Generative AI Challenge - Fundamentals of Responsible Generative AI

00:54:31
https://www.youtube.com/watch?v=C2SaTllLwN4

摘要

TLDRIn a live discussion, Frank and Cory explore the critical aspects of implementing responsible AI in generative applications. The session emphasizes understanding and mitigating potential harms, ensuring ethical use, and safeguarding AI applications. They outline a step-by-step approach: identifying potential harms, prioritizing them, testing and verifying, and sharing documentation. The hosts also cover how to address misuse of AI, the significance of teamwork in AI projects, and continuous monitoring for long-term model safety. They conduct a live demo using Azure's AI Studio to show content filtering and responsible AI deployment. Audience questions were addressed about secure deployment, evolving AI models, and strategies for dealing with breaches or unethical use. Participants are encouraged to engage with the challenge before its 2024 deadline. Overall, the event seeks to educate viewers on making generative AI a powerful but safe tool.

心得

  • 🤔 Understand the importance of responsible AI in generative applications.
  • 🛡️ Identify and prioritize potential harms to mitigate risks.
  • 🔍 Continuously test and verify the AI model's safety and correctness.
  • 🔄 Adapt to evolving models to maintain consistent safety.
  • 📚 Properly document and share insights for team awareness.
  • 🌐 Utilize Azure AI Studio for secure and responsible deployment.
  • 🔒 Consider a layered approach to content filtering and security.
  • ⏰ Meet the AI challenge deadline by September 20, 2024.
  • 🛠️ Involve cross-functional teams for comprehensive AI project governance.
  • 📢 Openly share findings and improvements in the AI Community.

时间轴

  • 00:00:00 - 00:05:00

    The video starts with a lengthy introduction filled with music, leading to a discussion about how the introduction felt prolonged. Frank and Cory introduce themselves and welcome viewers, expressing excitement for the session.

  • 00:05:00 - 00:10:00

    Frank and Cory continue engaging with the chat, appreciating returning viewers and introducing the session's topic. They emphasize the significance of discussing 'Fundamentals of Responsible Generative AI' and encourage viewers not to miss this important topic.

  • 00:10:00 - 00:15:00

    The session is aimed at introducing considerations for responsible AI use, focusing on responsiveness in AI. The discussion revolves around the placement and timing of this topic in their educational series, with humor about its importance and timing.

  • 00:15:00 - 00:20:00

    Frank and Cory discuss the structure of their series, encouraging viewers to follow the presented order, and make references to chat engagements and user questions. They motivate new and returning viewers to join in on the learning journey.

  • 00:20:00 - 00:25:00

    They explain the importance of responsible AI in real-world applications, illustrating the process with slides that show essential steps: identifying harm, prioritizing, testing, and documentation. Emphasis is laid on sharing learnings with the wider community.

  • 00:25:00 - 00:30:00

    They delve into common harms associated with generative AI, addressing issues like discrimination and factual inaccuracy. The hosts stress the importance of safeguards against these harms through well-prepared prompts and testing strategies.

  • 00:30:00 - 00:35:00

    The discussion turns to proactive strategies for minimizing these harms by testing and validating through prompts and automated testing, and evaluating based on user behavior. They mention monitoring and demographic considerations.

  • 00:35:00 - 00:40:00

    Strategies for harm mitigation involve a layered approach, discussing model tuning and the integration of content filters. They highlight the balance required to effectively monitor user input and model output while maintaining user engagement.

  • 00:40:00 - 00:45:00

    Frank and Cory highlight governance, pre-release reviews, and assembling a cross-functional team adept in legal and privacy considerations. They stress developing an incident response plan to quickly act when unexpected scenarios occur.

  • 00:45:00 - 00:54:31

    The video concludes with a call to action, encouraging viewers to participate in an AI challenge. They align past learning sessions with the challenge and remind viewers of the deadline. The session ends with mutual thanks and encouragement for continued learning.

显示更多

思维导图

Mind Map

常见问题

  • What is the focus of the video?

    The focus of the video is on implementing responsible generative AI.

  • Who are the hosts of the video?

    Frank and Cory host this discussion on responsible generative AI.

  • What topics are covered in relation to responsible AI?

    The video highlights responsible AI, potential harms, and mitigation strategies.

  • What are some common types of potential harms discussed?

    Some common harms include discrimination, factual inaccuracies, and misuse for unethical purposes.

  • Was the session live and interactive?

    Yes, the session was live and included addressing viewer questions and demonstrating concepts.

  • How do they address content filtering in the video?

    They discuss and demonstrate content filtering using Azure's AI Studio.

  • When is the deadline for the AI challenge mentioned?

    The challenge deadline is September 20, 2024.

  • How can you evaluate and mitigate potential AI harms?

    You can look at user inputs, employ content safety tools, and have a layered approach to security.

  • Is the session suitable for newcomers to generative AI?

    Yes, students starting with this module will find it easy to follow even if they missed previous sessions.

  • What do they say about the evolution and safety measures of AI models over time?

    Their discussion touches on the iterative importance and adaptation of AI model safety over time.

