Alan from Tryolabs - AI Is Not Magic: Here’s What Real Deployment Looks Like | DSH #1768

It's like within the species I can tell like if this lion is Bob or if this lion is Alice based on the whisker pattern

So they will conserve the whisker patterns since their cubs and the idea here is like

these organizations have a data set of many photographers of lions taking through the years

and whenever there's a new photo we can actually use a eye to match it to the existing database

so they can see all this lion was actually found in another reservoir like hundreds of miles away

and now it's moved here so that gives them data to be able to protect those lands

so fast yeah and those are kind of the use cases of AI that AI for good of the world that's

it's something that is really interesting for us

okay guys we got Alan from trial labs here we are at the AI 4 conference great to meet you Alan

what is thanks so much yeah what is trial labs about so trial labs is an AI consulting

and services company we've been in the space for 15 years so we before AI was a thing that

everybody was talking about in fact AI was called machine learning back then so we started around

2009-2010 initially serving clients in the Bay Area startup San Francisco which was the only

place in the world where there were crazy enough founders to kind of use these technologies

and through the years we've grown nowadays we serve mostly corporate and enterprise clients

and also some big non-profit organizations and what is the exact service that you're

giving these clients so we help them build custom solutions using AI and data to get some

desired outcome in general it's like business outcomes but it can also be some outcome that's

good for the world or some other initiative that they're pursuing right so yeah these companies

have like massive amount of data there's a lot of things that they need to hold and the application

of AI for these problems is something that's actually non-trivial and it's not just a technical

aspect there's of course a big technical aspect to that but there's a lot of like processes

things that have been always done in a certain way there's people involved and there's like

complex systems and you need to make AI fitting those complex systems and actually provide

business value for them interesting and you guys have been featured in a lot of major outlets

for solving interesting problems yeah we actually have a depending on on specific initiatives

that would pursue there's some that are pretty pretty high profile yeah sorry

there's there's initiatives that are pretty high profile so we've done some work with unicep for

example around the impact of heat waves on developing children we've also done some other things

that are super interesting for example a couple of years back we did an initiative with unicep

when we published a paper on this which was how do you use satellite imagery to detect schools

so the problem that that the world has is that in most developing countries the schools are not

100% mapped by the government so like there's countries in which like they don't know more than

50% of the schools where they're at that's crazy so unicep has this initiative called giga

and their intention is to find these schools and go and connect them to the internet so that

these children can have the best education possible and what we're doing is taking satellite imagery

and turning custom computer vision models to actually understand what buildings are actually

likely to be schools and map them that's fascinating yeah we're also doing some interesting work

with an organization called the Nature Conservancy around the sustainable fishing practice so I think

it's like close to 38% of all the fishing stocks in the world are overfished and that is a big

problem for conservation of ecosystem and there's also like push from governments in the regulation

of electronic monitoring and also push for from some retailers like Walmart for example they

announced that every fish that will be in the shelves in the next year or two will have to be

fish sustainably so what we're doing there is like we're putting cameras on the fishing vessels

and whenever they're they're fishing we can classify like what's catch like the intended catch

on the bike catch and get independent metrics wow so yeah this is an industry that the electronic

monitoring industry exists for a long time but the problem is that the review cycles for the videos

takes like several months so you ship a hard drive and then three months later yeah you have over

a fish tier and clearly you cannot act rapidly with that so we are shortening the cycles and

making sure that this ship can report independent metrics in real time yeah 38% is a lot

is that a worldwide assure that's a worldwide issue for sure okay I was reading the other day that

even in the that is like putting selective pressure on some species so the fish are actually shrinking

due to the overfishing so imagine that these smaller individuals they can kind of escape the nets

so that makes the species overall like through the years smaller and that also of course

causes problems for conservation and yeah interesting so that's a major problem yeah you guys

also track some lions too right we have done that too there's an organization called lion guardians

and basically the issue is like conservation is in Africa they they need to understand where

the lions roam to be able to protect those lands and to prevent like human development in there

and the issue is that tracking lions is a complex endeavor there's like two tracking mechanisms one

is very invasive like you put a color on the lion so you have to go there yeah good luck with that

you have to go there set eight the animal the colors are expensive then you need to replace

a battery every a year and there's the other method which is tracking that's not invasive so

photographers will go with a telephoto capture these photos from very far away and in terms of

that the lions they can be uniquely identified so this is like within the species I can tell like

if this lion is Bob or if this lion is Alice based on their whisker patterns so they will conserve

the whisker patterns since their cubs and the idea here is like these organizations have a data set

of many photographers of lions taking through the years and whenever there's a new photo we can

actually use a eye to match it to the existing database so they can see all this lion was actually

found in another reservoir like hundreds of miles away and now it's moved here so that gives them

data to be able to protect those lands so fast yeah and those are kind of the use cases of a eye

that they're for good of the world that's it's something that is really interesting for us one of my

favorite ones was my friend Walter O'Brien he used AI to solve the Boston bomber at the marathon

