Latest / Azadeh and Guy on AI / AI Insights

Transcript

AI Insights · Azadeh and Guy on AI

• 0:00
  SPEAKER_00: Hello everyone, hello from the first episode of Azadeh and Gayan.
• 0:05
  SPEAKER_01: AI.
• 0:06
  SPEAKER_00: AI, perfectly.
• 0:07
  SPEAKER_00: I'm Azadeh,
• 0:08
  I'm a third year PhD student here at the University of
• 0:11
  Nebraska-Lincoln and...
• 0:13
  SPEAKER_01: I'm Gai Tranin,
• 0:14
  SPEAKER_01: I'm a professor at the Innovative Learning Technologies...
• 0:19
  SPEAKER_01: Today we're going to talk about two things.
• 0:22
  SPEAKER_01: Azadeh is going to introduce me to perplexity,
• 0:24
  SPEAKER_01: which I kind of downloaded once but never actually did anything with it.
• 0:28
  SPEAKER_01: It sits on my phone still, but not much.
• 0:31
  And we're going to talk about what it can and can't do from
• 0:36
  very casual terms.
• 0:39
  And then the second thing is we're going to actually talk about
• 0:42
  the research paper
• 0:43
  SPEAKER_01: that I found that I find incredibly helpful.
• 0:46
  And that is a paper about generative artificial intelligence
• 0:49
  and misinformation acceptance.
• 0:52
  SPEAKER_00: Okay, perfect.
• 0:54
  SPEAKER_00: So the Gen AI tool that we are talking about today is perplexity.
• 0:59
  SPEAKER_00: And I think many people have heard about that.
• 1:02
  SPEAKER_00: Actually, in our last episode, I think we kind of talked about that.
• 1:07
  SPEAKER_00: We touched slightly on perplexity.
• 1:10
  SPEAKER_00: But today we tried to explore it more.
• 1:13
  SPEAKER_00: I find it a very useful tool for doing research.
• 1:17
  To be honest,
• 1:18
  I haven't explored all the features,
• 1:20
  but I'm very excited to try it more.
• 1:25
  SPEAKER_00: And I didn't download it.
• 1:29
  SPEAKER_00: I just used the desktop version of Perplexity.
• 1:32
  SPEAKER_00: I wanted to get more used to the different features of this tool,
• 1:36
  SPEAKER_00: and then I will download it on my phone.
• 1:39
  SPEAKER_00: So for perplexity, I think you have the pro version guy.
• 1:45
  SPEAKER_00: Am I right?
• 1:46
  SPEAKER_01: I do not have the pro version yet.
• 1:48
  SPEAKER_01: I have the version they share with educators.
• 1:50
  SPEAKER_01: So they have a version they share with educators.
• 1:53
  SPEAKER_01: They have a version for students who do get the pro version.
• 1:57
  SPEAKER_01: And so students,
• 1:58
  SPEAKER_01: if you are students and you can verify your university students,
• 2:01
  SPEAKER_01: they will give you a version of the pro version.
• 2:03
  SPEAKER_01: I'm not sure if it's the full pro version.
• 2:06
  SPEAKER_01: I'm a little bit lost in the versions.
• 2:07
  SPEAKER_01: But I'm right now playing with a...
• 2:11
  SPEAKER_01: Free version?
• 2:12
  SPEAKER_01: No, it's a pro version.
• 2:13
  SPEAKER_00: It's a pro version.
• 2:14
  SPEAKER_01: I'm a pro for anybody who didn't know.
• 2:18
  SPEAKER_00: Yes, you are a pro version.
• 2:20
  SPEAKER_00: And as you said,
• 2:21
  they offer a one-month free pro version for students as long as
• 2:25
  you just enter your
• 2:28
  SPEAKER_00: university email and they are able to verify your student status.
• 2:33
  SPEAKER_00: And then they provide you with some interesting features.
• 2:37
  SPEAKER_00: One of the features that I'm really interested about is,
• 2:40
  SPEAKER_00: we were just talking about that,
• 2:44
  SPEAKER_00: is spaces.
• 2:45
  SPEAKER_00: So if you look at the left navigation menu,
• 2:50
  SPEAKER_00: you can find spaces in a kind of,
• 2:53
  SPEAKER_00: they have a cubic icon for spaces.
• 2:55
  SPEAKER_00: And...
