area of
computer science to get decent at and then there's you know the actual algorithmic side and then there's just a lot of intuition it's it's very much still art more than than than anything else so it's almost like you're going to be a combination of jobs and bosniaki effectively I think that's what keeps it interesting though you know it's like why I mean I was chatting with my wife the other day and I'm like yeah I don't I feel like I'm supposed to do something else with my
life in addition to this but I really don't know what else it would be and every time I you know try to do something else I just wind up back here because it's just too interesting maybe that's something you can chat about with uh from a behavioral therapy point of viewp which I know you've done in the past um that's really interesting in itself I I just wanted to talk a little bit but you've mentioned model a couple times and I think that's the one that we hear quite often from the outside so so
what is a model yeah so let's so let me see if I can come up with a simple example so like everybody think about your you know your eighth grade uh math class where you might have you know had an X and a y axis and so your X is like a variable like let's say square footage and um the Y is a different variable um you know let's say number of stories and we're describing a house right so in this little X Y axis like really big mansions are going to have a lot of square footage presumably and maybe a
number of stories and really tiny houses might be like one story with not so much squares footage and if you look at that across like a bunch of you know houses um now you can make a quote model where a given square footage and um and number of stories you can quote predict you know know what the let's say let's add another variable like price you know what the price might be yeah and so the so the model is sort of the act so like you can imagine in a simplest case like everything kind of Falls along that diagonal like those
points like you get bigger houses you have more square footage and maybe it like jumps up a step every time you add a new floor you can throw on another couple thousand square feet or something and so you wind up with some kind of curve or function or some some smattering of dots and like a simple line like a linear regression you just pop a line on there and now your model is just that line um and so you know given one or two variables You can predict the third so when it comes to machine learning it's basically that but instead of like those
nice little two-dimensional things that we can Envision or we could in like eighth grade math you know now it's you know maybe three or four or five or a million or a billion dimensionality so it gets to be a much larger dimensional space and maybe we can't just throw a line on it so we have to have a network that learns how to minimize the error from the thing it's predicting to the other thing and that that all that stuff we call the model the act of training the model is training the part to be able to do the
prediction and then the inference is the the act of once we've trained the model being able to give it some input like in this case the square footage and the number of stories and have it spit out the you know the price or some category or whatever awesome no that's even I understood that so that's that's good so yeah I think that's a great way to look at it um let's let's move quickly to you know chat JDP because obviously I think you've mentioned it a couple of times we know that you know in December it kind
of took over the world um and I think from that point of view it's just brought AI uh to the Forefront right and people are generally excited and scared about it so um you know what it's what sets open ai's check JDP apart and obviously there was mid-journey as well for image processing which was crazy in itself um but yeah what's for you kind of had chat jelly Pages come on and take the World by storm like it did yeah I mean I think that's a good question because we've had large language models for a few years now and
for those of us who are familiar with working them with them day to day there was nothing super radical about chat GPT but they did put a really important kind of reinforcement learning layer on top so one of the things about so so remember that model I described where you can predict future sequences of text so you take that thing and you can ask it you know like questions um and it you know like all the kinds of stuff you would ask chat GPT and it'll give you like you can ask it repeatedly and you'll get different answers each
time and a good chunk of time it will wow you and a not insignificant chunk of time it will completely butcher it right and so you kind of have to really tailor it and so what openai did was they actually put a whole layer on top that is this reinforcement learning that said it and they they as far as I understand it they you know they hired tons of folks I think in Kenya to actually do a lot of this labeling nobody really knows for sure they've been pretty tight-lipped about most stuff but they but they you
know they just started like asking it you know a lot of uh questions and it gives answers and then these humans are like selecting towards the you know the optimal um answers are like high quality answers across a set of normal large language permutations and so that layer of uh on top of the llm is like a really important layer that uh other folks are now you know racing to reproduce yeah and um yeah right so I would say that's like a good part of what makes Chad GPT different and the other part that I think was pretty
um important was just simply having a a simple little interface on it yeah so a it's a text box it's super intuitive but B um so before chat gbt with the llms you know you had to you had to kind of do these promptings so you would have to give it a few examples it was a little bit nerdy like you had to use some you know some kind of cryptic stuff and they the chat interface it turns out is pretty good at like kind of coaxing the the prompts out and it's really intuitive for humans and um the other
thing is they kind of maintain some state so so it retains some of the prior information that isn't necessarily what's happening with this kind of like one and done um llm stuff so I think all of that kind of contributed um to oh and that and then I think they also put some guard rails on it which I think probably it would have taken off without the guard rails um but that's a lot of work actually it's a tame these beasts because they can say completely a name and stupid things so yeah because basically I've
just seen it chat jdp4 is starting to trend on Twitter so there you go it must be released um but no but I think um to that end I think if the guardrails are important because obviously you know the data that it's sourcing yeah there can be a lot of dodgy stuff in there right and I think that's part of the art of what they've done it's been able to protect the responses to a certain extent certainly people have been able to you know to close it into you know giving some you know some some dodgy
answers and some not so great answers because that's the nature of of the Beast right but I I know actually it's a great example from the coaching was that I actually asked it yesterday you know tell me some stuff about zionics right and at first it said I don't know about that company and I was like okay that's interesting and then I kind of said well hey zionics do they're an AI and ml they're based out of Seattle founded in 2016.
