0:02 I'd like to share you some tips for
0:05 prompting launch language models if
0:07 you're using the web user interface of a
0:09 LM provider hopefully these tips will be
0:12 useful to you right away and it turns
0:14 out that similar tips are also useful
0:15 for if you're ever involved in building
0:18 a software application that uses LMS
0:21 let's dive in in this video we'll go
0:23 through three main tips for prompting
0:26 first is be detailed and specific second
0:28 is guide the model to think through his
0:30 answer and third is EXP experiment and
0:33 iterate this starts with be detailed and
0:36 specific using the fresh college grad
0:39 analogy I would often think about how to
0:41 make sure the OM has sufficient context
0:44 or sufficient background information to
0:47 complete the task so for example if you
0:49 were to ask it help me write an email
0:51 asking to be assigned to the legal
0:53 documents project well given only a
0:55 prompt like this and on doesn't really
0:58 know how to write a compelling case to
1:00 advocate for you to be assigned to that
1:01 project but if you give it additional
1:04 context such as I'm apply for a job in
1:05 the legal documents project we check
1:07 legal documents Vamp for experience on
1:09 prompting Els to get after text a
1:11 professional tone then this gives the LM
1:13 more relevant context to write that
1:16 email to help you ask to be assigned to
1:19 the project further if you can also
1:22 describe the desired task in detail so
1:24 if you tell it instead of saying help me
1:25 write an email if you ask it write a
1:27 paragraph of text explaining why my
1:28 background makes me a strong candidate
1:31 on this project an Africa from candidacy
1:33 then this type of prompts would not only
1:36 give the own sufficient context but also
1:38 tell it quite clearly what you wanted to
1:40 do and this is more likely to get you
1:43 the result that you want second tip is
1:46 to guide the model to think through his
1:49 answer so if you were to tell it
1:51 brainstorm five names for a new cat toy
1:55 it actually could do pretty well but if
1:57 say you have in mind you want a rhyming
2:00 cat toy name with a relevant emoji this
2:02 is what I might try I might tell it
2:05 brain Stone five names and tell it step
2:07 one come with five joyful words ready to
2:09 cats that for each work come with a
2:11 rhyming name and finally for each toy
2:14 name at a fun relevant emoji and with a
2:16 prompt like this you might get result
2:19 like this where the LM follows your
2:21 instructions to first come with pearl
2:23 whisker and so on and then Pearl 12
2:26 whisk whisper feline beine with fun
2:29 emojis add it to the end so if you
2:31 already have in mind a process by which
2:33 you think the own could get to the
2:35 answer that you want giving a clear
2:37 step-by-step instructions to follow
2:39 could be quite effective finally there
2:41 have been a bunch of articles that I've
2:43 seen on social media that say things
2:45 like 20 problems that everyone must know
2:48 or 17 problems that will help you grow
2:50 your career I don't think there a
2:52 perfect prompt for everyone instead I
2:54 find it more useful to have a process by
2:57 which you can write the prompt that
3:00 generate the result for you so so when
3:03 I'm prompting on myself I will often
3:06 experiment and iterate and try something
3:08 like my S off say help me rewrite this
3:11 and if I don't like the result I might
3:12 clarify and I say correct any
3:15 grammatical and spelling errors in this
3:17 and if it still doesn't give me exactly
3:19 result I want I might clarify even
3:21 further to say correcly grammatic spring
3:23 errors and rewrite in the tone
3:25 appropriate for a professional resume so
3:28 very frequently the process of prompting
3:30 is not about starting off for the right
3:32 prompts is about starting off with
3:34 something and then seeing if the results
3:37 are satisfactory and knowing how to
3:38 adjust the prompt to get it closer to
3:41 the answer that you want I think of the
3:43 process of prompting as like this you
3:45 start up of an idea of what you want the
3:48 LM to do and you just Express that in a
3:51 prompt and then based on the prompt the
3:53 LM will give a response and it may may
3:56 not be what you want if it is then great
3:58 you're done but if it isn't satisfactory
4:00 then that initial response wants helps
4:02 you shape your idea and modify the
4:04 prompt and iterate maybe a few times
4:06 before you get to the results that you
4:09 want so I think of the prompting process
4:10 as when I start off I try to be
4:12 reasonably clear and specific but to
4:14 save time I'll often start off with a
4:16 short prompt that may be as frankly less
4:18 specific than this desire but I just
4:19 want to get going quickly after you get
4:21 a result back if it's not what you want
4:23 then think about why the result isn't
4:26 the desired outputs and based on that
4:28 refine your prompt to clarify your
4:31 instructions and keep on repeating until
4:33 hopefully you get the El response that
4:36 you want one tip I want to share is I've
4:38 seen some people overthink the initial
4:40 prompt I think it's better to usually
4:42 just try something quickly and if it
4:43 doesn't give you the result you want
4:45 it's fine go ahead and approve it over
4:48 time you will not break the Internet by
4:50 just accidentally having a slightly
4:52 incorrectly worded prompt so go ahead
4:54 and try what you want two important
4:57 cards first if you're in possession of
4:59 Highly confidential information I would
5:02 make sure I understand how a large
5:04 language model provider does or does not
5:06 use or keep that information
5:08 confidential before copy pasting highly
5:09 confidential information into the web
5:13 user interface of an LM and second as we
5:15 saw in the last video with the lawyer
5:17 they got into trouble submitting court
5:21 filings with facts made up by ANM before
5:24 you count to L's result it may be worth
5:25 double-checking and deciding for
5:27 yourself whether or not you can trust
5:30 and act on tlm's output but with these
5:33 two caveat when prompting I will often
5:34 just jump in and try something and see
5:37 it not work but then use the initial
5:39 result to decide how to refine the
5:41 prompt to get a better result and that's
5:43 why we say prompting is a highly
5:45 iterative process sometimes you have to
5:46 try a few things before you get the
5:49 result you want so that's it for tips on
5:51 prompting I hope that you go to some of
5:54 the web user interfaces of the large
5:55 language water providers and try out
5:57 some of these ideas yourself and in this
6:00 course we provide some links some of the
6:02 popular om providers um and I hope you
6:04 go play with them and have fun with them
6:06 that brings us to the end of the main
6:09 set of videos for this week um there's
6:11 one optional video to follow where I'll
6:13 talk a little bit about image generation
6:15 or diffusion models so take a look at
6:17 that if you want and then look forward
6:19 to seeing you back next week where we'll
6:22 talk more about how to build projects
6:24 using lunch language models look forward
