0:04 in the last video we looked at writing
0:06 tasks where you would specify a prompt
0:09 to the LA L model and have it generate a
0:11 comparatively longer output than the
0:14 input prompt it turns out LMS also
0:17 useful for many reading task and by that
0:19 I mean task where you would input a
0:21 prompt and then have it generate usually
0:24 a similar length or often shorter output
0:26 than the input prompt let's take a look
0:28 at some reading tasks starting with
0:30 something I use myself all the time
0:33 which is proof reading many times if I'm
0:35 writing a piece of text I will read
0:37 through it carefully three or four times
0:39 myself for spelling and grammatical
0:42 errors and even though I thought I Pro
0:44 read carefully myself a l language model
0:46 we'll still find errors in it that
0:49 somehow I had missed here's an example
0:52 of a prompts that you could try proof
0:54 read the following text and I find that
0:57 if you tell it what you want the text
0:58 before so here's text intended for
1:01 website selling TR of toys and sometimes
1:02 ask it to check for spelling and
1:04 grammatical errors as well as awkward
1:07 sentences and then have it rewrited with
1:09 corrections and this is a piece of text
1:12 with some errors and the output of the
1:15 launch language model fixes snuggle was
1:17 misspelled and it fixes this little
1:19 piece of grammar over here when I'm
1:21 writing text myself that I want to be
1:24 quite confident is free of spelling and
1:26 grammatical errors and sometimes also
1:29 awkward sentences I actually use this
1:32 myself to proof read what I write a
1:34 second reading task that large languish
1:36 models are often used for is to
1:38 summarize long articles one of my
1:41 collaborators Eric Brinson who's a sord
1:43 professor once sent me an email linking
1:46 to an article that he had written titled
1:48 the tour ring trap and I knew it was a
1:50 good article but it was a very long
1:52 article and I didn't have time to read
1:54 the whole thing before I responded to
1:57 his email so I actually used the
1:59 following prompt and copy pasted his
2:02 entire article into an web interface of
2:05 a large langage model and had it quickly
2:07 generate a summary for me it turns out
2:09 this paper that he had written talks
2:12 about how human like AI offers benefits
2:14 but there's a lot to be done by having
2:17 AI augment humans rather than automate
2:20 but the point of brov's article on the
2:23 touring trap was he was advocating that
2:26 instead of having AI automate or replace
2:28 human work we should put more effort
2:31 into having AI complement augment human
2:33 work and so with a lar action model
2:35 summarizing this long article I was able
2:38 to get back on this faster than if I had
2:41 to read the entire article myself and by
2:42 the way this is a good article
2:44 eventually I did read the entire article
2:46 myself and really enjoyed it but today I
2:48 do sometimes use large language models
2:51 to summarize for me things that I don't
2:53 have time to read in this entirety and
2:55 so this is a use case that you could go
2:57 to one of the web interfaces of a lar
2:58 language model and use relatively
3:00 quickly yourself
3:03 now it turns out there's a software
3:05 application version of this too that is
3:08 taking off in businesses let me
3:10 illustrate this with an example say
3:13 you're a manager of a customer service
3:16 call center where you have many customer
3:18 service agents like this person shown on
3:20 the left with the microphone having
3:22 phone calls with customers like this
3:24 person shown on the right if you have
3:26 permission to record these phone calls
3:28 between the agents and the customers you
3:31 can then run the phone calls through a
3:33 speech recognition system to get a text
3:35 transcript of the conversation and if
3:36 you have many customer service agents
3:38 having conversations you end up with a
3:41 lot of text transcripts if you want to
3:42 review what's going on in your call
3:44 center you priority end up with too much
3:47 text to read given a text transcript
3:48 like this between a customer and an
3:50 agent you what really happened in this
3:53 call one use of lar language models
3:57 would be to have it summarize this
3:59 entire conversation and generate a short
4:03 summary like mk41 127 KX was reward is
4:06 broken and so on and if you were to take
4:08 all of these different text transcripts
4:10 and have a software application to
4:13 generate these summaries then you as the
4:16 manager of this can take a quick look at
4:18 all of these summaries and just maybe
4:19 spots if there are any issues or any
4:22 trends that you want to be aware of a
4:24 system like this would be implemented as
4:26 a software application that uses a large
4:28 language model because it doesn't really
4:30 make sense for you or anyone else to
4:33 copy paste these conversations one at a
4:36 time into the website of a large
4:38 language model provider in terms of
4:41 customer service interactions L langage
4:44 models are also used for customer email
4:47 analysis in an earlier video you saw the
4:50 example of taking a customer email and
4:52 deciding if is a complaint um in this
4:55 case no as well as what department to
4:57 Route this email and this will be
4:59 another software application that uses a
5:01 l language model let's take a deeper
5:03 look at how one could build this
5:05 application focusing on the parts of
5:07 deciding what department to Route this
5:10 email one thing you could do is writeit
5:12 prompt to tell the L to read the email
5:15 and decide which department to Route it
5:16 to so you can specify the task and
5:18 provide the email but it turns out that
5:20 with a prompt like this you may find
5:23 that the algorithm routes it to the
5:25 complaints department in this case which
5:27 may or may not be a department that
5:30 exists in your organization
5:32 so this would be an example of where the
5:35 L has been given insufficient context to
5:37 know what are the names of the actual
5:39 departments in your company that they
5:42 should choose from in contrast if you
5:44 were to update the promise follows we
5:47 say read the email choose the most
5:48 appropriate Department to Rouse it to
5:50 and choose Department only from the
5:53 following list in this case given the
5:56 set of choices you wanted to choose from
5:58 routs it to the apparel Department
6:01 correctly it process of building an
6:03 application using the launch n model is
6:05 again not at all in commmon to write a
6:07 prompt that doesn't quite work right the
6:09 first time and when you find it routs it
6:12 to a non-existent complaint department
6:13 then just update the prompt and that
6:15 fixes the problem one last application I
6:17 want to touch on is reputation
6:20 monitoring where you can use an LM to
6:22 build a dashboard to track your customer
6:25 sentiment positive or negative of your
6:28 business or your products over time so
6:31 for example if you run a restaurant and
6:33 occasionally your customers write online
6:35 reviews or send you emails describing
6:37 their experience you can then use a
6:39 prompt like this read the following
6:40 review and classify as having the
6:42 positive negative sentiment to have it
6:45 decide automatically if each review was
6:48 positive or negative so in this case if
6:49 the food was amazing or service are
6:51 friendly that would be classified as
6:53 having a positive sentiment then by
6:56 having software counts the number of
6:58 positive reviews per day as well as the
7:00 number of negative negative reviews per
7:02 day you can build a dashboard that
7:04 tracks per day over time how the
7:07 sentiments are trending and looks like
7:10 the customer senent is pretty positive
7:12 but if ever it starts trending negative
7:15 like this with more negative reviews
7:17 then this dashboard can alert you to
7:19 that maybe something's happening that we
7:21 should pay attention to and see if
7:22 there's something we need to fix at the
7:25 restaurant so in this video we looked at
7:27 a number of reading applications
7:29 including proof reading summarization
7:31 email routing restaurant review sens
7:34 analysis if you can think of task where
7:36 you wish you had someone that could read
7:38 a piece of text and just say a few
7:40 things or give a few quick indications
7:42 of what was in that piece of text that
7:45 could be a good candidate for a reading
7:48 toss to get home to do for you next
7:50 let's go to the next video to take a
