0:04 in the last video we looked at writing
0:06 reading and chatting as three major
0:08 categories of tasks that you can go to
0:10 all them to do given that large language
0:12 models were trained to repeatedly
0:13 predict the next word maybe it's no
0:15 surprise that they're pretty good at
0:18 writing at generating words and it turns
0:20 out that many writing tasks can be done
0:23 via web user interface so I hope you
0:25 find this video where will dive more
0:28 into writing tasks immediately useful
0:30 for writing task broadly what we to do
0:32 is start with a prompt and use a
0:34 relatively short prompt to write so to
0:37 generate a much longer piece of text so
0:40 let's take a look at some writing
0:42 applications I often use the web
0:44 interface of large language models as a
0:46 brainstorming partner if you ask it
0:48 brain some five creative names for
0:50 peanut butter cookies it actually comes
0:54 up with some pretty creative names noty
0:57 Nirvana Nevels I would eat that or if
0:59 you ask it to brainstorm ideas for
1:02 increasing sales then it comes up with a
1:05 few ideas and you can take a look to see
1:07 if any of these may be useful you can
1:10 also use a large language model again
1:13 maybe the web interface version to write
1:16 some copy for you let's start to an
1:19 example if you were to ask it to write a
1:21 press release announcing the hire of a
1:23 new coo a new Chief Operating Officer
1:25 for your company it may come up with a
1:28 piece of text like this company name
1:31 welcomes a CO's full name as so on and
1:34 so forth and this is a pretty generic
1:37 press release when writing a prompt you
1:38 find that if you can give the large
1:41 language model more context or more
1:43 background information then it will
1:47 write more specific and better copy for
1:49 you if all that the large language model
1:52 sees is this write a press release at
1:53 this point in time it doesn't know
1:55 anything about your company about the
1:58 new C's name or their qualifications and
2:00 so ends up writing some very generic
2:03 like this if you end up prompting a l
2:06 model like this is not a problem you may
2:08 realize that you wound up with a very
2:12 generic press release and decide to
2:14 update the prompt to give it more
2:16 information and serve you at a prompted
2:18 and say use the following information
2:21 for the press release this is a CO bio
2:24 this is the name of our company and some
2:26 details about our company then it would
2:28 write a much more detailed and
2:32 insightful press release specific to the
2:36 Coos joining this company I find that
2:38 when prompting an LM I'll often not get
2:40 the prompt right the first time like
2:42 what we saw just now where we had the
2:43 prompt R press release announcing the
2:46 new hire of coo without giving any
2:49 context and that's totally fine if you
2:51 see the result isn't what you want just
2:54 revise the prompt and try again I'll say
2:56 more about this in the later video this
2:58 week when we talk about tips for writing
3:00 effective prompts let's look at one more
3:02 example another writing task that I
3:06 sometimes use l for is translation in
3:07 fact some of the lar language models you
3:11 can access VI web UI are competitive and
3:13 sometimes even better than the dedicated
3:16 machine translation engines already
3:18 especially for languages with a lot of
3:20 text on the internet and so where the
3:23 large language model had a lot of data
3:25 to learn how to generate text in that
3:27 particular language it tends to do less
3:29 well in languages also called Low
3:32 resource languages with less text on the
3:34 internet in that language but if you're
3:36 operating a hotel and you want to
3:39 translate the welcome message into
3:42 formal Hindi to welcome guest then a
3:44 large language model may be able to
3:46 Output text like this for you um
3:48 unfortunately I don't speak Hindi I wish
3:50 I did but it turns out that this
3:52 particular translation is only so so the
3:55 word front desk it translates into the
3:58 desk at the front rather than you know
4:00 the reception which is what we mean when
4:03 we say the front Des of a hotel so if
4:06 you're working with a Hindi speaker and
4:08 I was when preparing the slide then they
4:10 able to give you some tips to say oh
4:13 this is some sort of not quite the best
4:16 formal Hindi but they were to tell it to
4:19 translate this into formal spoken Hindi
4:22 then it Updates this text to make front
4:25 desk translate into the Hindi word for
4:27 reception which is a much better
4:29 translation now here's one fun thing
4:32 I've seen recently in the AI Community
4:34 which is a lot of us that are working
4:37 with translation often need to translate
4:39 text into languages that we don't speak
4:42 ourselves so how can we tell if the
4:44 large language model is doing something
4:47 reasonable and in fact even if you have
4:49 say one Hindi speaker on your team if
4:50 other members of the team don't speak
4:52 Hindi how can they figure out what's
4:55 going on so what I'm seeing multiple
4:57 teams in AI Community do is translate
5:00 text into pirate English
5:02 for testing purposes and so if you were
5:05 to prompt Lish model to translate this
5:09 into pirate English you get Oho mate we
5:11 be hoping you relish your time B the
5:13 ocean view in that sounds a pretty good
5:17 pirate English to me so that howby grand
5:20 worthy models be used for writing let's