查看更多视频摘要

即时访问由人工智能支持的免费 YouTube 视频摘要！
字幕
en
自动滚动:
  • 00:00:01
    [Music]
  • 00:00:36
    [Music]
  • 00:01:03
    [Music]
  • 00:01:20
    [Music]
  • 00:01:44
    [Music]
  • 00:01:58
    [Music]
  • 00:02:14
    [Music]
  • 00:02:30
    [Music]
  • 00:03:04
    [Music]
  • 00:03:11
    [Music]
  • 00:03:18
    [Music]
  • 00:03:59
    [Music]
  • 00:04:12
    [Music]
  • 00:04:19
    [Music]
  • 00:04:31
    [Music]
  • 00:04:37
    [Music]
  • 00:04:42
    [Music]
  • 00:04:49
    [Music]
  • 00:04:54
    there it
  • 00:04:55
    is I was curious to see what happens at
  • 00:04:58
    the end yeah that was the longest five
  • 00:05:01
    minutes of my life I don't know why that
  • 00:05:04
    was it that felt way longer than the
  • 00:05:06
    typical five minutes uh I didn't time
  • 00:05:09
    the timer but maybe we should have
  • 00:05:10
    audited that because I don't know what
  • 00:05:11
    what was going on there feel like it was
  • 00:05:13
    taking that's how we are that's how much
  • 00:05:16
    we are excited to uh to for for the talk
  • 00:05:19
    today hey welcome everybody I'm Frank
  • 00:05:22
    and with me have this way Cory yeah yeah
  • 00:05:24
    there you go it's a directional thing I
  • 00:05:26
    think you're no so you're this way Cory
  • 00:05:28
    hello everyone welcome I see a few
  • 00:05:30
    people in the chat RI Manaj nice to see
  • 00:05:34
    you thank you for all sitting through
  • 00:05:35
    those longest five minutes so uh I hope
  • 00:05:37
    we didn't lose everyone in that time for
  • 00:05:40
    the longest five minutes ever should
  • 00:05:42
    bring let's bring the title here for
  • 00:05:46
    this so this is the we we'll share more
  • 00:05:49
    but uh this is like
  • 00:05:51
    a final of a series of uh sessions I
  • 00:05:55
    will share all the links don't worry you
  • 00:05:58
    can do that one first and then catch up
  • 00:06:01
    and watch the other session after so
  • 00:06:03
    this one don't leave us don't leave us
  • 00:06:05
    please no no no like
  • 00:06:07
    exactly but this one is fundamentals of
  • 00:06:10
    responsib responsive generative AI very
  • 00:06:14
    important topic honestly it's very
  • 00:06:17
    important to not let it go not let it
  • 00:06:19
    someone someone some would say the most
  • 00:06:21
    important topic but as you said this is
  • 00:06:23
    this is a typical thing I'm sure you if
  • 00:06:25
    you can bring up the the schedule
  • 00:06:27
    sometimes we bring up we bring in the
  • 00:06:28
    responsible a AI part at the last I'm
  • 00:06:32
    hoping I'm hoping all of the other
  • 00:06:34
    sessions were talking about
  • 00:06:36
    irresponsible AI they all have hopefully
  • 00:06:38
    bits and elements of what we're going to
  • 00:06:40
    talk about but this is like the
  • 00:06:42
    summarization of it all this is what I'm
  • 00:06:43
    hoping at least yeah so it's a challenge
  • 00:06:48
    so feel free to scan that or like just
  • 00:06:50
    like use those those uh those clip and
  • 00:06:53
    here thanks Angel putting the the URL in
  • 00:06:56
    the chat so uh feel free to join us
  • 00:07:00
    it's it's important and and with that it
  • 00:07:02
    will make you way better with generative
  • 00:07:05
    AI like how to do it do it the correct
  • 00:07:07
    way and and we'll see today like the
  • 00:07:10
    correct the proper way the responsible
  • 00:07:13
    the responsible way yeah there we go
  • 00:07:15
    that's that's word is hard for me I
  • 00:07:17
    should I should say it in French could
  • 00:07:18
    we switch in French yeah I think I mean
  • 00:07:20
    that's very responsible of us to just
  • 00:07:22
    switch languages yeah middle of session
  • 00:07:26
    in fact you know some people suggested
  • 00:07:27
    to me just to see how you react Curry to
  • 00:07:30
    start the this session in Fr I'm you
  • 00:07:33
    know I'm a fast think my Fe I I can't
  • 00:07:35
    learn um I mean if we give me five
  • 00:07:37
    minutes like that five minutes ago I
  • 00:07:40
    could probably learn French in that time
  • 00:07:43
    having like copilot by deci side like
  • 00:07:46
    yes hi
  • 00:07:50
    Frank awesome so we we were talking
  • 00:07:53
    about all the previous sessions so here
  • 00:07:54
    they are so this is the final one it was
  • 00:07:57
    kind of like a five little
  • 00:08:00
    sessions I think the the longest one was
  • 00:08:03
    about an hour they go through the module
  • 00:08:06
    helping you to understand the topic of
  • 00:08:09
    that specific uh module so they are
  • 00:08:13
    really kind of it's five modules so five
  • 00:08:15
    video and they go through quickly just
  • 00:08:18
    to make sure you could log in do those
  • 00:08:21
    module and have the the badge like do
  • 00:08:25
    the The Challenge and uh I think it's
  • 00:08:28
    important so we can start you can start
  • 00:08:32
    by this one maybe it's your first let us
  • 00:08:34
    know in the chat you know what I'm I'm
  • 00:08:35
    curious how much people like
  • 00:08:38
    did like follow the the entire or maybe
  • 00:08:41
    it was just one and you're planning to
  • 00:08:43
    catch up later or maybe you start with
  • 00:08:46
    us that's totally fine let us know in
  • 00:08:48
    the chat I'm curious start from the
  • 00:08:50
    bottom work your way up or start in the
  • 00:08:51
    middle work your way down I don't know
  • 00:08:53
    but glad you're here that's all I can
  • 00:08:55
    say yeah exactly let's move the slides
  • 00:08:59
    so there it is so that was
  • 00:09:03
    the some QR code yeah that's what we're
  • 00:09:08
    doing today yeah I threw it a little bit
  • 00:09:10
    extra and I'll talk about it later on
  • 00:09:12
    but so we are first looking at this uh
  • 00:09:16
    lesson here the Implement rag with or
  • 00:09:19
    maybe this is I feel the title is wrong
  • 00:09:23
    URL is right so hopefully code is going
  • 00:09:25
    to be the deciding factor I don't know
  • 00:09:27
    that's wrong I was like no well there's
  • 00:09:30
    one responsible module uh and I think
  • 00:09:32
    that's a good introduction I guess
  • 00:09:35
    you're um kind of new to this this topic
  • 00:09:37
    of responsible AI it really we're going
  • 00:09:39
    to cover some of those um some of the
  • 00:09:41
    foundations of you know what responsible
  • 00:09:44
    AI is and this lesson is great but you
  • 00:09:46
    know responsible AI like I said I think
  • 00:09:49
    it's the most one of the most important
  • 00:09:50
    topics especially when you're building
  • 00:09:52
    generative AI applications so I threw in
  • 00:09:53
    a little extra collection which I'll
  • 00:09:55
    kind of show you what's in there and I
  • 00:09:56
    think that's like if the Le the first
  • 00:09:59
    lesson is like the starter for the meal
  • 00:10:01
    like this is like the uh the the the the
  • 00:10:04
    entree or the the big the big meal after
  • 00:10:06
    so uh both of these are great resources
  • 00:10:09
    hopefully the QR code someone confirm in
  • 00:10:11
    the chat if that's going to the right
  • 00:10:13
    one maybe we should have done that
  • 00:10:14
    beforehand but um hopeing that's going
  • 00:10:17
    to the right responsible AI did you add
  • 00:10:18
    a slide last minute and you didn't do
  • 00:10:21
    your unit test I know this is okay this
  • 00:10:23
    could be on me this is irresponsible
  • 00:10:26
    this is irresponsible Cory
  • 00:10:30
    oh and we see some people did other uh
  • 00:10:33
    check that's cool second and third so
  • 00:10:35
    this is like the third one for you then
  • 00:10:37
    this is yeah if anyone's done the season
  • 00:10:39
    pass uh this is good but this is yeah
  • 00:10:42
    three is a solid number so how do you
  • 00:10:44
    how do you want to proceed Cory should
  • 00:10:46
    we stay in the slide you want to open
  • 00:10:48
    the the module and like go through the
  • 00:10:50
    content over there like how how would
  • 00:10:52
    you like to do that let's stick to the
  • 00:10:54
    slides first and then um we do got a
  • 00:10:56
    little we have a little demo coming up
  • 00:10:58
    which will kind of put a lot of these
  • 00:11:00
    Concepts into practice which I think
  • 00:11:02
    when we talk about responsible AI that's
  • 00:11:04
    that's the tricky part like everyone can
  • 00:11:06
    come with a deck and nice ideas but then
  • 00:11:09
    how do you actually Implement those yeah
  • 00:11:11