and then it came back and said oh yeah I actually I do know a little bit about it you know sorry about that and actually like apologize to me and then it basically started to get a little bit more information but that first information was basically what I fed it okay and then I said are you sure you don't know anything more about zionics at that point it's like and I was surprised at this because I didn't know I always thought it was a finite sort of cut-off date and whatever but it basically then came up and said yep here's all the information on zionics
this is what they do this is the services they offer so it was interesting that at that first prompt I'm not sure if it was a competitive thing or whatever but it basically just said you don't exist and then I had to basically manipulate it a little bit to say hey it actually does exist for what what what part of it is that all about because that's quite interesting the whole whole coaxing thing yeah I don't know for sure but my guess is um they you know when they Define these guard rails like in addition to you know
really assigning credibility to content so which is one of the reasons it answers things so well is they they I think they did spend a lot of time and effort on that it seems clear from the output but um but in addition to that you know those guard rails part of their guard rails was to kind of steer away from instances of things so if you asked Chad gbt at least pre um gpt4 or so like 3.
5 if you ask it about specific politicians or specific um companies or financial like anything that gets too much into details their guardrail system kind of tries to steer it away from that and I think that's their kind of first major attempt to avoid being accused of giving Financial advice or avoid being confused of giving doctorally advice you know or things like that so I'm guessing that that's probably what is going on there so yeah and look just on that obviously you know when we're talking about Good and Evil
um to put it into that sort of biblical term and we talked about the fact that AI breeds all this emotion what I'm saying is a proliferation of AI companies out there at the moment it seems to be the in thing and I've I've kind of linked it to the D5 crypto craze where everyone was trying to get in on that and make money and obviously there's always some Bad actors in any space and any technology that comes out so in your sort of opinion what's the difference between good Ai and bad AI you mean from like a morally good or bad
standpoint or do you mean from a quality standpoint I think I think both work I think you know if I I think you can answer it as both parts right morally yeah let's start with there because obviously that's quite interesting and then I also think just from the point of view of people just trying to cash in I'm trying to say that they are an AI company which may be at the back end they're all just using the same the same system effectively and that's trying to try to sort of put only skin on it so to
speak so from a moral standpoint it's not obvious how to think about it exactly because it's it's kind of complicated but I'll tell you about the ways I think about it um there's the it's a hammer um analogy ask yourself are you building a house for you know a family to live in or are you bludgeoning somebody to death and that from that lens it's pretty obvious you can use machine learning to you know help you break into people's credit cards or you know all like all kinds of Nefarious activity and you can
use uh you know AI or ml to do all kinds of good things like identify heartbeat anomaly and anomalous conditions or make you know a physician a little bit faster or help address you know help a therapist be a little more efficient or whatever then there's there's like a different so that's that one seems kind of straightforward um but there's there's another lens that I think is a little bit different it's like um so you've got the the hollywoody thing which is you know you've got this
Terminator like Vision which I think is so grandiose that it's kind of pointless to even think about because I think as far as an arch villain goes that's not how it exactly happens it sort of happens more innocently right like so the example I'd like to give is if you go um back in time a little bit you know on in YouTube on Google and you look at um so somebody did an experiment where they started with a campaign speech of uh Hillary Clinton uh on the left of center and uh Donald Trump on the right
and you just looked at the recommendations that come back from this system right so you so you take that campaign video and then you click on one of the recommended videos at random watch that video click on one of those at random and it turns out in both cases you wind up going down the Allison Wonderland rabbit hole and at the end you're in completely insane conspiracy theory BS land yeah and that is an example of real harm that was done and if you ask for the causes of it conspiracy theories will come up with all well you know
those evil Folks at Google tried to whatever but that's not really how it happened right like what happens is you have to come up with a cost function to optimize right when you build these algorithms you have to say like well hey what am I optimizing for and at YouTube obviously they're probably optimizing for engagement like they want people to watch as much as possible because ultimately their money comes from from ads and they're running ads and nobody sat down and said hey let's make an evil