• 2:57
  SPEAKER_00: What I really like about that feature is that you can add your files,
• 3:03
  SPEAKER_00: you can add links,
• 3:04
  and you can create a kind of very personalized and maybe
• 3:09
  contextualized repository
• 3:11
  SPEAKER_00: of all of the things that you are trying to do research about.
• 3:17
  And you can also add your teammates,
• 3:20
  and you can all start a group project.
• 3:23
  SPEAKER_00: Yeah.
• 3:23
  SPEAKER_01: So this is,
• 3:24
  SPEAKER_01: I think,
• 3:25
  SPEAKER_01: the feature that's important because you could always share,
• 3:28
  but here is a feature where you have a shared space where you
• 3:32
  can all contribute
• 3:33
  SPEAKER_01: and it's more collaborative.
• 3:35
  SPEAKER_01: Kind of reminds me of the idea behind Google Docs or something like that.
• 3:39
  SPEAKER_01: Everybody can contribute.
• 3:41
  It's not just me sending it to you,
• 3:43
  you doing something with it and sending it back to me.
• 3:45
  SPEAKER_01: It's a shared space.
• 3:48
  Or I can say Google Docs or Google Drive,
• 3:52
  but maybe three times more flexible.
• 3:55
  SPEAKER_01: Yeah.
• 3:56
  SPEAKER_01: But the idea is a collaborative space where you can work concurrently.
• 4:01
  SPEAKER_01: That's my analogy at least.
• 4:04
  SPEAKER_00: Yeah, yeah, yeah.
• 4:05
  SPEAKER_01: So what can you do there?
• 4:06
  SPEAKER_01: So you can add links.
• 4:07
  SPEAKER_01: I see you can add links.
• 4:09
  SPEAKER_01: You can add instructions.
• 4:10
  SPEAKER_01: You can add files.
• 4:11
  SPEAKER_01: So this is a productive idea.
• 4:14
  a productive space,
• 4:16
  and I love that notion that we can do some of our exploration.
• 4:22
  So the things we used to use the browser for or the search
• 4:27
  function in Google for,
• 4:29
  SPEAKER_01: I see this very,
• 4:30
  SPEAKER_01: very useful,
• 4:31
  SPEAKER_01: is you can ask a question and get the information that is more detailed.
• 4:38
  SPEAKER_00: Yeah, that's true.
• 4:39
  SPEAKER_00: And the other thing about the spaces feature is that
• 4:47
  They have settings,
• 4:49
  and in that space,
• 4:50
  you can formulate your instructions.
• 4:54
  SPEAKER_00: You can say,
• 4:55
  SPEAKER_00: I mean,
• 4:56
  SPEAKER_00: the description says,
• 4:57
  SPEAKER_00: give instructions to perplexity to customize how answers are focused and
• 5:01
  SPEAKER_00: structured.
• 5:03
  So I really like that kind of,
• 5:04
  because you basically dictate Gen.AI about the outputs.
• 5:08
  SPEAKER_00: Okay.
• 5:09
  SPEAKER_00: And you can also include web by default.
• 5:12
  SPEAKER_00: So you cannot just rely on your files.
• 5:16
  SPEAKER_00: You can ask Gen.ai,
• 5:17
  SPEAKER_00: or in that case Perplexity,
• 5:19
  SPEAKER_00: to go through the different websites to just amplify the output.
• 5:26
  SPEAKER_01: And so I,
• 5:27
  SPEAKER_01: for example,
• 5:28
  did a search about the relationship between state CS education
• 5:32
  policies and the
• 5:33
  SPEAKER_01: median salary in the state.
• 5:35
  SPEAKER_01: I just threw a question that I was interested about at the time.
• 5:40
  SPEAKER_01: And what I love about that is I got visuals, I got analysis.
• 5:45
  SPEAKER_01: I've got interesting analysis because it actually found that the –
• 5:50
  SPEAKER_01: The correlation is not very high.
• 5:53
  SPEAKER_01: So the R squared is 0.04.
• 5:55
  SPEAKER_01: That's nothing.
• 5:56
  SPEAKER_01: But it says there's actually detail behind it.
• 5:59
  SPEAKER_01: So it explores a little bit further and it says,
• 6:02
  SPEAKER_01: well,
• 6:02
  SPEAKER_01: you know,
• 6:03
  SPEAKER_01: there's an interesting correlation.