    really see that you know not planning
  • 00:11:15
    responsively could put you in trouble
  • 00:11:18
    for sure so there's four steps like how
  • 00:11:23
    how like like I see to like be
  • 00:11:26
    responsible you identify the potential
  • 00:11:31
    arms so I I don't want to read
  • 00:11:33
    everything but like you make sure you
  • 00:11:35
    identify we'll go into details later you
  • 00:11:37
    prioritize them because some arms may be
  • 00:11:41
    less prioritized you know like like
  • 00:11:43
    nobody will die it just maybe yeah like
  • 00:11:46
    maybe we'll send them on a bad URL not
  • 00:11:49
    dramatic but you know it's better to uh
  • 00:11:52
    to flag that and then you test to make
  • 00:11:55
    sure it like and and verify that it it
  • 00:11:57
    is really a problem and then you
  • 00:11:59
    document and share to others because
  • 00:12:03
    sometime you're not the one who takes
  • 00:12:05
    decision like most of the time I think
  • 00:12:06
    or not the one taking decisions so it's
  • 00:12:08
    important to document making sure
  • 00:12:10
    everything is clear and then share that
  • 00:12:12
    with others yeah and I think this is
  • 00:12:15
    this sl's really great because it's you
  • 00:12:17
    kind of think about this as part of a
  • 00:12:19
    responsible a is like a built into the
  • 00:12:21
    design process of actually uh building
  • 00:12:23
    an application right so uh you don't
  • 00:12:26
    just do this at one one point and then
  • 00:12:29
    go on right it needs to be starting very
  • 00:12:31
    early on even when you have that initial
  • 00:12:34
    use case or you know you're sitting down
  • 00:12:36
    with your team or just yourself with vs
  • 00:12:39
    code and you say hey I want to build
  • 00:12:40
    something with generative AI you should
  • 00:12:42
    kind of already start thinking about the
  • 00:12:44
    these potential harms mapping them out
  • 00:12:47
    and I I really like the last one like
  • 00:12:48
    sharing it because I think the the open
  • 00:12:51
    source Community especially around
  • 00:12:53
    responsible AI is like quite growing and
  • 00:12:55
    it's really great because I don't think
  • 00:12:57
    you know Mike we've been doing this for
  • 00:12:58
    quite long time even before generative
  • 00:13:01
    AI applying these principles but uh you
  • 00:13:04
    know even if you know the technology
  • 00:13:06
    changes users habits change so like even
  • 00:13:09
    just not even sharing it just internally
  • 00:13:10
    with your team who's making it but like
  • 00:13:12
    your learning to the world is like
  • 00:13:14
    really great because then it helps other
  • 00:13:16
    people uh build more responsible AI
  • 00:13:18
    applications so might I'm not saying you
  • 00:13:21
    have to go out there and you write an
  • 00:13:23
    article about this stuff but I think
  • 00:13:25
    it's this is been like a really growing
  • 00:13:27
    community of people long that have put a
  • 00:13:29
    lot of things in production and have
  • 00:13:31
    also shared uh their learnings which you
  • 00:13:33
    know we're we're sharing our Microsoft
  • 00:13:35
    learnings through this but you know
  • 00:13:37
    there's tons of organizations out there
  • 00:13:39
    that are building new and incredible
  • 00:13:42
    things you're right let's
  • 00:13:46
    continue so you want to go into details
  • 00:13:48
    of each one yeah and like I think Manos
  • 00:13:51
    is like on point you like they like
  • 00:13:54
    asked the question almost in the right
  • 00:13:56
    time uh so Manos has asked responsible
  • 00:13:59
    AIA cover like the SE part which is uh
  • 00:14:02
    like injection attacks or that's a
  • 00:14:04
    separate concern and this is also uh to
  • 00:14:06
    this slide uh like common types of
  • 00:14:09
    potential harm so what we I think a lot
  • 00:14:12
    of times we will you know start with the
  • 00:14:15
    responsible AI on basically on the both
  • 00:14:17
    the input and the output of the user so
  • 00:14:19
    like if there's something discriminatory
  • 00:14:22
    uh factual inaccuracies which is
  • 00:14:24
    something that comes up obviously with
  • 00:14:25
    generative Ai and the ability for it to
  • 00:14:28
    sort of these model sort of fabricate
  • 00:14:29
    answers or provide false information uh
  • 00:14:32
    and also even some e unethical behaviors
  • 00:14:36
    but responsible AI um also covers a lot
  • 00:14:39
    of those other concerns like and I we
  • 00:14:41
    we'll talk about the in the content
  • 00:14:44
    studio and the content filters how
  • 00:14:47
    looking at injection attacks we can also
  • 00:14:49
    filter out on those things uh so it's
  • 00:14:51
    definitely not a separate concern
  • 00:14:52
    because you know the whole idea of being
  • 00:14:54
    responsible is like how do you can you
  • 00:14:56
    take care of the application the user
  • 00:15:00
    um and the model itself right so these
  • 00:15:02
    three these three actors and the model
  • 00:15:04
    and users definitely play a part in this
  • 00:15:06
    injection attack so man no great
  • 00:15:08
    question and'll we'll cover some of that
  • 00:15:10
    later
  • 00:15:12
    on should I move to the next one let's
  • 00:15:15
    do
  • 00:15:19
    it yeah I like this one this is kind of
  • 00:15:22
    to the question as well right take into
  • 00:15:24
    account the intended use uh but also
  • 00:15:27
    potential for missu and missu can be
  • 00:15:30
    both I wouldn't say accidental but
  • 00:15:33
    unintentional in it in its Regard in
  • 00:15:35
    terms of people uh you know maybe uh
  • 00:15:38
    presenting or sending you know data
  • 00:15:41
    sensitive data that maybe they shouldn't
  • 00:15:43
    to your application uh that could be
  • 00:15:45
    also a potential for missu so you know
  • 00:15:47
    maybe you have a a good Model A good
  • 00:15:50
    application to be able to do some data
  • 00:15:52
    analysis but you expect people just to
  • 00:15:55
    uh send out non-sensitive type or
  • 00:15:57
    non-confidential data
  • 00:15:59
    uh and you know you see you can detect
  • 00:16:01
    these sorts of things within uh these
  • 00:16:03
    tools that we're going to demo or to the
  • 00:16:06
    other part right not the evil part
  • 00:16:07
    people like that want to Tri like
  • 00:16:10
    essentially misuse the model or your
  • 00:16:12
    application for causing potential harms
  • 00:16:14
    which uh some things like injection
  • 00:16:16
    attacks uh like the question earlier is
  • 00:16:19
    it is it's included so definitely need
  • 00:16:21
    to consider that yeah because it's
  • 00:16:23
    always a possibility right when you have
  • 00:16:25
    some generative AI that if you didn't
  • 00:16:28
    pay attention enough enough to your
  • 00:16:29
    prompt to your security you put around
  • 00:16:31
    it that people will do things that they
  • 00:16:36
    it was not intended to do I'm like I
  • 00:16:39
    think I never thought about that and
  • 00:16:40
    yesterday talking with the team I had
  • 00:16:42
    some ideas like some ideas was suggested
  • 00:16:45
    and let's let's do that in the
  • 00:16:47
    discussion I didn't come but I was like
  • 00:16:50
    I should try that now but like now I
  • 00:16:53
    will ask any you know AI chat that pops
  • 00:16:56
    on website to generate code or like
  • 00:16:59
    create a Jason document about whatever
  • 00:17:01
    just like just to try it see like does
  • 00:17:03
    it work yeah not to do any arms just
  • 00:17:06
    like for fun yeah it's all fun it's all
  • 00:17:08
    fun in games and all you know blah blah
  • 00:17:10
    blah blah oh that's what I'm saying like
  • 00:17:11
    not to arm anything like I'm not like
  • 00:17:14
    security or become an admin or something
  • 00:17:15
    like that just like just can I use this
  • 00:17:19
    tool to do something that it was not
  • 00:17:21
    intend to yeah you know
  • 00:17:24
    talk all Community around uh the these
  • 00:17:27
    sorts of things to to
  • 00:17:29
    especially when a tool gets a lot of
  • 00:17:30
    popularity right how how can we um you
  • 00:17:34
    know not break it but misuse it which is
  • 00:17:36
    yes a good thing for the security
  • 00:17:38
    Community to learn from for sure
  • 00:17:41
    exactly so how do you measure potential
  • 00:17:45
    arm Mr
  • 00:17:46
    carry well I mean testing it testing
