YT YouTube2Text

YouTube2Text → Transcripts → Session 2: Potential Applications of AI in Various Purposes

Session 2: Potential Applications of AI in Various Purposes — Transcript

by Development Academy of the Philippines · 4,666 words · 723 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:21So in the morning I just I upon touch on
0:23this uh prompt engineering but because
0:26now you have a new uh participants so
0:29that's I want to also emphasize
0:32uh the importance of this prompt
0:34engineering so this is a structure so
0:37prompt engineering so This is the
0:40beautiness of uh improving the
0:43productivity of our outcome after we use
0:48the AI. So there is a uh structure. So
0:53we needed a details
0:56we need we need to create a state and
0:58also we need a structure and also if you
1:01have a background information
1:03uh you need to add as a huot it's called
1:07huh short additional learning additional
1:10information but for example um AI maybe
1:15like current ch is pre trend since I
1:20think December 2023 I
1:23So like last year information is um
1:27uncertain. So they don't know.
1:30So that's in that case for example like
1:33if
1:36we ask about the prime minister of like
1:41uh Japanese new prime minister maybe
1:42they might misunderstand or maybe uh
1:47current current model is also connected
1:50with the internet. So that's why I think
1:52should be okay but I think old model uh
1:55there there's a cutting off cutting
1:57cutting off point like in terms of the
2:00learning experiences learning time so
2:04that in that case we need to educate
2:06them so it's called a few short so I
2:09think you have your own data set right
2:11so in in that case you can you know add
2:15your own information or like a document
2:18or like your publication
2:21So you can educate them
2:24to improve the cor the outcome.
2:29So there are many models but the
2:32prompting uh types but I think I just
2:35focus on four.
2:37There is a four uh ways to communicate
2:42with AI. It's a just a type of prompting
2:47not so difficult just few short few
2:50short means just we need to uh provide
2:55additional information
2:57we need up to date we need to educate AI
3:02it's called fshot prompt zerosot
3:06just on preed
3:09information pre
3:11so if you are talking about maybe
3:14history or like past
3:17or maybe like a literature or if you are
3:20a researcher or you you if you want to
3:22have some publication uh like you need
3:25to publish
3:26uh like a books or academics academic
3:30research paper uh you if you need to uh
3:33summarize
3:35maybe like you can use zerot
3:39from but of course like if you have
3:42additional information you can add some
3:45information. So this is the new uh this
3:49is the different chain of salt. Chain of
3:52salt or sometimes like it's called like
3:55abbreviation is C O T chain of salt.
3:58This is very important chain of thought
4:01is like there is a like a l communic
4:04like communication. So we so imagine
4:07like you have a system you have a friend
4:10you have a family. So you you have a
4:13communication right? If you say
4:15something big maybe your friend or your
4:20family do not understand for example if
4:23you say like oh today uh you pass I
4:27participate with some something like AI
4:31seminar but if you say oh something AI
4:35what what kind of seminar did you
4:37attend? So maybe like if you provide
4:40some vague information AI cannot
4:44understand. So then AI might ask oh like
4:48uh what you want is this or that.
4:54So maybe still AI is not uncertain but
4:57if you can uh clearly communicate or
5:02state what you want you can directly get
5:05the uh answer or what you want to get
5:10from AI that's why this is a
5:12communication skill it's called prompt
5:14engineering prompt engineering so let's
5:17let's learn it's not so difficult so
5:20chain of salt is there are many rarities
5:22many raries is so
5:25structured.
5:28So however this pain of salt is a part
5:32of
5:34CH GPT or model chat GPT or model is not
5:40generated preformer.
5:44So maybe not I just explain.
5:48So open AI has two model
5:52GPT is pre generated train transformer
5:57right do you got it but this is
6:01AI is not thinking at all AI is just
6:05generating
6:08the something
6:11which might be relevant to the previous
6:16work.
6:17So just they continue to generate
6:21generate generate generate generate then
6:25this kind of continuous generation
6:28will make a sentence or paragraph or
6:34chapter
6:36but
6:37we need to check whether it's true or
6:41not. So that's why it's called
6:43hallucination. So this is because if
6:46some somebody say oh AI might make some
6:49mistakes and also if you see the at the
6:52bottom of the chpt they say they
6:55strictly they state that GPT might make
6:58some mistakes but because GPT model is
7:03designed like that because just generate
7:06generate generate generate generate but
7:10it might make some mistakes. So this is
7:13the GPT model but it's okay.