you know recommendation engine that takes that starts spewing nonsense but the models are so powerful and so um good that they figured out that if you want to optimize engagement the way to do that is to give somebody something slightly kookier and slightly weirder and slightly scarier than the time before yeah and if you do that then you get them to stick around you also get them to storm the capital you know and just be Nutters like beyond belief and that's the kind of evil that I think is very real and very problematic in every
data scientist needs to push back hard with their product managers and others and all every everyone in Tech needs to like like figure out where our Collective you know where the good conscience lies yeah that's that's really interesting in itself I know I listened to a couple of interviews Mr based um who's obviously a top top YouTuber in the world right from that point of view and he was openly saying that he knows that he has to put some some level of negativity um or at least excitement that is on the
negative side into his titles and into his um thumbnails as well because that's what draws people in and that's what sends people down that rabbit hole so yeah you're spot on there it's it seems to be you know the fact that if it's driven by the algorithm driven by money then there's the potential for it to be not that great so from a moral standpoint you know everyone who's in this in the space needs to kind of be aware of that and you know be okay with effectively what they do right people
won't end up you know getting paid for their their service their specialty and I guess if they're being paid to do some that a company wants them to do they might justify that in themselves and therefore they can't avoid you know the the nefarious activity that might come of it um but you're kind of saying that in the back of everyone's mind in this space they should be aware of the potential impact yeah and I think it's more than in the back of your mind right like if we look at the area of you know ethics
and AI which is sort of a really important emerging area there's like there's it needs to be more than just in the back of your your head like you have to really ask yourself how is this algorithm going to bias negatively particular populations of humans too like that's another aspect right like if you're training up you know an algorithm to like automatically label people as beautiful or not I don't know like some weird dating app and you only use people from a particular you know ethnic
demographic group and then you take it you know into a you know like some other completely different place like you train it in Norway and you take it to Botswana or something and then it always says like not not hot not hot not hot well you know you can argue like well oh well the bio algorithms obviously bias it's like no that's on the data scientist and that's on the product managers because that's an obvious thing to like think through but you just chose not to think it through and your startup
chose not to like it chose to like or your company chose to move it into a new area without really thinking things through so I I think that the data itself and the potential like bias is what these models do I mean that that's the point but there's there's you know harmless bias and then there's like potentially harmful bias like another example you know that I think is is out there so you know I used to um uh you know work for a company that did a lot we we build a whole platform to help governments
um you know mine data effectively and uh make it available for other people to build apps so you know stuff like hey you know all of the taxi license data in the city of Chicago would go online and then anyone could somebody built an app for an iPhone to check on whether or not you have a license that day I mean like it's it's kind of a weird Quirk of that particular City yeah but that you get a license for a day as a taxi driver but um but there's like another application which is like this Arena called
predictive policing so you know you turns out you can pretty accurately predict where crimes are going to happen and even when they're going to happen and so cities to their credit said like hey you know you you can't use like ethno-demographic data in those models to predict that crime so everyone thought okay well we're off the hook as long as we use like a lat long coordinate uh you know some some temporal history and the magnitude like the crime type uh and then the the crime location like we're good well it turns
out that those are still like proxy variables for all this other stuff and so what's going to end up happening is it's going to end up routing the police to go sit in a few like camp out in a few neighborhoods that you know that other variables that you in that you rightfully so omitted from the data set are still being represented even though the data looks sort of tame yeah and that's very minority reporter that's that's totally fine and there's companies that that make predictive policing apps and it works so crazy