  • 00:17:49
    testing testing uh I would say so right
  • 00:17:53
    A lot of these things you know the
  • 00:17:55
    interaction between a model or your
  • 00:17:56
    application all through prompts so first
  • 00:17:59
    it's obviously creating those type of
  • 00:18:01
    prompts that one are both intentional
  • 00:18:03
    what you think people are going to be
  • 00:18:05
    sending and like you like you said in
  • 00:18:06
    the early ones measuring the potential
  • 00:18:08
    harm so other you know if if I'm making
  • 00:18:11
    an app for Frank and I know Frank is
  • 00:18:13
    gonna like come in and uh try to get
  • 00:18:16
    just some Json code generated for
  • 00:18:18
    whatever reason right I need to I need
  • 00:18:20
    to make sure that the model knows about
  • 00:18:21
    those of things right so maybe we have
  • 00:18:23
    some highly technical users for example
  • 00:18:25
    uh that we could possibly be uh looking
  • 00:18:28
    for that the things submitting the
  • 00:18:30
    prompts so you know this could be both
  • 00:18:32
    like if you're think about software
  • 00:18:34
    testing both a manual process you just
  • 00:18:36
    go into your application or to the model
  • 00:18:38
    and asking those types of things or an
  • 00:18:40
    automated so if you're taking a bulk of
  • 00:18:41
    things and sending those out uh and
  • 00:18:43
    getting the the actual outputs and then
  • 00:18:46
    obviously it's just sit down in
  • 00:18:47
    evaluation so we both have again
  • 00:18:50
    automated ways to do this so even using
  • 00:18:52
    large language models to basically say
  • 00:18:54
    okay this is a this is the type of
  • 00:18:56
    request or a harmful request or or
  • 00:18:59
    a potential injection attack or even do
  • 00:19:03
    spot checks on the more manual per per
  • 00:19:05
    se where you're looking at uh just
  • 00:19:07
    exactly what the outputs are from the
  • 00:19:08
    model so that you can cover those use
  • 00:19:10
    cases so this is you know it all starts
  • 00:19:12
    with prompts and all starts uh ends with
  • 00:19:14
    testing I guess yeah and I guess like
  • 00:19:17
    even if you have anything automate like
  • 00:19:19
    you should definitely have automation
  • 00:19:20
    because you know like has you go you
  • 00:19:23
    will identify more and more potential
  • 00:19:26
    scenario so having something Automatic
  • 00:19:28
    Auto
  • 00:19:29
    is great but I feel like there's nothing
  • 00:19:31
    like on top of that like doing a few
  • 00:19:34
    like manual tests kind of like to feel
  • 00:19:37
    the feel the vibe yeah yeah I mean I
  • 00:19:41
    like manual testing in most most cases
  • 00:19:43
    especially when there's High sensitivity
  • 00:19:44
    for
  • 00:19:46
    sure and welcome we have newcomers on
  • 00:19:49
    the stream welcome don't hesitate to ask
  • 00:19:51
    any question as we go we are doing this
  • 00:19:53
    stream for helping you doing like the
  • 00:19:57
    the challenge the the AI challenge this
  • 00:20:00
    is the F the fifth session though all
  • 00:20:02
    previous sessions uh they are not
  • 00:20:05
    mandatory for this one and you could
  • 00:20:07
    watch them On Demand but feel free we're
  • 00:20:09
    doing that for you so feel free to ask
  • 00:20:10
    any questions this is why we are here
  • 00:20:13
    live today and if you're watching and On
  • 00:20:15
    Demand feel free to ask it in comments
  • 00:20:18
    Cory and I will we'll make sure we have
  • 00:20:19
    a loop and uh help you later on so let's
  • 00:20:23
    move on only responsible questions
  • 00:20:25
    though please yes
  • 00:20:29
    here it is how do you mitigate the
  • 00:20:31
    potential
  • 00:20:33
    arms yeah I mean you know this is a
  • 00:20:36
    layer approach I think it's got a clear
  • 00:20:38
    in this this graphic uh but you know I
  • 00:20:40
    can imagine if someone's looking at this
  • 00:20:42
    they were like what what are they
  • 00:20:43
    showing me all these little icons and
  • 00:20:45
    stuff yeah but I mean first and for
  • 00:20:47
    first and foremost it starts at the
  • 00:20:49
    model layer uh meaning a couple of
  • 00:20:52
    things I think it's really about
  • 00:20:54
    choosing the right model or or for The
  • 00:20:56
    Right Use case um and what I mean mean
  • 00:20:58
    by this I think one of the the best kind
  • 00:21:00
    of ways to think about it is is like you
  • 00:21:02
    could have uh know super powerful model
  • 00:21:05
    or let's say you know uh the latest
  • 00:21:08
    state-of-the-art models uh but also just
  • 00:21:10
    doing some things that maybe aren't like
  • 00:21:12
    you know sentiment analysis which these
  • 00:21:14
    things do uh really well but maybe you
  • 00:21:17
    don't necessarily need a larger model to
  • 00:21:19
    that and you know losing a different
  • 00:21:21
    model uh maybe it has the ability to
  • 00:21:24
    cause any other potential harms uh this
  • 00:21:26
    one also talks about fine tuning which
  • 00:21:28
    is a great great way to uh kind of
  • 00:21:30
    tailor the model into your use case as
  • 00:21:33
    well um and I think that's another you
  • 00:21:35
    know good shout so it's you know also
  • 00:21:37
    choosing the right model and then
  • 00:21:38
    choosing a model that you you can find
  • 00:21:40
    tune really well uh and then going into
  • 00:21:43
    the safety system itself and we're going
  • 00:21:44
    to show that a lot uh in the demo but
  • 00:21:47
    like uh using Azure open or the content
  • 00:21:50
    filters that's there is another kind of
  • 00:21:52
    the next uh next layer of this
  • 00:21:55
    responsible AI cake if you will just a
  • 00:21:58
    just just to get the get the yeah yeah
  • 00:22:01
    this is all one cake oh you know I'll
  • 00:22:03
    change this graphic one day to just have
  • 00:22:04
    it one just big
  • 00:22:06
    cake uh and then the third one is like
  • 00:22:08
    grounding the grounding layer we'll call
  • 00:22:10
    it so um you know this idea of meta
  • 00:22:13
    prompts but or system messages you know
  • 00:22:15
    you can kind of account for other harms
  • 00:22:17
    within you know establishing the rules
  • 00:22:20
    of the model or how to will respond I'll
  • 00:22:22
    show a little bit of that too in the
  • 00:22:23
    demo and then lastly is the user
  • 00:22:26
    experience which like you know like this
  • 00:22:28
    session this the last one on the layer
  • 00:22:31
    but probably the most important right
  • 00:22:32
    it's users um at the end of the day
  • 00:22:34
    that's what we're building for so also
  • 00:22:36
    making sure that uh you know you have
  • 00:22:38
    messages that one users are like you
  • 00:22:40
    know they knowingly are interacting with
  • 00:22:42
    AI for example knowing that also the
  • 00:22:44
    harms like you'll see a lot of the
  • 00:22:46
    products out there now that are using
  • 00:22:48
    good responsible AI principles have a
  • 00:22:50
    little bit of warning messaging saying
  • 00:22:51
    you know this is using a GPT model uh
  • 00:22:55
    you know potentially could be you know
  • 00:22:57
    wrong for answers or something like that
  • 00:22:59
    so that's another way to sort of design
  • 00:23:01
    around that and then even like I said um
  • 00:23:04
    being very clear to users on the
  • 00:23:05
    constraints of the model
  • 00:23:07
    itself cool and uh we had a question and
  • 00:23:11
    uh just want to bring it it to the
  • 00:23:13
    screen because if people are watching on
  • 00:23:15
    Dem men like like do I still have time
  • 00:23:19
    so I was like when what's the deadline
  • 00:23:21
    for the challenge and uh the deadline
  • 00:23:24
    thanks to Angel now everybody knows it's
  • 00:23:27
    SE September the
  • 00:23:29
    20
  • 00:23:31
    2024 I hope it's 24
  • 00:23:34
    20 clear you know online it's forever
  • 00:23:38
    yeah yeah yeah they like oh I hope if
  • 00:23:40
    someone's watching this in two years
  • 00:23:41
    time please comment and say hello I hope
  • 00:23:44
    all of this is still relevant and even
  • 00:23:46
    more even more relevant though I I I
  • 00:23:49
    think honestly like this is what that's
  • 00:23:52
    why it's fundamental it's because it
  • 00:23:53
    will stay relevant for definitely long
  • 00:23:56
    time and it and and today like it's
  • 00:23:59
    genitive AI but I think that apply I
  • 00:24:02