7:17Then all model is this different all
7:20model all model it's called using chain
7:23of salt actually they uh it's called a
7:26deep sync deep sync not a deep sync deep
7:30deep sync so all model is just uh just
7:35uh my intuitive understanding is all
7:39model is there is a a lot of agent
7:42players a lot of agent just imagine you
7:46have 100 assistant
7:49100 assistant
7:51simultaneously start thinking
7:54about to to work for you to provide
7:59the outcome better outcome. So you have
8:02a uh option A B C D E you have a 100 100
8:08options then oh this all model is using
8:13the chain of salt and uh reinforcement
8:16training it's called but just I have no
8:18time to explain but just this model is
8:21just start thinking using chain of salt
8:25then reach to some outcome
8:29then option A option B option C maybe
8:34three option might be good for your
8:37request.
8:39Then they put a score. Option A 100
8:43point, option B 90 point, option C 60
8:48points. Okay, option A is a winner. Then
8:52they will reply to you together with
8:55option A. So they are like know compete
8:58like there's a competition.
9:01So all model is at the same times the
9:04many agents start thinking using chain
9:07of thought then they put the score
9:12uh recommend the better answers. So this
9:17is not thinking but looks like thinking
9:19right. So there is the option then they
9:22will find the better answer for you not
9:26not just generating generating
9:28generating because they are comp uh
9:31there's a competition
9:33so there's a put a score so they are
9:36comparing the option so that's why it
9:40looks like thinking like also like we we
9:43are also like kind of like human
9:45thinking right we have some option
9:48okay like for dinner or what kind of
9:51like dinner shall I eat to maximize my
9:55satisfaction utility or kare adobo or
9:59shiniga or pit or what can be increase
10:03my satisfaction so this is like in my
10:06mind there is a many agent are fighting
10:08each other right then I might find kare
10:12to maximize my yeah satisfaction so this
10:15is looks like my thinking right so this
10:18is all
10:19So
10:21the thing is GPT can be used for
10:26brainstorming.
10:29It's good for brainstorming because in
10:31the case of brainstorming it's okay any
10:34idea is welcome right welcome. So that's
10:37why you can use GPT model for
10:40brainstorming. So my my my ways also now
10:44now I'm also using such kind of for
10:46brainstorming or a lot of idea where I
10:49need so I use GPG model then I get them
10:54a lot of like item or information then I
10:58input this generated information into
11:03our model
11:05then I will like put some structure
11:10or also target
11:13then I have more better structured
11:19document.
11:21So this is what I'm writing a book
11:27that is recommended.
11:30So I think like from here so yeah easy
11:33easy to like uh yeah it's easy to follow
11:37actually. So you can change dramatically
11:40you can change the productivity.
11:42So I think maybe so my target is yeah I
11:46think like a one yeah one book per month
11:52my
11:53maybe I maybe next I say my one one book
11:56per year
11:59because the facteing is it takes time
12:01for facteing so I realize that so
12:03unfortunate so still we yeah human task
12:06is remained
12:08so anyway so anyh you can speed up the
12:11publication or like you know generating
12:14process. So you even you can improve the
12:17quality of your work. So in my case like
12:21I I need publication or like a text or
12:25educational materials.
12:27So the last zero short chain of thought
12:30is no no additional train right.
12:34So prompt this is a very important
12:36right. So in the in the morning session
12:38like I explained like there is a
12:40hashtag. So we can use the hashtag for
12:43to structure to create the structure. So
12:46like from the participants from the uh
12:49afternoon session so you can still catch
12:52up so you don't miss so just I will
12:56explain. So this is the p like my main
12:59message
13:01important. Yeah, I just so instruction
13:05to the AI are recommended to be placed
13:07at the beginning. But I think the role
13:09play low ro is also very important. Then
13:13can use a hash hash sign hash hash sign
13:17to uh create the section section. So we
13:21need a structure.
13:24So this is the recommended structure.