• 6:04
  And part of it is that states that have lower income seems to
• 6:09
  have more CSEs.
• 6:12
  education policy because they seem to be looking for ways to
• 6:18
  improve things in their state.
• 6:19
  And so I got this really nice report with visuals and text and
• 6:25
  all of these things
• 6:26
  SPEAKER_01: that would take me quite a while to correlate and collect together.
• 6:30
  SPEAKER_01: And that's something I appreciate.
• 6:32
  SPEAKER_00: Yeah, yeah, yeah.
• 6:32
  SPEAKER_00: That's very true.
• 6:34
  SPEAKER_00: The other option or feature of this Gen AI tool is labs.
• 6:40
  So Perplexity itself defined labs as advanced tools for
• 6:46
  developers and power users
• 6:48
  SPEAKER_00: to create custom apps,
• 6:49
  SPEAKER_00: perform deep analysis,
• 6:51
  SPEAKER_00: and automate complex tasks.
• 6:54
  SPEAKER_00: Sorry.
• 6:55
  SPEAKER_00: So if you go to Perplexity website,
• 6:58
  SPEAKER_00: they also provide you with a kind of sample of different labs.
• 7:04
  SPEAKER_00: And one of the samples that I checked was,
• 7:08
  SPEAKER_00: I can just read you the first few sentences of the form.
• 7:11
  SPEAKER_00: So they say you are a product designer,
• 7:14
  SPEAKER_00: AI researcher,
• 7:14
  SPEAKER_00: and behavioral psychologist in the year 2025.
• 7:18
  Your task is to design a futuristic social media platform made
• 7:22
  for the year 2030.
• 7:25
  SPEAKER_00: So you can talk about the core concept,
• 7:28
  SPEAKER_00: key features,
• 7:29
  SPEAKER_00: growth mechanics,
• 7:30
  SPEAKER_00: and it creates a whole project for you,
• 7:34
  SPEAKER_00: answering all of those questions,
• 7:37
  talking about possibilities of connecting that kind of social
• 7:42
  media platform to VR
• 7:44
  SPEAKER_00: even experiences.
• 7:46
  SPEAKER_00: So this is also a very interesting feature of that Gen.ai tool.
• 7:51
  SPEAKER_01: And much like the other AIs,
• 7:53
  it does have a learn section where it supports you in learning
• 7:59
  step by step and
• 8:00
  SPEAKER_01: really making these kind of supports for learning.
• 8:03
  This is something that many of the AI companies are starting to
• 8:08
  figure out,
• 8:08
  and that is...
• 8:10
  SPEAKER_01: People want to study with AI.
• 8:11
  SPEAKER_01: They don't just want the answer.
• 8:13
  And that learn space is usually the space where it can scaffold
• 8:17
  your understanding
• 8:19
  and try to help you and give you tools instead of just giving
• 8:22
  you answers.
• 8:23
  SPEAKER_00: Yeah, yeah.
• 8:24
  And it's interesting because even in spaces,
• 8:27
  you can also activate a study mode lab.
• 8:30
  SPEAKER_00: And you can, by using that, you can do the step-by-step learning.
• 8:35
  SPEAKER_00: Yeah.
• 8:36
  SPEAKER_00: So I think most Gen.AI tools are becoming more self-explanatory.
• 8:40
  SPEAKER_00: They are trying to offer you some features.
• 8:44
  And at the same time,
• 8:45
  you can use Gen.AI to learn about Gen.AI and its features.
• 8:50
  SPEAKER_01: Yeah, that seems kind of meta.
• 8:52
  But the piece that I liked the best out of all of this is the
• 8:57
  ability to create automated tasks.
• 8:59
  SPEAKER_01: And this is available in a few of the generative AIs.
• 9:03
  SPEAKER_01: Actually, I think in most of them.
• 9:04
  SPEAKER_01: But I find perplexity's approach very, very simple.
• 9:09
  SPEAKER_01: So you can just give it a regular task.
• 9:11
  For example,
• 9:12
  I said,
• 9:13
  search for airline tickets under a certain price every Friday.
• 9:18
  I want to maybe at least dream about getting on an airplane and
• 9:21
  disappearing for the weekend.
• 9:23
  SPEAKER_01: Yeah.
• 9:23
  SPEAKER_01: So you can get a report at 1 p.m.