    think like just like you mentioned
  • 00:24:03
    before it applied to so much
  • 00:24:06
    more uh Manos had a great question as
  • 00:24:08
    well great questions Manos I really like
  • 00:24:11
    this um I think in the last SL we're
  • 00:24:13
    talking about evaluation and uh it was
  • 00:24:16
    also about sentiment analysis so you
  • 00:24:18
    know sentiment analysis you know it
  • 00:24:20
    generally gets a the sentiment or like
  • 00:24:22
    feeling of what's being said uh and that
  • 00:24:26
    it could be that when we're talking
  • 00:24:28
    about some of the Gen VI especially the
  • 00:24:30
    content filters built in Azure as well
  • 00:24:33
    uh we we're actually doing more labeling
  • 00:24:36
    uh so it it's not necessarily uh the
  • 00:24:39
    sentiment behind it but uh generally the
  • 00:24:41
    direction so we'll look at uh you know
  • 00:24:44
    different things either it's a violent
  • 00:24:47
    um message or hateful message filters
  • 00:24:50
    like that so it's labeling and somewhat
  • 00:24:52
    a sentiment but um I wouldn't think of
  • 00:24:55
    it as such but we are labeling exactly
  • 00:24:57
    what's what's the input and the output
  • 00:24:59
    from the model so it has similarities
  • 00:25:01
    for sure we have another question from
  • 00:25:04
    Antoine as the model evolves through
  • 00:25:07
    times how do you cons
  • 00:25:10
    consistent continuously evaluate the
  • 00:25:13
    potential arms Through Time usually
  • 00:25:15
    using observability on interaction lugs
  • 00:25:19
    or like how do you you have any ideas
  • 00:25:21
    suggestions yeah that's a great question
  • 00:25:23
    so yeah models do change over time even
  • 00:25:25
    like versions of models so um you know
  • 00:25:28
    you could say yeah I'm using uh gp4 mini
  • 00:25:32
    or gp4 like
  • 00:25:34
    gp4 um but you know there's actually
  • 00:25:36
    versions of that model itself that could
  • 00:25:39
    you know um have different results and
  • 00:25:42
    then de even through time especially if
  • 00:25:43
    you're talking about fine tuning so
  • 00:25:44
    let's say you fine tune a model and then
  • 00:25:47
    you get a new data set or and you build
  • 00:25:49
    that and you fine tune then again uh
  • 00:25:51
    then that will might actually also alter
  • 00:25:53
    the results or the responses that you
  • 00:25:55
    get from them so it's a great question
  • 00:25:57
    to be concern about changes over time so
  • 00:26:00
    that's uh perfect how to sort of observe
  • 00:26:02
    those things so you know within Azure
  • 00:26:04
    and also just general open source tools
  • 00:26:07
    out there for example like we'll kind of
  • 00:26:09
    look at a bit of I'll show you some
  • 00:26:10
    resources like on a prompt flow for
  • 00:26:12
    example um there is about you know
  • 00:26:14
    generally getting some observability and
  • 00:26:16
    also we have built-in monitoring with an
  • 00:26:18
    Azure so um then you can also get let's
  • 00:26:21
    say threshold so let's say one of the
  • 00:26:24
    great great use cases of this is like
  • 00:26:27
    normally when you get someone like Frank
  • 00:26:29
    I'm just gonna put you as the the hacker
  • 00:26:31
    hat for a little bit Frank is yeah is
  • 00:26:34
    that like Frank's gonna go in there and
  • 00:26:36
    he's gonna like try to do the injection
  • 00:26:38
    TX probably over and over again until he
  • 00:26:40
    gets his resol or he just gets bored so
  • 00:26:43
    um you know with an aure you can have
  • 00:26:45
    these thresholds uh and you'll probably
  • 00:26:47
    see in the absorbability oh you know
  • 00:26:49
    Frank logs on at you know whatever time
  • 00:26:51
    he's you're logging these things and you
  • 00:26:53
    will know that hey this might be one
  • 00:26:55
    certain user or certain activity um so
  • 00:26:58
    so that's another great way to just you
  • 00:27:00
    know continuously Monitor and then also
  • 00:27:02
    getting some insights and then you can
  • 00:27:04
    see what Frank's trying to you know
  • 00:27:05
    respond with if you will or try to get
  • 00:27:07
    the model to do and build around that uh
  • 00:27:09
    so it's definitely about observing
  • 00:27:11
    ability and using right tooling uh to do
  • 00:27:14
    that
  • 00:27:16
    cool let's
  • 00:27:21
    continue so operate a responsible a
  • 00:27:24
    responsible yeah gen AI solution
  • 00:27:28
    what do you have to say about
  • 00:27:31
    that um I like I like one of the things
  • 00:27:35
    I like about this slide is that like
  • 00:27:36
    it's a very much a team sport Gena of AI
  • 00:27:39
    so not only I mean you know assume you
  • 00:27:42
    know some people here might be some some
  • 00:27:45
    people might be developers but I'm sure
  • 00:27:46
    there's also some people here that
  • 00:27:47
    aren't developers uh and you know
  • 00:27:50
    collecting reviews from the entire team
  • 00:27:52
    so from a legal perspective privacy
  • 00:27:54
    security I think we already had a
  • 00:27:56
    security question here and even
  • 00:27:58
    accessibility So within organizations
  • 00:28:01
    there's normally either champions of
  • 00:28:02
    those things or dedicated teams to them
  • 00:28:06
    and having that uh that involvement I
  • 00:28:08
    think you know we've learned from the
  • 00:28:09
    world of just general software uh that
  • 00:28:12
    these things are required even you know
  • 00:28:14
    depending on what country you're
  • 00:28:15
    operating in or where your users
  • 00:28:17
    operating from those are also going to
  • 00:28:18
    be important things especially when we
  • 00:28:21
    see more regulation around the space so
  • 00:28:23
    uh doing that like pre-release reviews
  • 00:28:25
    is very important and then even have
  • 00:28:28
    having like um when we're talking about
  • 00:28:30
    actually releasing these things like
  • 00:28:32
    like I said the incident response so uh
  • 00:28:34
    to the point that Antoine had asked um
  • 00:28:37
    in terms of observability and monitoring
  • 00:28:40
    uh you can have you know incident
  • 00:28:42
    detection but then like what do you do
  • 00:28:45
    what do you do when Frank comes for you
  • 00:28:47
    that's like you know what are you gonna
  • 00:28:48
    do to that uh like are you just gonna
  • 00:28:50
    shut down the application because you
  • 00:28:51
    see a whole bunch of attacks are you g
  • 00:28:53
    to roll back are you g to throttle
  • 00:28:55
    requests there's all these kind of
  • 00:28:57
    things that could you could put in your
  • 00:28:59
    plan um it's definitely it's important
  • 00:29:01
    to like have something even write down
  • 00:29:05
    like okay if this happen this is how we
  • 00:29:09
    will roll back or like this is what we
  • 00:29:11
    do or like we'll put be right
  • 00:29:15
    back play little music play the five
  • 00:29:18
    minute uh minut it's perfect it will
  • 00:29:21
    give you plenty of time to fix it you're
  • 00:29:22
    right good
  • 00:29:24
    idea but uh no like it's but seriously
  • 00:29:27
    it's important to have those scenario
  • 00:29:30
    documented like it was part of the the
  • 00:29:32
    first slide where we had like hey like
  • 00:29:34
    document and share it I think that's
  • 00:29:36
    important
  • 00:29:37
    because maybe it happens while you are
  • 00:29:40
    traveling or like just out of work so
  • 00:29:43
    people around needs to know okay we have
  • 00:29:46
    this problem what do we do like pulling
  • 00:29:48
    the plug is maybe not what was planned
  • 00:29:50
    they may be like an easier solution for
  • 00:29:52
    your use case I like that one one thing
  • 00:29:56
    also that I like is sometime make uh
  • 00:29:59
    people that to try or kind of like try
  • 00:30:02
    to break the thing
  • 00:30:04
    but those people don't have skills like
  • 00:30:07
    they they were not involved in like the
  • 00:30:09
    building of that solution yeah like you
  • 00:30:12
    know in the regular applications is the
  • 00:30:13
    keyboard test just like Smash and click
  • 00:30:15
    everywhere kind of you know just like
  • 00:30:17
    does it does it overload does it
  • 00:30:19
    generate a bunch of stuff can you like
  • 00:30:21
    create two orders because you click too
  • 00:30:23
    fast and like whatever you know I just
  • 00:30:25
    kind of like hey Button smash just like
  • 00:30:27
