0:04 welcome back last week we discussed how
0:07 generative AI can be used either via web
0:10 user interface or be built into a
0:13 software application in this week we'll
0:15 take a look at how many amazing software
0:18 applications are being built using G of
0:20 AI and we'll also take a look at some
0:23 technology options that go beyond just
0:25 prompting and that allow you to do much
0:28 more with gent of AI for example having
0:31 it operate on your own propri documents
0:33 rather than just on what is learned from
0:35 public sources on the internet let's
0:38 take a look we saw last week a few
0:40 examples of generative AI applications
0:43 such as writing answers to questions
0:46 that may require access to information
0:49 about your company's parking policy in
0:51 this example or reading restaurant
0:53 reviews on the internet to help with
0:56 reputation monitoring or building a
0:59 chatbot to help take food orders it
1:01 turns out that that while some
1:03 applications like this did exist and
1:06 were built before the rise of gen AI gen
1:09 AI has made building these applications
1:11 much easier and in many cases it's made
1:14 them work much better as well let me
1:17 illustrate with the example of reading
1:19 restaurant reviews for reputation
1:22 monitoring a few years ago if you wanted
1:25 to build a system for reading restaurant
1:28 reviews it would have taken writing a
1:30 lot of software code that looks like
1:32 this pages and pages of software that
1:35 you need machine learning Engineers to
1:38 write and specifically the process of
1:39 building a restaurant reputation
1:41 monitoring review system would have
1:43 looked like this you would use
1:44 supervised learning that's that
1:46 technology that maps from inputs a to
1:49 outputs B and if I were building this
1:53 system I would start by collecting maybe
1:55 a few hundred or a few thousand data
1:58 points with examples like this I would
2:00 have a review that suit Dum things ever
2:02 eaten that sounds delicious and label
2:04 that as a positive review the colorful
2:07 table cost made me smile that's positive
2:09 or not worth the three Monon wait that
2:12 be a negative review and the process of
2:14 building the system would involve first
2:17 getting label data then finding AI team
2:21 to help train an AI model on the data to
2:24 learn how to Output positive or negative
2:27 depending on different inputs a and then
2:29 finally you might have to find a Cloud Server
2:30 Server
2:33 like AWS or Google cloud or a zuro to
2:35 deploy and run the model so that when
2:38 you then input best bub I've ever had
2:40 that would hopefully recognize this as
2:42 having a positive sentiment and this
2:45 process would often take months in
2:48 contrast if you were to use prompt based
2:50 development this is the code you would
2:53 need to develop a sentiment classifier
2:56 first here's how I we specify a prompt
2:58 in code my prompt which have set equal
3:00 to 2 positive text there's the
3:02 instruction text classify the following
3:03 reviewers having the positive negative
3:05 sentiment and then here is the review
3:08 text and after specifying The Prompt in
3:11 codes I just need one line of code to
3:13 call the large language model to get a
3:16 response back and then I'm going to have
3:20 it display or print the response so this
3:23 is pretty much all the code it takes to
3:26 build such a system and in fact in the
3:29 next video I'll share you an optional
3:31 exercise where you can try out this code
3:34 yourself whereas with the traditional
3:36 approach to building a sentiment
3:39 classifier using supervised learning the
3:41 timeline for the project might have been
3:45 a month to get say a th labeled examples
3:46 with a th000 reviews and positive
3:49 negative labels after collecting the
3:51 data it might have taken a team say
3:52 three months to train the AI model on
3:55 data and then another three months to
3:57 deploy it and make sure it's running
3:59 well and it's regular and robust I don't
4:01 know if this seems like a long time to
4:03 you but for many really good machine
4:06 learning teams I've worked with this 6
4:08 to 12 month timeline was pretty
4:10 realistic for what it took to build and
4:13 deploy a valuable AI model and this
4:15 worked and this was very valuable for a
4:17 lot of applications but this just took a
4:20 long time in contrast for prompt based
4:22 AI this is what it feels like you can
4:24 specify a prompt in minutes or maybe
4:28 hours and then deploy the model in hours
4:31 or maybe days so there are now many
4:33 applications that have previously taken
4:36 me and very good machine learning teams
4:38 maybe six to 12 months to build that
4:40 today I think there are millions of
4:42 people around the world that can now
4:45 build in maybe days or a week and this
4:47 is fantastic because this lowering of
4:49 the barrier to entry to building such
4:51 applications is leading to flourishing
4:54 of a lot more AI applications with one
4:57 important caveat which is that as we
4:59 discussed last week gen of AI tends to
5:02 work much better for unstructured data
5:04 like text and images and audio um but
5:07 with that admittedly important caveat
5:09 the number of AI applications built on
5:12 top of gender AI is just letting the
5:14 community do much more than ever before
5:17 in the next optional video I'd like to
5:20 invite you to try out some codes with me
5:22 for reading restaurant reviews and
5:24 classifying sentiments it's fine if
5:26 you've never seen or written a line of
5:27 code before in your life but I'm hoping
5:30 to convey to you how code is needed to
5:32 do this now and let you try it out
5:34 yourself so I hope you take a look
5:36 though also feel free to skip it if you
5:38 wish and after that we'll come back and
5:41 talk about what building a genitive AI
5:43 software project feels like when we talk