• 9:27
  SPEAKER_01: every Friday to say,
• 9:29
  SPEAKER_01: hey,
• 9:29
  SPEAKER_01: there's a cheap ticket from,
• 9:30
  SPEAKER_01: in my case,
• 9:31
  SPEAKER_01: Omaha,
• 9:32
  SPEAKER_01: most likely,
• 9:34
  SPEAKER_01: going to Cleveland,
• 9:36
  SPEAKER_01: which I had fun at this summer,
• 9:39
  SPEAKER_01: so I want to go again.
• 9:41
  And it's going to be only $250,
• 9:44
  and I'm in the airport that evening,
• 9:46
  and that'll be that.
• 9:47
  SPEAKER_01: So you can create this whole series of automatic tasks.
• 9:50
  So if there's something that you need to update on a regular
• 9:53
  basis in your
• 9:53
  SPEAKER_01: business,
• 9:54
  SPEAKER_01: in school or something like that,
• 9:56
  SPEAKER_01: this is a great opportunity to have that.
• 10:00
  SPEAKER_01: You can connect it to your email.
• 10:01
  SPEAKER_01: So,
• 10:01
  SPEAKER_01: for example,
• 10:02
  SPEAKER_01: you can say,
• 10:02
  SPEAKER_01: hey,
• 10:02
  can you give me a report on Friday about all the emails I
• 10:05
  missed and I should
• 10:06
  SPEAKER_01: probably respond to before the weekend sets in or things like that.
• 10:10
  SPEAKER_01: That seems to be a really reasonable way to work your workflows.
• 10:15
  SPEAKER_00: So they are not basically just tasks.
• 10:18
  SPEAKER_00: They can be customized or personalized agents, right?
• 10:21
  SPEAKER_00: Yes.
• 10:22
  But the other thing that I want to mention that they kind of
• 10:27
  hide that very useful feature.
• 10:31
  SPEAKER_00: So if you want to access tasks,
• 10:33
  SPEAKER_00: you should click on your icon in the left navigation menu,
• 10:38
  and then you can click on task and then define the task and
• 10:41
  even talk about the
• 10:43
  SPEAKER_00: timing of the task.
• 10:45
  SPEAKER_01: Yeah.
• 10:46
  SPEAKER_01: And I think that in some ways,
• 10:48
  SPEAKER_01: I find that the secondary AI labs,
• 10:52
  SPEAKER_01: and we have to be honest,
• 10:53
  SPEAKER_01: OpenAI owns most of the market.
• 10:57
  SPEAKER_01: And so that forces the other AI companies to offer you something extra,
• 11:03
  SPEAKER_01: either cheaper,
• 11:04
  SPEAKER_01: which is happening...
• 11:05
  SPEAKER_01: Or something extra.
• 11:07
  SPEAKER_01: And in this case,
• 11:08
  I think Perplexity is doing some really interesting things with
• 11:11
  their interface
• 11:12
  that allows you to get quicker into having an agent or doing
• 11:18
  something on a regular
• 11:19
  SPEAKER_01: basis like in Tasks.
• 11:20
  SPEAKER_00: Sure, sure, yes.
• 11:22
  SPEAKER_01: All right, I think it's time to talk about the paper.
• 11:24
  SPEAKER_00: Sure, yeah.
• 11:25
  All right,
• 11:25
  so we talked about perplexity,
• 11:27
  and now I want to talk about the paper.
• 11:29
  SPEAKER_01: This is a problem everybody's talking about, hallucinations, which...
• 11:36
  SPEAKER_01: Just recently,
• 11:37
  SPEAKER_01: we found out from some research,
• 11:38
  SPEAKER_01: may actually be a result of the training,
• 11:43
  SPEAKER_01: the post-training reinforcement and not the actual LLM.
• 11:49
  So there has been this hypothesis with some support that the
• 11:55
  LLMs hallucinate
• 11:56
  SPEAKER_01: because in reinforcement learning.
• 11:59
  So when people interact with them...
• 12:01
  Based on the feedback that they get from...
• 12:04
  SPEAKER_01: And that feedback is always on the positive side.
• 12:08
  SPEAKER_01: It never rewards them when they say just, I don't know.
• 12:14
  So they are much more likely to try and invent and maybe hit a
• 12:18
  home run than not.
• 12:20
  So it's a reinforcement thing,
• 12:22
  which is really interesting to think about.