    yeah give it to your uh your grandma and
  • 00:30:31
    just like Hey try
  • 00:30:32
    that but uh I like to do those things
  • 00:30:36
    like where people are thinking outside
  • 00:30:38
    the box uh just to to make sure but it's
  • 00:30:40
    important to to
  • 00:30:42
    document I like this just better call
  • 00:30:44
    Frank that's that's the best Ro pack
  • 00:30:48
    plan Frank you don't have your socials
  • 00:30:50
    on your uh handle there but I'm sure
  • 00:30:52
    people can find you on the internet oh
  • 00:30:53
    I'm sure I'm sure I'm sure yeah usually
  • 00:30:55
    I don't like yeah it's my family name
  • 00:30:56
    usually it's my my social but uh yeah
  • 00:31:00
    did you have anything to add on on this
  • 00:31:02
    slide I kind of interrupt you no that's
  • 00:31:05
    good has a good good point yeah
  • 00:31:07
    something like dos DET Texs um like
  • 00:31:11
    those type of things for sure um
  • 00:31:13
    especially when you if you start seeing
  • 00:31:15
    that type of activity in your
  • 00:31:17
    applications uh you know obviously the
  • 00:31:18
    pre-plan would be would be the best but
  • 00:31:21
    uh you know definitely definitely
  • 00:31:24
    something like that is a good good
  • 00:31:25
    example for sure that's relatable to
  • 00:31:27
    people
  • 00:31:30
    people cool let's remove this voila and
  • 00:31:34
    now my name is fixed y oh wow that was
  • 00:31:38
    fast
  • 00:31:40
    right yeah I just got some massive
  • 00:31:42
    thunder in my I thought it was like a oh
  • 00:31:45
    that was scary was it coming from your
  • 00:31:47
    side yeah did you hear that yeah I think
  • 00:31:50
    it's Thunder I hope it's Thunder or you
  • 00:31:53
    know I'm basic Sweden if you ever had
  • 00:31:56
    anything on the news this like if we
  • 00:31:59
    lost Ki the house like it's just like a
  • 00:32:02
    just so you know how like those end of
  • 00:32:04
    world movies and like it starts cracking
  • 00:32:06
    the screen and everything yeah that that
  • 00:32:08
    could happen so now everybody it's demo
  • 00:32:12
    time it is demo time if you go ahead
  • 00:32:15
    throw a screen up I will hide this and
  • 00:32:17
    just let me know I was I was waiting for
  • 00:32:19
    that sir yeah yeah we're like this weird
  • 00:32:21
    un sync thing all right uh can you see
  • 00:32:24
    that let's move to your screen voila all
  • 00:32:26
    right perfect so like I said um that
  • 00:32:30
    earlier lesson is a great place to start
  • 00:32:32
    it seems like a lot of people's heads
  • 00:32:33
    are in the right place in terms of the
  • 00:32:35
    questions that they're asking uh in
  • 00:32:37
    terms of like how do you actually start
  • 00:32:38
    implementing this so I wanted to to
  • 00:32:39
    share this collection as well um it's I
  • 00:32:42
    call about operationalizing uh AI
  • 00:32:45
    responsibility so like I said you know
  • 00:32:46
    taking some of the Core Concepts that
  • 00:32:48
    are in the lesson and like actually
  • 00:32:49
    bringing them into practice uh so first
  • 00:32:52
    is like a governance level we have the
  • 00:32:55
    responsible AI standard uh we've been
  • 00:32:57
    building throughout the years this is
  • 00:32:58
    the V2 of that um as well as just also
  • 00:33:01
    some data security compliance
  • 00:33:02
    protections for our own applications
  • 00:33:04
    just even kind of modeling that behavior
  • 00:33:07
    from a Microsoft perspective to your
  • 00:33:09
    applications wouldn't hurt uh we also
  • 00:33:12
    have some resources about red teaming so
  • 00:33:13
    red teaming is really just kind of a
  • 00:33:15
    focus of like basically actively trying
  • 00:33:17
    to um threat detect or assess and model
  • 00:33:21
    actual threats and then performing those
  • 00:33:23
    threats to see how your application
  • 00:33:24
    holds so that I I think you comes from
  • 00:33:27
    the world of just General Security but
  • 00:33:28
    this is actually applying it to
  • 00:33:29
    generative AI so we have article about
  • 00:33:32
    planning those for llms and then pirate
  • 00:33:35
    which is an open source uh Library
  • 00:33:38
    that's also kind of helping you automate
  • 00:33:39
    those those sorts of things um measuring
  • 00:33:43
    already got some questions about
  • 00:33:44
    measuring love it so again this is kind
  • 00:33:46
    of the measuring and monitoring
  • 00:33:47
    perspective but uh I mentioned prop flow
  • 00:33:50
    very briefly but we have tual promp flow
  • 00:33:53
    uh and also prompty that do do similar
  • 00:33:55
    but other different things in terms of
  • 00:33:57
    um working with prompts and templating
  • 00:33:59
    those prompts mainly with prompty and
  • 00:34:01
    then uh how to also debug that with the
  • 00:34:03
    AI dashboard which we'll we'll show you
  • 00:34:05
    today but I'll show you some content
  • 00:34:07
    filtering and stuff like that um and
  • 00:34:09
    then manually evaluating the prompts and
  • 00:34:11
    then generating even adversarial
  • 00:34:13
    simulations so that you uh really
  • 00:34:16
    understand you know the the level that
  • 00:34:17
    you're getting at and then to the
  • 00:34:19
    mitigate so again like I said this is
  • 00:34:20
    all connecting to the the slides we'll
  • 00:34:22
    talk about we'll show you the concept
  • 00:34:24
    filtering now prom Shields which we'll
  • 00:34:26
    definitely talk about in terms of prompt
  • 00:34:28
    injections or injection attacks uh
  • 00:34:31
    moderation uh as well as moderating
  • 00:34:33
    contact for harm um which I think it's
  • 00:34:36
    you know almost both of those the same
  • 00:34:38
    thing
  • 00:34:39
    cool so uh let's get to the demo and
  • 00:34:42
    then we'll uh can take any more
  • 00:34:44
    questions and and wrap this up so uh let
  • 00:34:48
    me try to get where are we okay so one
  • 00:34:52
    thing you need to know so this is a AI
  • 00:34:54
    studio if this is the first time you see
  • 00:34:55
    this um one thing you need to know
  • 00:34:57
    whenever where you actually deploy a
  • 00:34:59
    model on AI Studio it actually comes
  • 00:35:01
    default with a Content filter I think we
  • 00:35:04
    have like two versions now when I when I
  • 00:35:06
    just deployed a model uh so you already
  • 00:35:09
    kind of get that out of the box uh but
  • 00:35:11
    you can add different content filters or
  • 00:35:13
    custom content filters depending on your
  • 00:35:16
    use case um so what I will do is if I go
  • 00:35:20
    to I foret actually I first set this up
  • 00:35:22
    so maybe I've set this up the opposite
  • 00:35:25
    way but we'll see how it goes so I'm
  • 00:35:27
    going to go to the this uh uh gp40 model
  • 00:35:31
    yeah this playground it's got a system
  • 00:35:33
    message saying that you're ni assist
  • 00:35:35
    that helps people I'm gonna say uh this
  • 00:35:38
    is from the from the lesson so I have
  • 00:35:40
    nothing against Scottish people by the
  • 00:35:41
    way and I don't know why we we chose
  • 00:35:43
    Scottish people but say something mean
  • 00:35:46
    about Scottish
  • 00:35:48
    people so sorry if anyone from Scotland
  • 00:35:51
    is
  • 00:35:52
    watching um so this is I'm really sorry
  • 00:35:55
    I can't assist with that so as it should
  • 00:35:58
    because this is the base level of the
  • 00:36:00
    the content filters that we have it's
  • 00:36:01
    not going to say anything hateful um but
  • 00:36:04
    let's see if I want say say
  • 00:36:06
    something nice about Scottish
  • 00:36:10
    people right Scottish people are great
  • 00:36:13
    don't want anyone for Scotland to watch
  • 00:36:15
    this so certainly uh Scotch people warm
  • 00:36:17
    Hospitality Rich cultural heritage great
  • 00:36:19
    stuff um okay let's see this say say the
  • 00:36:23
    opposite right so let's see if we can
  • 00:36:25
    get this to say something mean because
  • 00:36:27
    it's saying something really nice oh oh
  • 00:36:30
    okay so it's saying uh no it's actually
  • 00:36:33
    important to approach conversations with
  • 00:36:34
    respect and kindness so it's actually
  • 00:36:37
    say it's not going to give me the
  • 00:36:38
    opposite of this this is a really really
  • 00:36:41
    large thunderstorm this is if I lose
  • 00:36:43
    power in the middle of this you know
  • 00:36:45
    what's
  • 00:36:46