wow search through thousands of hours of footage and they were able to find his his patterns on

how he was reacting because everyone else was running away while he was acting casually wow

crazy right yeah that's fantastic I mean you also hear the kind of dystopian stories about like

what the potential of that technology is like look at China and what they're doing right mass

surveillance massive surveillance yeah but I feel like underlying the underlying technology is not

not bad or or anything like right if it's used for a good purpose I think it can make a huge

difference agreed yeah I am worried about mass surveillance right I think everybody got cameras

everywhere yeah have all the traffic lights they do they do I mean there's there's of course

another aspect of the AI that's called edge AI so there's like some specific algorithms that can

run on device so for example retailers can use that for analytics winning the store and that

doesn't necessarily have to identify any individual person they will just like count food traffic

understand like even classified gender or something that the retailer might care about around

how people are moving on their stores and that does not mean that the fact that there's a camera

that's not mean that you're real surveilled this can be something that runs on device and then

just like reports aggregates statistics oh clearly depends on how you actually implement that

yeah that to me is more I guess approachable right yeah rather than someone watching you all day

wherever you are correct yeah I'm good on that you also used your technology to track some fires

yeah we did actually we worked a couple of years back so you everybody knows that in California like

the wildfires are a massive problem yeah that not only jeopardize like property and value but

also human lives this affects community deeply and we worked for this startup in San Francisco that

went on to raise a lot of money in developing a system that can detect the early signs of wildfire

so these these companies in selling cameras in many different locations they can actually triangulate

so where we see signals of smoke they can triangulate and call the fireman so in the very first

moments of the of the fire wow and if you can like actually detect it on the first five to 15 minutes

like the chances of it not becoming a massive wildfire is is extremely increment so so the idea

here is that with the eye we could train custom computer vision models to detect signals of smoke

in the early minutes of the fires that are very difficult to detect for the human eye so it's

interesting that when if you see like the photos of how they look like yeah you would never be

able to tell there's a fire but with that like the algorithm can say oh there's something here and

when you see the video play out yeah there's a tiny signal of smoke over there wow and the

interesting part is like how do you differentiate smoke from fog or from somebody doing a barbecue like

there's a lot of challenging aspect that come with with developing these models like you you

start to generate some false positives and if your system generates too many false positives then

nobody will pay attention to that so you have to kind of balance it out in a way that it's actually

useful but it's like it's it can be really accurate and these companies expanding internationally

they're having like great success nice so the one that happened last week in California did they

track that on I'm not sure like we were we developed the initial version of that solution and then

left them to its original because that's something that we do it's not just like about

building the solution and staying there forever it's like we can do what's called project delivery

with knowledge transfers so we will train their own teams to continue developing the solutions

through the years what was the most difficult model to train out of your 15 years doing this?

that's a great question I think the fire one is particularly tricky because it's like the data

collection effort that needs to undergo a VAD in order to to have a nice ratio of false positive

that is manageable is really tricky there's like any single phenomenon like there's a bird on the

camera that it can really things that you you don't think about can happen also like lens flare and

they can fog or some specific pattern of clouds and then you take your model to another location and

maybe there's snow and when you're training California in these parts there's never the models

has never seen snow so there's a lot of custom things that can that can go wrong and there's a

long tail of data that you need to collect for it to be really accurate so I say that's an

interesting one of course there's many more areas in which we've worked so there's like

work that we're not in forecasting like how can a retailer forecast how many items they need to

buy so that they don't run out of stock next week and there's a lot of complexities in that kind

of scenarios but but yeah I don't know if there's like a single most difficult model but this is

something worth mentioning yeah I'm sure you're working on so many different models at the moment

we are yeah I bet business is really booming business is thriving so we're having I mean we've

been in the space for 15 years and that's every company under the sun right now says they do AI

and they're all like mostly users of chat GPT and similar APIs like Chennai APIs shout out to

today sponsor quince as the weather cools I'm swapping in the pieces that actually gets the job done

that are warm durable and built to last quince delivers every time with wardrobe staples

they'll carry you through the season they have false staples that you'll actually want to wear like

the 100% Mongolian cashmere for just $60 they also got classic fit denim and real leather and wool

out of wear that looks sharp and holds up by partnering directly with ethical factories and top

artisans quince cuts out the middle man still a two deliver premium quality at half the cost

of similar brands they've really become a go-to across the board you guys know how I love linen and

how I've talked about it on previous episodes I picked up some linen pants and they feel incredible

quality is definitely noticeable compared to other brands layer up this fall with pieces that

feel as good as they look go to quince.com slash DSH for free shipping on your order and 365

their returns they're also available in Canada too but we have developed like a very fine

understanding of how really machine learning works from the fundamentals and we've done that all

over the years so we have like hundreds of different use cases and yeah right now we're working for

many different industries on the airline business we're doing models from like optimization of

contingency fuel we're doing revenue management systems we're doing like GNII making developers more

more efficient we're doing also things in the manufacturing space in the automotive space wow