• 12:23
  SPEAKER_01: But here are two researchers...
• 12:26
  SPEAKER_01: Huang and Zhong,
• 12:28
  and they are looking at generative artificial intelligence and
• 12:33
  misinformation
• 12:34
  SPEAKER_01: acceptance,
• 12:35
  an experimental test of the effect of forewarning about
• 12:38
  artificial intelligence
• 12:39
  SPEAKER_01: hallucination.
• 12:41
  SPEAKER_01: Long title.
• 12:43
  SPEAKER_01: First, I want to say that...
• 12:46
  SPEAKER_01: Area of AI in general,
• 12:48
  SPEAKER_01: especially AI in education and in society,
• 12:50
  SPEAKER_01: is seeing an explosion of articles.
• 12:53
  SPEAKER_01: Not all of them are great, even if they are peer-reviewed.
• 12:56
  And too many papers are getting air and publicity before they
• 13:01
  are peer-reviewed.
• 13:02
  SPEAKER_01: This is a peer-reviewed article.
• 13:04
  SPEAKER_01: And at least from reading it, it is fairly well executed.
• 13:09
  SPEAKER_01: So the sample size, for example, is 208 Korean adults.
• 13:13
  So we're talking about a large enough sample to learn something
• 13:18
  about,
• 13:18
  obviously.
• 13:19
  SPEAKER_01: And they say that these are Korean adults.
• 13:21
  It might work or not work in other countries,
• 13:24
  but you've got to start somewhere.
• 13:25
  SPEAKER_01: You've got to start with a place that you're at.
• 13:27
  SPEAKER_01: And what they did for them is they really had them...
• 13:32
  look at AI results and examine whether they were accepting
• 13:38
  misinformation or not,
• 13:39
  SPEAKER_01: and they had two conditions.
• 13:41
  SPEAKER_01: One of them was just,
• 13:42
  SPEAKER_01: here it is,
• 13:43
  SPEAKER_01: here's the AI,
• 13:44
  SPEAKER_01: here are the results,
• 13:45
  SPEAKER_01: and in the other one,
• 13:45
  SPEAKER_01: they just did a reminder.
• 13:47
  SPEAKER_01: Remember,
• 13:48
  SPEAKER_01: AI sometimes gives misinformation,
• 13:52
  SPEAKER_01: and we as consumers should be really careful because...
• 13:59
  SPEAKER_01: It does that.
• 14:00
  SPEAKER_01: And what they found is something as simple as just telling people,
• 14:06
  SPEAKER_01: you need to be careful,
• 14:07
  SPEAKER_01: you need to think about what you're consuming,
• 14:09
  SPEAKER_01: did make a big difference.
• 14:12
  SPEAKER_01: And it made an interesting difference because they examined this.
• 14:15
  SPEAKER_01: It made people not trust hallucinations when they were wrong,
• 14:19
  SPEAKER_01: but actually trust the results differently.
• 14:21
  SPEAKER_01: When they were right.
• 14:23
  SPEAKER_01: So it just made them better consumers.
• 14:25
  SPEAKER_01: It didn't,
• 14:26
  SPEAKER_01: because,
• 14:26
  SPEAKER_01: I mean,
• 14:27
  the fear everybody's talking about right now is that with so
• 14:31
  much AI slop,
• 14:33
  SPEAKER_01: the fear is not that we will believe lies,
• 14:36
  SPEAKER_01: but the fear is that we will not believe anything.
• 14:38
  SPEAKER_00: Anything, yeah.
• 14:39
  SPEAKER_00: Yeah, yeah, yeah.
• 14:40
  SPEAKER_00: So raising people's awareness about, yeah.
• 14:43
  SPEAKER_00: Mm-hmm.
• 14:44
  SPEAKER_01: And that's a really good design from a study perspective.
• 14:49
  SPEAKER_00: It's an online experiment, right?
• 14:51
  SPEAKER_00: Yes.
• 14:51
  SPEAKER_00: They had two groups.
• 14:53
  SPEAKER_01: It is an online experiment.
• 14:55
  SPEAKER_01: They worked it out.
• 14:56
  SPEAKER_01: They're working with adults.
• 14:57
  So this is not a result for kids,
• 14:59
  but we're working often with college students and adults.
• 15:02
  SPEAKER_00: They also talked about the age range of the participants, right?
• 15:06
  SPEAKER_00: Yes.