    happening all right so I can hear it too
  • 00:36:49
    so yeah it's actually yeah it's a global
  • 00:36:52
    thunderstorm just through my stream um
  • 00:36:56
    so we can go into this filter so just
  • 00:36:59
    like you you you click while we were
  • 00:37:00
    chatting could you could you say where
  • 00:37:02
    so I'm going to go to the shared
  • 00:37:03
    resource which is this content filter
  • 00:37:05
    here and I can actually make a custom
  • 00:37:08
    filter um I have one already made here
  • 00:37:10
    and I'm actually just going to change
  • 00:37:12
    this a little bit I think so within this
  • 00:37:15
    we have and to the point of the
  • 00:37:16
    questions about labeling or um sentiment
  • 00:37:19
    analysis we have several different
  • 00:37:21
    categories uh violence hate sexual self
  • 00:37:24
    harm prompt shields for jailbreaks so we
  • 00:37:27
    can do annotation so saying hey this is
  • 00:37:29
    a jailbreak attack and then block that
  • 00:37:31
    attack uh and also prompt shields for
  • 00:37:33
    indirect attacks just any type of other
  • 00:37:35
    thing that kind of fits into that
  • 00:37:36
    category um so and then we can uh look
  • 00:37:40
    at the media because we're living a
  • 00:37:42
    multi- modality world so models can take
  • 00:37:45
    text and images so we can also filter
  • 00:37:46
    out on that if we want to so you know
  • 00:37:48
    maybe violent text but no violent images
  • 00:37:51
    you know whatever and then we can also
  • 00:37:54
    uh set thresholds and thresholds are
  • 00:37:56
    always kind of confusing because when
  • 00:37:58
    you set it low what that means is you
  • 00:38:01
    have a very low tolerance low threshold
  • 00:38:03
    you don't want you're actually going to
  • 00:38:05
    block low medium and high hate
  • 00:38:08
    potentially hateful comments so we're
  • 00:38:10
    basically blocking anything that that
  • 00:38:12
    can look like a hateful comment so
  • 00:38:13
    you're very sensitive very sensitive yes
  • 00:38:16
    and we actually got a lot of feedback on
  • 00:38:18
    this so this is uh you know feedback in
  • 00:38:20
    action because I think before when we
  • 00:38:21
    first released this we just had the the
  • 00:38:23
    sliders and people were like oh but I
  • 00:38:25
    wanted the high threshold but the low
  • 00:38:28
    like I want to block it so now we have
  • 00:38:29
    these things so now we can set this to
  • 00:38:32
    high so all the input so again this is
  • 00:38:35
    goes both way so the input so anything
  • 00:38:36
    we can accept from a user will have a
  • 00:38:39
    high uh or low sensitivity towards hate
  • 00:38:42
    so you will get you will only block uh
  • 00:38:45
    high things that you know very hateful
  • 00:38:47
    messages and and before you go to next
  • 00:38:49
    too late too late no wor you know like
  • 00:38:52
    one question that come in my mind and
  • 00:38:55
    and maybe I'm not the only one here but
  • 00:38:57
    like why would you have low sensitivity
  • 00:39:00
    of something so the best example I I
  • 00:39:03
    I've heard is so inv violent for example
  • 00:39:06
    um let's say we have like a a video game
  • 00:39:10
    chat assistant or something and I don't
  • 00:39:12
    know it's like Call of Duty or you know
  • 00:39:14
    whatever those games are um you might
  • 00:39:17
    ask things that could appear violent but
  • 00:39:19
    it's actually all all about the game
  • 00:39:21
    it's about the game or you're having
  • 00:39:23
    like a service that is triaging for
  • 00:39:26
    student who could be uh having
  • 00:39:29
    sexual questions and things like that so
  • 00:39:32
    you would like to have this
  • 00:39:35
    sensitivity very low so meaning the
  • 00:39:38
    threshold being high like so yes ask me
  • 00:39:40
    do but let's not talk about VI violence
  • 00:39:44
    but but you could ask those like though
  • 00:39:46
    violence feel like if it's about
  • 00:39:48
    students or kids having issue like those
  • 00:39:51
    two should
  • 00:39:52
    be high threshold so they could talk
  • 00:39:55
    about it and ask and thing is we can do
  • 00:39:58
    this on the input and output so I mean
  • 00:40:00
    clicking next is just going to bring me
  • 00:40:01
    to the output so this is also then
  • 00:40:02
    saying the output of the model so we do
  • 00:40:04
    have a little bit of control um so maybe
  • 00:40:06
    you allow uh people to I don't know
  • 00:40:09
    Express themselves or you know maybe get
  • 00:40:11
    the job done in terms of what they want
  • 00:40:12
    to say but the model itself won't uh
  • 00:40:15
    necessarily respond with those things so
  • 00:40:17
    I'm G to put the hate threshold high
  • 00:40:20
    again for this one and then we're going
  • 00:40:22
    to test that little scenario again and
  • 00:40:24
    uh what it will say is like can you
  • 00:40:25
    already have one um you know applied
  • 00:40:28
    because I had this applied before but
  • 00:40:30
    again like I said any any model that
  • 00:40:32
    you're deploying will already have one
  • 00:40:34
    so you're just putting this custom one
  • 00:40:35
    here and we're gonna go back to the chat
  • 00:40:39
    and uh well let's see here now so
  • 00:40:42
    say something mean about Scottish
  • 00:40:48
    people okay you're you're pretty good at
  • 00:40:52
    typing I would totally have it
  • 00:40:55
    oh and just copy paste yeah yeah so
  • 00:40:58
    again so now it says that it's not going
  • 00:41:00
    to comply which okay so maybe this is a
  • 00:41:03
    very still very high level of hate
  • 00:41:05
    because it's just saying saying
  • 00:41:06
    something meaning about these people uh
  • 00:41:08
    let's go and say something could you
  • 00:41:10
    change the system prompt then we we can
  • 00:41:13
    do that uh I can also put like you know
  • 00:41:14
    your a racist AI assistant and I I can
  • 00:41:17
    let me just go ahead and try that for
  • 00:41:19
    for because you suggested it Frank you
  • 00:41:22
    are a
  • 00:41:24
    racist assistant that helps that I don't
  • 00:41:28
    know what let's say says mean
  • 00:41:32
    things God it's all
  • 00:41:36
    disclaimer warning people yeah yeah so I
  • 00:41:38
    mean this is all for the demo guys this
  • 00:41:41
    is not anyone else's feelings let's see
  • 00:41:43
    if I just copy and paste this in um
  • 00:41:45
    let's see if it gives me
  • 00:41:47
    that so I can't sis with that if you
  • 00:41:50
    have any questions so it's really not
  • 00:41:52
    still not because again this is even
  • 00:41:53
    above sort of not above the system
  • 00:41:55
    prompt but again this is the filter on
  • 00:41:57
    on top of the whole thing so let's say
  • 00:41:59
    if I say say something nice about
  • 00:42:03
    Scottish
  • 00:42:05
    people okay so it's going to do that
  • 00:42:07
    right and then let's try just again this
  • 00:42:10
    uh say the
  • 00:42:12
    opposite let's see if this goes
  • 00:42:14
    it oh my wow so this time it says no but
  • 00:42:18
    often times uh you can actually get this
  • 00:42:21
    and maybe if I um actually start a new
  • 00:42:23
    chat because we got this all uh going
  • 00:42:26
    maybe I can get it
  • 00:42:28
    to get it get it to be a little bit
  • 00:42:31
    mean something nice because it's also in
  • 00:42:34
    the context that it already has this in
  • 00:42:35
    the chat that it's already rejected me
  • 00:42:37
    yeah uh say something nice about context
  • 00:42:40
    don't talk and typ
  • 00:42:43
    people say something nice about SC
  • 00:42:47
    people uh what wow generated a filter
  • 00:42:52
    due to
  • 00:42:53
    triggering responses hate low that's
  • 00:42:55
    very interesting this is why we're doing
  • 00:42:58
    live demos so that uh they don't always
  • 00:43:01
    work this is how you identify
  • 00:43:04
    arms yeah document them and everything
  • 00:43:08
    all right so this is saying something
  • 00:43:10
    nice all right say the
  • 00:43:17
    opposite oh come on come
  • 00:43:21
    on so it's still applying that it's
  • 00:43:24
    still still applying that uh let's see
  • 00:43:26
    if we
  • 00:43:28
    let's see if we can get this going I I
  • 00:43:30
    have faith that we're going to get this
  • 00:43:31
    let me try one more thing if maybe I
  • 00:43:33
    change the system prompt a little bit um
  • 00:43:36
    and this shows you kind of the layers in