so these companies that have like massive amounts of data they can like a 1% for one of these

companies is like hundreds of millions of dollars savings each every year yeah 1% yeah 1% so it's

like a project we did for a major airline last year they it saved them over 120 million dollars

in fuel cost savings wow and that is something that yeah that project took like a year of development

but that is like the when you optimize the whole value chain like massive unlocks happen

and that is like I think it takes like a change of mindset in the leadership of a company from

the top to make sure that there's actually like budgets for R&D and for like what's going on in bigger

companies is they have so many opportunities to use AI however their data layer is so far behind

and I was selling my colleague that I met right here in the conference that the other day we started

a project for a massive retailer and the way that we got the access to the data set for it was a

forecasting project was a hard drive mail to our office so imagine if you want to get like the most

most valued of AI what can you do if your data layer implies that you'd have to ship a hard drive

rather than giving me cloud access and everything right so there's like I think we're leaving kind of

in a bubble in the sense that the what we see when we see that AI moves so fast everybody's doing

agents everybody's having the ROI whatever like the vast majorities of companies are like so far

behind of that that they actually need to invest years in in these unifying the data layer and make

everything ready for extractive value out of AI so I think that's the most the massive part of

market is over there that's that's an interesting problem right yeah for sure for sure what would

your advice be to younger people that are just getting into AI starting companies and getting

into this business world so a question um I think right now it's very easy to be to be overwhelmed

by everything going on it's very easy to have imposter syndrome in the sense that yeah you're

you read the news you read read it you read like x everybody's doing cool stuff everybody's

getting things shipped everybody's like making a lot of money and it's again it's really easy to

to say okay am I wrong or something am I too slow and I think my answer would be just like

my advice would be try to do things to create a mental model of what the limits of the technology are

words best to use it word won't actually work at at scale there's many it's very easy to do

like flashy demos but then when you go to the real world there's a lot of complexities right so yeah

my advice would be don't don't think you are the imposter just like go play with the tools get

things done um and and yeah build that mental model of how the future might look like and enjoy

the ride like it's a it's a fantastic time to be working in this space great advice I struggle with

imposter syndrome grown I think a lot of people in my generation yeah it's very common yeah because

you're comparing yourself all day when you've grown on social media correct correct yeah the

grass is always greener on the other side that's like what people don't post their bad moments

exactly exactly I mean some do but it's not like yeah I use a i for these 15 different use cases

and it actually didn't work for 13 of those right like who actually post that yeah and that is

like the reality of everything that's going on behind the scenes you start understanding yeah my

demo will be very nicely but then when I want to scale this I will hit roadblocks yeah and I think

in like going back to the enterprise and everybody's promising agents and there's certainly like

use cases where agents are really useful especially if they're like kind of narrow

but I think the expectations of some of people on the leadership are like agents are the

magic silver bullet that will solve all your problems and again an agent that is 99% reliable

on a business might be completely useless imagine a self driving car that drives perfectly

nicely without crashes 99% of the time it's it's a one out of a hundred will crash it's useless

so getting to the 99.999999% that you need for the agent to be really self-sufficient in a generic

scenario it's really really hard and for many verticals like the technology is not yet there

it might get there eventually but there's no clear path to that however there's maybe like if I

used Chad GPD for brainstorming or whatever I don't care if it's just like hallucinates

some answers to me I will I will know I will read them yeah these are useless these are useful

and I will take that and do my things but for other use cases that really add business value to

the enterprise you need like and the level of real reliability that's not yet there well said so

that being said you think a lot of these companies are gonna fail do you feel like we're in an AI

bubble right now I don't I don't think we are like in a bubble as we were before I think we are

in a period of realization so they're going to realize that the AI as a silver bullet promise the

magical Chad GPD moment just not really translate to I plug in these models and then I don't need to

hire more people then I like completely rebound my business processes like the real world like

any super messy and it doesn't run like that but and I think the industries will take time to

adapt yeah you still need that human touch yeah absolutely absolutely probably for years to come

I mean it's hard to predict but it's it's very hard to predict I would bet that for the foreseeable

future I there's a lot of barrying timelines for whatever people call AGI yeah I think that from

the next at least like 10 years oh 10 years yeah 10 or maybe maybe even more like I think people are

the this this notion about exponential progress might not be such when you're trying to

apply all these to a very complex system with a lot of moving parts right so even if you had

these magical models they can do everything the all the industries being disrupted might take like

a longer time yeah they still implement it yeah yeah that's a good point wow and it's been

awesome I work at people find what you're doing and percent work with you yeah perfect so you

can follow us on LinkedIn try you'll have CRY we also have a website tryout.com we have a

technical blog we post a lot of interesting content around like these projects that I mentioned

and others like for technical folks there's like how do we do these things for business people

is how they can help your business so yeah I'd be glad if people check it out and thanks so

much for having absolutely man check them out guys we'll link it below see you next time

I hope you guys are enjoying the show please don't forget to like and subscribe it hopes

to show a lot with the algorithm thank you