13:27But first like uh I think later later
13:29there's a coming there's a role
13:32we can uh allocate the role
13:36we need to specify what kind of role AI
13:41will play but then instruction the
13:44background information data if you if
13:47any then output format.
13:52So just a simply this is just example
13:55just a simp simple one simple one but I
13:58think more more detail prompt is coming
14:01later. So academic research assistant
14:04I'm writing a paper blah blah blah then
14:07requirement or uh
14:11uh constraints
14:14so purpose content detail. So this is
14:17one example. So we need a section break.
14:20Section break. So this is the
14:23instruction. Then section break. Then
14:27requirement or constraint or conditions.
14:33The job interview preparation.
14:35This is also very good like like for job
14:38interview preparation. Uh maybe there
14:42might be a mock training but I think we
14:45can you know we can communicate with our
14:48assistant to train yeah by myself. So I
14:52have some job interview or maybe for
14:55promotion maybe we can have a uh we can
14:59prepare for interview practice. So then
15:02I have a job interview blah blah blah
15:05then there's there's a condition
15:08that travel itinary for example if you
15:10are planning to go to Tokyo maybe there
15:13are some condition so you you don't like
15:15the meat of your vegetarian or I don't
15:18know but then also like budget
15:20constraint or time constraint you can
15:23put create because the AI is connected
15:26generative AI is connected with now the
15:29internet so you can also get uh uh
15:32accurate information.
15:35So this is uh I think that many people
15:38might be using just simply
15:42uh asking the task
15:46uh ask the task
15:49uh to your assistant AI. But just also
15:54like we need a Lord play
15:58uh to hit the target. we need to uh
16:02specify the role. So in that case uh we
16:06easy to hit the target.
16:08So because I think my my understanding
16:11is generative AI is a generous general
16:13intelligence. So they can answer they
16:16try to answer
16:20uh for your questions but maybe their
16:24reply might not be put on not hit hit
16:28your target. So, so just we we we we
16:32want to like you know increase the
16:34productivity right we don't want to
16:35waste our time yeah chain of thought is
16:38sometimes good but it there's a many
16:40rally so it does not I think we can save
16:43our time so in that case we need to ask
16:46the we need to ask to play a certain
16:49role okay translator data analysis or
16:54like financial uh uh advisor
17:00Yeah, for my my personal advice, but
17:03just unfortunately
17:05I'm a maybe typical Japanese
17:09uh
17:10husband. I have no freedom about my
17:14financial decision making.
17:19Everything is controlled by my wife.
17:23Is it is it the same here?
17:27Because I think because I when I check
17:31in Yeah. in Yeah. We are staying we are
17:34staying near here. So when I check in
17:37the hotel so automatically the message
17:40will deliver to my wife's
17:43email. So then like oh and then he she
17:46said oh Akira you checked in
17:51no no freedom. So but
17:56so it's my typ I'm typical Japanese. So
18:02the role play is very important for AI.
18:06we need to provide a certain load so we
18:09can hit the target
18:12and also maybe you don't know like u
18:15some there is a Japanese uh investors
18:18Masayoshi song the founder of soft bank
18:22or bjon fund it's he's very crazy about
18:27investing uh for AI then also I'm
18:31attending the soft bank uh I don't know
18:34like a conference and He is he is also
18:38he he was enthusiastic about yeah chach3
18:43or model and also
18:45um he also mentioned about he also
18:48shared how he
18:51uh doing the decision making he said
18:54like he is using role play every for any
18:59decision making he create he asked the
19:02chat GPT as a board member
19:05as if as a as if they were board member.
19:09So he he asked some advices. So so
19:14that's why he always like encourage
19:18also like us Japanese to uh they he
19:24already encourage us to use uh
19:28generative AI for better decision
19:30making. So this is also one example.
19:38So in that case definitely it's a game
19:41changer. Definitely it's a game changer.
19:45So in that case prompt the engineer
19:50it's a new maybe title or new skill or
19:56job
19:58and also
20:00GPT model might make some mistakes. So
20:04in that case we need to allocate someone
20:09for factchecking.
20:11So this is also a new law
20:16maybe in your authority or in your
20:18institution right?
