Building generative AI software applications is an iterative, empirical, and experimental process that involves rapid prototyping, internal and external evaluation, and continuous improvement through techniques like prompt engineering, retrieval-augmented generation (RAG), and fine-tuning.
Mind Map
Klik untuk perbesar
Klik untuk menjelajahi mind map interaktif lengkap
I'd like to share with you what the
process of building a gent of AI
software application feels like let's
take a look here's what the life cycle
of a gent VI project to build a software
application feels like we would start
off by scoping a project to decide what
do we want this software to do so for
example say you decide you want to build
a restaurant reputation monitoring
system The Next Step would be to
actually try to implement it and given
the ease of building AI applications
using gen of AI which you may have seen
in the optional video that came before
this one very often you be able to build
a prototype quite quickly and then plan
to over time improve this software
prototype for some applications I've
worked on we would build the initial
prototype in one or two days and that
initial prototype frankly isn't that
good initially but building it quickly
lets us then take it into internal
evaluation where we might have our own
internal team write different restaurant
reviews and test the system to see how
often it is giving a correct response
and sometimes the internal evaluation
will turn out some examples where it
doesn't give the right results in this
case with my pastor was code it outputs
this as a positive sentiment and you
know sometimes C Pastor is delicious but
this sounds like a negative sentiment to
me and based on problems that we
discover internally will then go back to
continue to improve the system as you
saw last week writing prompts is a
highly aerv process where you have to
try something see if it works and then
improve it and building a generative AI
software application also tends to be a
very iterative process after a
sufficient internal evaluation to give
you confidence that the systems working
well enough then we would deploy it out
in the wild and continue to monitor its
performance and it would not surprise me
if you deploy something and initially
external uses also generates input that
causes your system to make some mistakes
for example maybe a user right my Ramen
tastes like Ramen is this good or bad
well if you're not familiar with Ramen
or Japanese Cuisine you may not know is
this a good thing a bad thing and if
your system rates this as a positive
sentiment but it turns out that if
you're ordering miso Ramen on the menu
you probably don't want it to taste like
donot Ramen which tastes more like pork
based soup broth and when you find
incorrect responses like this other in
the while you might decide to go back to
internal evaluation for example to
systematically understand if your system
is say underperforming on certain types
of Cuisine or or you might decide to go
back to take these learnings to improve
the prompt or improve the system further
assuming you decide that these types of
errors are
unacceptable so it turns out that
building generative AI software is a
highly empirical and by that I mean
highly experimental process meaning that
we repeatedly try something and then
find and fix
mistakes we've already seen how
prompting itself is a highly empirical
process where you would have an idea try
the prompt see the El response then
maybe update your idea in the prompt and
go again but other than updating the
prompts there are other tools that we'll
talk about this week for improving the
performance of your genter AI system one
to that we talk about later this week is
rag or retrieval augmented generation
that gives the large language model
access external data sources we'll also
talk later this week about a technique
called fine-tuning that allows you to
adapt to Lar language model to your task
and then finally pre-training models
which refers to training a lar language
model from scratch don't worry about it
if you don't know what any of these
terms mean we'll go through each of them
in depth later this week but they're all
key techniques that in addition to
prompting gives you different ways to
improve the performance of your genf AI
systems performance just to walk through
a second example of the life cycle
regenerative project let's look at what
building a system to take food orders
might look like say you decide to scope
a food order customer service chatbot to
take orders what you would do is start
by building the system and quickly throw
together a chatbot to take food orders
then because we don't know how well this
is doing internally you might let your
internal team try it out and place
different orders and see how well it
does and sometimes they'll generate good
responses like do you have pickles on a
cheeseburger and also if you want some
and sometimes it will give an unexpected
poor response such as if you do have
mushrooms on your Burgers but for some
reason the chat bought says I'm sorry we
don't have mushrooms similar to what we
saw for the restaurant reputation
monitoring system it would be by
discovering mistakes like these that
helps you to improve the system and
after you're sufficient confidence that
this is safe to deploy externally you
can then deploy it and let customers
place real orders and monitor the large
language model's responses to make sure
that if it still says anything it isn't
quite supposed to that you can continue
to improve his performance having built
a number of gen of AI projects I've
often been surprised and delighted by
the strange and wonderful things that
the users will try to do with your
system for example if a user asks how
many calories are there in your burger
initially the system may not know but if
you discover this you can then update
the system using perhaps a technique
called rag that I mentioned just now and
it will go into depth later this week to
allow your software application to give
the correct answer so that's what
building a gent VI software application
feels like and if you work at a company
with a few or lots of software
Developers is and if you ever come up
with a cool idea for a G of AI
application that your company could
build this hopefully gives you a sense
of what that process of getting it built
might be like now one of the worries I
sometimes hear about is is it really
expensive to use these large language
models hosted by companies on the
internet it turns out that the use of
these large language models is probably
cheaper than many people think in the
next video I'd like to share with you
some intuitions about how expensive it
is or isn't to actually use these large
language models let's go on to the next video
Klik teks atau cap waktu mana pun untuk melompat ke momen tersebut dalam video
Bagikan:
Sebagian besar transkrip siap dalam waktu kurang dari 5 detik
Salin Satu Klik125+ BahasaCari KontenLoncat ke Cap Waktu
Tempel URL YouTube
Masukkan link video YouTube apa saja untuk mendapatkan transkrip lengkap
Formulir Ekstraksi Transkrip
Sebagian besar transkrip siap dalam waktu kurang dari 5 detik
Pasang Ekstensi Chrome Kami
Dapatkan transkrip seketika tanpa meninggalkan YouTube. Pasang ekstensi Chrome kami untuk akses satu klik ke transkrip video apa pun langsung di halaman tontonan.