yeah that's cool hey we could we could talk about a lot of this stuff for another couple of hours but I do want to kind of try and sort of wrap up and tie it back to zionics and you know you've obviously you've got this wealth of knowledge in this industry which you know is amazing I've learned so much just in this last sort of 30 or 40 minutes chatting to you um I just wanted to get your idea from a zionics point of view like what does the future hold for zionics as a company like short-term
medium term and then long term if you can if if we can go that far in I'm not going to give any dates on that so I'll let you kind of detect what that means yeah well like in the short term and medium term you know we love helping our clients um integrate machine learning and AI capabilities into their products that you know are the type of products that make the world you know a better place at least uh generally according to our you know our sort of ethical anchoring and um so that's what we do today
um being able to do that more like with more companies and more potential impact and uh and have a net even greater positive impact on the world is you know where we want to be you know in the future um I you know I see us over time getting you know a lot of uh sort of helping our customers you know fish for themselves you know and being able to to teach them to fish a lot faster and uh and get sort of a lot more of this capability of like ramping up a project really quickly getting training data really quickly
getting early outputs really quickly um being agile about it getting it integrated in your product getting it tested uh and validating your business hypothesis like all that methodology that's a part of our DNA and what we sort of do when we employ a project being able to like come up with you know capabilities to facilitate others being able to do that more so I feel like if we're successful like AI That's ethically considered and very impactful and filtered towards problems that are very kind of net beneficial to humanity
I feel like if we're successful then there will just be a whole heck of a lot more of that in the world meaning a better world that's less people getting sick faster safer more efficient surgeries you know um more predictive uh sort of Assessments of health conditions whether that's from like you know your blood or your you know your stool which sounds gross but that's you know that's a project um I've already said shoot on the episode you know yeah so like like all of our health indicators and being sent into to
look up a particular problem that is latent you know but now can suddenly um you know if addressed early you know extends your lifespan by quite a bit like all those kinds of and there's just the The Cape the possibilities are really just they're just so broad so you know we hope to play our small part in that you know that Evolution it's it's great it's on one level it's grandiose in a sense if you know what I mean but I think it also represents just how important and how critical this area is
to humanity in general right and I don't often get deep on this show but I think you've you've painted a really good picture of you know AI for good solution Made Simple it's kind of your catchphrase right from his ionic's point of view but generally speaking um this is where I think this field you know should be more excitable than scary for a lot of people um that are just thinking about how AI can impact you know not just the world but also our own personal daily lives and our lives moving forward for
ourselves and our kids and whatnot so that's why I'm really interested in it you know I'm trying to consume as much information as I can on this because I know that I can't contribute directly like like you can and like your company can but I certainly want to know enough about it so you know when it comes time to take a little bit of an advantage from it or even leverage it for something good that I can do I'm going to be ready for it at the same time so it's very commendable you know what
you're doing you know your life's work effectively here you kind of you questioned yourself a little bit in terms of what you you know what am I doing but I think you're doing great work and you know I'm really you know pleased that I had you on the show because I think it was just an awesome conversation not only about what zionics are doing but just Ai and general I think people that have listened to the show now have a really good almost layman terms understanding of what the AI is ml you know deep learning models
or everything that we got into even llms I know that acronym that's large language there you go awesome now hey thank you for being on the show um I'm just going to finish off by saying again once again if you're not subscribed or you're new to the show and you like what you hear you can go to the podcast client and subscribe go to YouTube hit like And subscribe there or go to gtwgt.
com and again register your interest there all the episodes are on there this episode will be there with the show notes at the same time thank you deep thank you to zionics and we will see you next time on great things with great Tech awesome thank you so much for having me it was a blast beautiful no worries and cut