20:21If we we want to encourage
20:25our staff or your colleague
20:30uh use AI
20:32maybe we need some break as well. So not
20:36not not not to dis not to disturb I mean
20:40just factchecking is for safe for safety
20:44right for safety driving.
20:48So factchecking there are many ways of
20:51course like we can check our outcome
20:55uh we can check just like internet or
20:58like a real world uh information but
21:02also there's a way we use rely on the
21:06different AI model generative AI if you
21:10use if you are worried about the outcome
21:14generated by chbpt I think you can use
21:18the same prompt uh with uh Google Gemini
21:24or Meta AI. So just just just in case if
21:28you worried about can check it for fact
21:31checking.
21:33It doesn't cost a lot.
21:36Then also there's some business but it's
21:39it's maybe outdated but there for the
21:41academia
21:44a year ago or one year ago there's a um
21:48just a
21:51a venture Japanese venture who who
21:53created some free software for you know
21:57like a Sherlock Holmes AI detective to
22:00check the students like a report
22:05but
22:06Yeah, we we we need I say many many
22:09institution actually installing this
22:11kind of thing. But anyway, so maybe we
22:15also need to change the the way of
22:18teaching or way of the assignment.
22:21If we just ask the student to write or
22:26write a report, maybe it's not enough.
22:30So we need to check oh exactly you wrote
22:34or not. Maybe we need some interview I
22:36think. So I think maybe such kind of uh
22:39but it it takes time right cost a lot if
22:42we interview all
22:46now we are going to hands-on practice
22:53I think first is because I'm a
22:55researcher and also like you you are
22:58public official official so just I I
23:00first I share some uh prompt
23:06for Proof reading
23:09just we can put the role here you are an
23:13advanced AI proof reading system or you
23:16are top-notch or excellent editor you
23:20can check also like for public
23:22publication also I use like you are
23:26excellent or well-known uh editor or
23:29publisher so we can use it so we need to
23:34uh edit Okay. Or like uh
23:39clearly state what kind of role AI need
23:43to do. So in that case it's easy to hit
23:46the target. All right.
23:49Then also like sometimes there we can
23:51put the guideline. So hashtag. So there
23:55is a low and also instruction please
23:59follow this guideline. Then break point
24:04like section break. Then hashtag we can
24:07put additional uh section next section.
24:12Then we can put the actual text at the
24:16end.
24:19Of course like you can upload the text
24:22file or like a word file or Excel file.
24:27But I think also like you can just uh
24:30copy and paste
24:32directly
24:34paste the text here.
24:37I think nowadays I'm I'm doing that. I
24:39think I I do it works well even for 100
24:43pages. I I just press
24:46because my book is my font book is um
24:52300 pages.
24:55So, so I just uh put the one chapter by
25:01chapter by chapter one. Yeah.
25:05Then summarizing sentence. I think also
25:08like you have a lot of document to read
25:10and check and summarize or policy brief.
25:13So you can use this kind of like prompt.
25:17So you can put some condition right. So
25:20but the structure matters. So just I I'm
25:22sharing the structure maybe uh maybe
25:25some some of you uh have already known
25:28about that but just this is a just a uh
25:32like information sharing.
25:36So then you can also make your sentence
25:40longer and shorter. You can also do
25:43shorter or longer.
25:47Then if you want to create the PPT
25:51presentation slides based on some
25:53document. So in that case my student use
25:57this kind of prompt like a create a
25:59bullet point summarize using the point
26:04appropriate English or new.
26:08So that might be a save your time and
26:12also I I mentioned like also my now
26:14nowadays I'm not maybe 90% of my email
26:18are uh drafted by my assistant.
26:24So turning break point into sentence.