• 15:07
  SPEAKER_00: Yeah.
• 15:07
  SPEAKER_00: They range from 20 to 59.
• 15:08
  SPEAKER_00: Yeah.
• 15:12
  SPEAKER_00: That's a good range.
• 15:13
  SPEAKER_00: Okay.
• 15:13
  SPEAKER_01: Yeah.
• 15:13
  SPEAKER_01: So it is very convincing and it seems to be well executed.
• 15:19
  SPEAKER_01: So thank you for the authors of that.
• 15:21
  SPEAKER_01: And basically what it says,
• 15:23
  SPEAKER_01: if you are using AI with your students,
• 15:25
  SPEAKER_01: if you are using AI in general,
• 15:29
  SPEAKER_01: just remembering it can hallucinate.
• 15:32
  SPEAKER_01: Even warning yourself, I think sometimes.
• 15:33
  SPEAKER_00: Yeah, yeah, yeah.
• 15:34
  SPEAKER_01: Remember.
• 15:35
  SPEAKER_01: And the AI sometimes do that.
• 15:37
  The problem is when the same warning comes up every time,
• 15:40
  you stop paying attention.
• 15:41
  SPEAKER_01: Mm-hmm.
• 15:42
  SPEAKER_01: So that specialized warning, it's like, this is great, but let's check.
• 15:47
  SPEAKER_00: Let's check, yes.
• 15:48
  SPEAKER_01: And I was thinking about this thing we did with perplexity.
• 15:51
  And with perplexity,
• 15:52
  for example,
• 15:53
  when I did my query,
• 15:54
  it did give me the sources.
• 15:56
  SPEAKER_01: It says, oh, this thing I got from code.org.
• 15:58
  SPEAKER_01: This thing I got from this.
• 16:00
  SPEAKER_01: So it showed me exactly where the data came from, and that was powerful.
• 16:03
  And it even showed the code that it created in Python to create
• 16:06
  the charts and all that.
• 16:07
  SPEAKER_01: So it was lovely.
• 16:09
  And you can always click on those links to make sure that they
• 16:13
  are directing you to
• 16:14
  SPEAKER_00: the right website.
• 16:16
  Or even sometimes the content is well aligned with what you
• 16:19
  have in your mind.
• 16:22
  SPEAKER_01: And so really, what am I trying to say?
• 16:24
  SPEAKER_01: I'm trying to say two things about this study.
• 16:27
  SPEAKER_01: One is that we can do something about hallucination,
• 16:31
  and that is be careful and use our human minds to sort through
• 16:35
  and make sure,
• 16:36
  SPEAKER_01: validate,
• 16:37
  SPEAKER_01: but also just process through,
• 16:39
  SPEAKER_01: is this really,
• 16:40
  SPEAKER_01: does this seem reasonable?
• 16:41
  SPEAKER_01: And the second thing is that there is good...
• 16:46
  SPEAKER_01: Because there's so much.
• 16:47
  SPEAKER_01: I mean, we were talking about this earlier today.
• 16:50
  SPEAKER_01: There's a lot of not so good research out there.
• 16:53
  SPEAKER_01: The sample sizes are too small.
• 16:55
  SPEAKER_01: It's selective.
• 16:56
  SPEAKER_01: There's bias.
• 16:58
  SPEAKER_01: The instrumentation is not very good.
• 17:01
  SPEAKER_00: The analysis is not... Yeah, solid and yeah.
• 17:04
  And this seems to be one of those studies where you're like,
• 17:06
  okay,
• 17:07
  there's quality out there.
• 17:08
  SPEAKER_01: We just have to go and look for it.
• 17:10
  SPEAKER_01: And if you're consuming that information, take a look at the methodology.
• 17:14
  SPEAKER_01: Make sure that it's peer-reviewed,
• 17:17
  but also make sure that the methodology makes sense and it
• 17:20
  really answers what they
• 17:22
  claim it answers and not something else because we want to
• 17:27
  learn from research,
• 17:27
  SPEAKER_01: but only from good research.
• 17:29
  SPEAKER_00: Yeah, yeah, sure.
• 17:31
  SPEAKER_01: All right, so today we talked about... Paraplexity.
• 17:34
  SPEAKER_01: And about avoiding believing in hallucinations.
• 17:40
  SPEAKER_01: And we'll see you next time on Azadeh and Guy on AI.