26:25This is the way around. If you have a
26:28presentation slides then based on your
26:31presentation slide you want to create
26:33some document. So in that case you can
26:36use this prompt to make bread point or
26:41powerpoints presentations
26:44based on PowerPoint presentation you can
26:47create the document
26:51then after getting the piece do you know
26:55Lego
26:57you know Lego so I'm also now nowadays
27:00I'm explaining uh maybe information is
27:04like a Lego for me. So I can generate
27:08some block using GPT model or maybe like
27:14I read some literature then I summarize
27:18I have a block
27:21then
27:23I have many information then but I need
27:26some
27:29structured or my product
27:32my product is sometimes
27:35research paper or academic paper
27:41what to do so my current
27:45way is I'm put all the block into all
27:51model
27:53to create a structure then O model as I
27:56mentioned O model is
28:00uh there is a competition right
28:03competition okay Akira
28:06This structure A might be okay but you
28:10can have a different scenario B but C
28:13might be also attractive then they are
28:16start competing then put the score
28:20then show oh I think after the chain of
28:26salt and reinforce reinforcement
28:29training so the latest uh O3 model
28:33so this might be a good structure or a
28:37good uh chapter.
28:42You might surprise you might be
28:43surprised.
28:47I mean I'm now you understand how much
28:50you I enjoy.
28:56So now so you you also have some you
28:59know repetive and also like you know uh
29:04some like uh ad work reply message or
29:14then sometime you need to write of
29:17course like for example like when I work
29:19for JA so I need to write like a uh
29:23opening speech for the executive right
29:27uh vice president or I don't know how to
29:30because I'm not polite enough polite to
29:33you know I don't know the polite
29:34Japanese language so so just maybe at
29:37that time because I I could not use AI
29:40because for a long time ago but now I I
29:44think just I have some you know my way
29:47to write the contents then I ask my you
29:52know assistant to change the format
29:55or to make you know executive feel
29:59satisfied.
30:01So because you know I can easily change
30:02the format. So for any if it's a
30:07international conference so my Japanese
30:10you know um casual conversation
30:15so based on that I can create a very
30:18sophisticated Japanese
30:21speech. then I can also create the
30:24English version. Right? So unfortunately
30:29at that time I have no so I think I have
30:31I have two days to to like think about
30:35like you know so but now maybe just
30:38within five minutes I can I can maybe
30:41make him satisfied. Yes. Feel satisfied.
30:45So you can time
30:48then also maybe many people are you know
30:52uh I need to reply email or chat so so
30:57you are also might be wondering like how
31:00I cannot I can reply so in that case you
31:03can also you know uh create a easy to
31:06read sentences and maybe sometimes like
31:09it's good like my student said like uh
31:13some university student like pretend
31:16to understand about their what they
31:19reported. So just I ask them oh what
31:23does that mean like on the presentation
31:25so could you explain then actually they
31:27don't they cannot explain so in that
31:29case I just I ask okay please take your
31:31time so you can ask your assistant okay
31:35please ask them about the term technical
31:38term you don't understand maybe like
31:41please explain
31:43uh in with with using the you know easy
31:47to inst manner or please explain to a
31:52high school or elementary school student
31:55about the technical terms. I think that
31:58that also works well. So sometimes you
32:00know
32:03then also maybe like you have many
32:06meeting right so you need assistant
32:10so it's a lovely like you know you can
32:12enjoy this no beautiful world now
32:16everybody everybody everybody can have a
32:20assistant now
32:22you can have a data assistant
32:25programming assistant translation
32:27assistant you
32:29like email creation
32:33or like you know Excel Excel seat like
32:36formatting
32:40we can do anything
32:42because I think when we u communicate
32:45using the CH GPT actually directly we
32:48can uh use u programming software like a
32:53Python. So Python code is also like
32:56learning uh behind but even you don't
33:00understand the Python code you can work
33:04you can use Python for your whatever for
33:08your purposes for your own purposes so
33:11that's very strong now
33:14translation English to Arabic English to
33:16Tagalog
33:20so
33:22but of course like you know Akira
33:24Oh, I know. I know. I know how to
33:26communicate,
33:28but
33:30Oh, what what do you think? Uh, you can
33:33you can follow my lecture or like my uh
33:38notes, but of course like it's not
33:41enough, right? It's not enough. You have
33:43your own tasks. So, in that case, we
33:46need to be flexible. In that case, I
33:49have a good prompt for you.
33:53How to create your own prompt original
33:58original prompt.
34:04So there is a way
34:07you can ask
34:09your assistant
34:11to create your own prompt.
34:17Did you get it? So I have assistant.
34:22I ask my assistant to create
34:27for himself for himself or for
34:30itself. So just I'm just asking oh you
34:35need to work for me but you need to
34:38think about your command or prompt by
34:42yourself.
34:45Lovely right? It's very lovely. I want
34:49you chach GPT. You can say chach.
34:52I want chach GPT to be my prompt
34:55engineer.
34:57Please help me. Help, help, helps create
35:01the best prompt for my needs.
35:05Very selfish, right?
35:08But this is a real real world. Yeah,
35:11this is a real real world. We are in the
35:13real world. The prompt is used by you.
35:19Please follow the process below. It's
35:21continue. It's continue.
35:25I'm very friendly, right? Number one,
35:29your first response is to ask me because
35:33I'm a I'm a command. I know I'm a
35:35conductor, right? So, ask me what the
35:38prompt is about.
35:42Like you think it's Yeah, you might
35:44think like I I I look like I'm I I I I'm
35:48seem to be stupid.
35:52Please please ask me what the prompt is
35:55about.
35:56I'll answer that question. Answer that
35:59question. But keep it iterating
36:04over the prompt through the next step.
36:07The next step, step two. Step one, step
36:09two. Based on my input, because I I I
36:13I'm a I'm a boss. I will reply based on
36:16my input, you generate the following two
36:20section. One device prompt
36:24divide by better to reach to the goal.
36:28Then questions
36:30just in case shat ask some question
36:34maybe to make me satisfied.
36:39So please provide the written prompt
36:42sorry for revision it is something that
36:45is real
36:47and is understood by you but you need to
36:50work. the question ask me any relevant
36:53additional information because I'm I'm
36:57I'm I'm stupid so I need I need maybe I
37:02might forget about that something
37:04something necessary I might so that's I
37:07think chach might think about oh my boss
37:09might forget about this information so I
37:12should work for that so he will chach
37:16will just in case remind me then step
37:20Three, continue this process until I say
37:26or I say okay.
37:30This works well.
37:32So this is also very important if you
37:38want
37:40become a book writer or if you want to
37:47uh generate a long long long sentences
37:51or a chapter
37:54because I will also explain
37:58with one prompt with one prompt
38:02I think on average you can get 1,00 or
38:071,200
38:10uh letters like letters words sorry the
38:14bars.
38:16So if you say
38:19please like
38:22for example I am writing a fantasy book
38:25right or economic economic economic
38:27textbook is okay. So I'm also actually
38:29writing I'm writing the economic
38:31textbooks.
38:32So if I say with one prompt chat GPT is
38:36on average reply
38:39uh some text I think I learn with around
38:431,000 words but with 1,000 words it's
38:47not enough to
38:50uh enough for consisting of one chapter
38:54right so what what what we need to do is
38:57we need to say please generate the text
39:01with stepby step.
39:04This is a
39:07I think a kind of secret or like a
39:12I think the very important
39:15prompt. Please generate
39:18blah blah blah with step by step or one
39:21by one.
39:23Then you can say like this.
39:27So I actually just I I changed this idea
39:32to make
39:34uh make such
39:37generating the long long long sentences.
39:40So then I can I can generate the 300page
39:44u book uh I think within half day. So,
39:50so then you can say because the CHP max
39:53output is 1,000 words. But then you can
39:57say one by one, one one section by
40:00section. Then you can continue, right?
40:04Then you can become a chapter. You can
40:06create a chapter. Then chapter become a
40:08uh you know book.
40:11Then you can say
40:14uh please generate the uh section or
40:18episode one by one with step by step.
40:21Then please ask me
40:24if you you produce the outcome. So then
40:28please ask me if whether your outcome is
40:32satisfactory or not. So I always use the
40:36same prompt. Please ask me
40:41if your outcome is satisfactory or not.
40:44Then ch generate the episode. Then ask
40:48me, oh are you satisfied? Do we need any
40:53other input or additional episode so we
40:57can improve? So this is a way to
41:00generate the very good quality of uh out
41:05output or like text generation
41:09within uh with minimum effort min with
41:14minimum effort.

About this transcript

This page contains the full transcript of Session 2: Potential Applications of AI in Various Purposes by Development Academy of the Philippines, generated from the public captions YouTube serves with the video. The transcript has 4,666 words across 723 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.