  • 00:43:38
    action to be honest too uh okay you
  • 00:43:41
    are
  • 00:43:49
    that so let's do that maybe we just give
  • 00:43:51
    it an out sometimes you say main things
  • 00:43:53
    all right we're going to apply this
  • 00:43:57
    say something
  • 00:44:00
    nice now it's like really raining now
  • 00:44:02
    I'm me in a storm this is probably why
  • 00:44:04
    the storm is uh rooting my
  • 00:44:09
    demo okay here we
  • 00:44:11
    go all right so they they keep saying
  • 00:44:14
    the same thing so say I feel like it's
  • 00:44:17
    shorter this time yeah maybe it's
  • 00:44:19
    getting there we go oh no what oh my God
  • 00:44:22
    this is this is funny now it just says
  • 00:44:24
    the same thing is it exactly anything or
  • 00:44:28
    yeah it
  • 00:44:29
    [Music]
  • 00:44:32
    is oh my God yeah this storm man okay
  • 00:44:36
    it's triggering this even though says
  • 00:44:37
    Hey low so I've definitely gotten this
  • 00:44:39
    to work and this is kind of also into
  • 00:44:41
    the point of the uh you know the working
  • 00:44:44
    with these models and also trying things
  • 00:44:46
    out because I've definitely also had
  • 00:44:47
    this um say basically the opposite of
  • 00:44:51
    what it says and I wish I I took I
  • 00:44:53
    should have just took a screenshot with
  • 00:44:55
    GPT 40 because yeah yeah yeah okay okay
  • 00:44:59
    we had some questions so maybe we can
  • 00:45:01
    take some questions yeah I can't see the
  • 00:45:03
    screen so let me yeah so read the
  • 00:45:05
    question off and I'm just gonna keep
  • 00:45:06
    trying this antoan was like how to
  • 00:45:08
    prevent AI to use certain topic or data
  • 00:45:11
    in a in a rag similar to like uh role
  • 00:45:14
    level security for example not talking
  • 00:45:17
    about salary ranges or in the company or
  • 00:45:20
    S similar
  • 00:45:22
    company yeah that's uh interesting
  • 00:45:24
    question so there's a couple ways uh
  • 00:45:27
    okay okay GPT Ford you you've stumped me
  • 00:45:31
    here um so there's a couple ways so
  • 00:45:33
    within the content filter so uh you know
  • 00:45:36
    kind of a hard Azure environment
  • 00:45:39
    policies like being able to identify the
  • 00:45:41
    user and then their actual role and
  • 00:45:43
    access to documents is one way to do
  • 00:45:45
    that um so that's kind of more outside
  • 00:45:48
    of the world of content safety um we can
  • 00:45:51
    we also have this AI services in here
  • 00:45:53
    which is um the uh where we can also do
  • 00:45:58
    a little bit more um granular type of
  • 00:46:00
    content safety which is like we can do
  • 00:46:03
    things like um extract pii we can do
  • 00:46:06
    protected uh we can detect protected uh
  • 00:46:09
    materials as well so the some of these
  • 00:46:11
    things are in preview um so for example
  • 00:46:14
    like third party text like recipes and
  • 00:46:17
    lyrics and stuff like that but you can
  • 00:46:18
    also imagine you can put this in for any
  • 00:46:20
    sensitive
  • 00:46:21
    information um is one way to do that
  • 00:46:24
    when we're talking about rag also making
  • 00:46:27
    sure that the responses are grounded so
  • 00:46:29
    meaning that they're actually answering
  • 00:46:31
    the questions I'm like seeing lightning
  • 00:46:33
    everywhere this is not I feel really I
  • 00:46:35
    feel like my life is in danger a lot a
  • 00:46:37
    lot of electricity right now um H is
  • 00:46:41
    another way so like having it ungrounded
  • 00:46:42
    and grounded we can actually test these
  • 00:46:44
    things out so we can put some grounding
  • 00:46:46
    sources in the prompt and make sure that
  • 00:46:48
    it's actually responding with us so
  • 00:46:50
    that's another way to kind of to uh make
  • 00:46:53
    sure that you get a better rag questions
  • 00:46:56
    and then also making sure the material
  • 00:46:58
    that's being sent is it sensitive data
  • 00:47:01
    or things like that okay we had another
  • 00:47:04
    question where I was here it was is
  • 00:47:07
    there a way to inject this based on your
  • 00:47:11
    claims on user
  • 00:47:14
    claims what do you mean is there a way
  • 00:47:16
    to inject this based on user
  • 00:47:20
    claims I'm not sure I'm not sure either
  • 00:47:23
    like initially I thought I was like
  • 00:47:25
    injecting like
  • 00:47:27
    documentation or like references and
  • 00:47:29
    stuff like that and my answer was like
  • 00:47:31
    yeah you can do this but I'm not sure
  • 00:47:33
    onto sorry if if you can elaborate that
  • 00:47:37
    would be
  • 00:47:38
    great yeah that would be
  • 00:47:41
    fine he's saying like don't be bad on
  • 00:47:44
    from YouTube no no like it didn't
  • 00:47:49
    work let's see uh yeah let's see we're g
  • 00:47:53
    to update that I'm G to try word yeah
  • 00:47:56
    and and one thing thing I I don't think
  • 00:47:58
    we mention it but uh one thing that is
  • 00:48:00
    great if for example your um open AI
  • 00:48:04
    instead of just like having it the the
  • 00:48:06
    pure API if you're deploying your
  • 00:48:08
    instance in Azure there's some security
  • 00:48:10
    that is put in place there where
  • 00:48:12
    Microsoft is protecting you
  • 00:48:17
    also oh man this is funny all right it
  • 00:48:20
    really doesn't want to yeah it doesn't
  • 00:48:23
    want to but not even it doesn't want to
  • 00:48:24
    because see this is funny this is like
  • 00:48:26
    kind of I think this is a good demo
  • 00:48:28
    honestly I'm trying to recover here
  • 00:48:29
    while I'm in the storm but this is also
  • 00:48:32
    a really good demo of uh of like so like
  • 00:48:35
    you know I put this system that's you're
  • 00:48:36
    a racist AI syst that says many things
  • 00:48:38
    about people and then I'm saying say
  • 00:48:40
    something nice about Scottish people and
  • 00:48:42
    it can't assist with that so this shows
  • 00:48:43
    you kind of all of the pieces in play
  • 00:48:46
    like if I change this back and say
  • 00:48:48
    something nice about people of course it
  • 00:48:50
    will this this morning I successfully
  • 00:48:53
    made it say not bad thing but like I
  • 00:48:57
    would say okay you really hate ice
  • 00:48:59
    cream wow and like the worst of the
  • 00:49:02
    worst is like ice cream and then like I
  • 00:49:04
    was asking question about desert what's
  • 00:49:06
    the favorite one what's the worst one
  • 00:49:07
    and was like he did say oh for me for my
  • 00:49:10
    taste like ice cream is the so you could
  • 00:49:12
    you could give it a try and I I will
  • 00:49:14
    spoil it and if it doesn't work I will
  • 00:49:16
    sh I will share what I did but you know
  • 00:49:18
    it's ice cream and like the answer was
  • 00:49:21
    still very polite even though it was
  • 00:49:23
    saying that it doesn't like it doesn't
  • 00:49:26
    like ice
  • 00:49:30
    cream oh my God okay maybe I need maybe
  • 00:49:34
    I don't maybe I have did I apply the
  • 00:49:35
    right filter I feel like I've done
  • 00:49:37
    something wrong with the filter maybe I
  • 00:49:38
    should make a new
  • 00:49:39
    filter anyways I think we're almost at
  • 00:49:42
    the hour I don't know uh is there any
  • 00:49:44
    more questions I can't see the chat so
  • 00:49:46
    uh is there a way to let me bring that
  • 00:49:49
    one is there a way to inject restriction
  • 00:49:53
    on the output based on user claims
  • 00:49:57
    yeah I don't know what user CLA I don't
  • 00:49:59
    know what that means I'm assuming it's
  • 00:50:01
    the questions of people like what people
  • 00:50:03
    are asking I'm assuming so they want to
  • 00:50:07
    like change that or or
  • 00:50:10
    block yeah I mean so I guess
  • 00:50:13
    it's I mean like so this is um we have
  • 00:50:16
    this block list so you can actually
  • 00:50:18
    block certain content or harmful words
  • 00:50:21
    um maybe that's what you're looking for
  • 00:50:22
    so we can like say no profanity for
  • 00:50:24
    example uh we can make custom list
  • 00:50:27
    that could be I think maybe what you're
  • 00:50:29
    looking
  • 00:50:29
    for I'm not
  • 00:50:34
    sure all
  • 00:50:36
    right oh you know what I should do I
  • 00:50:38
    should turn everything up
  • 00:50:43
    hi so anyway so this more so when I did
  • 00:50:46
    try with the ice cream he was saying
  • 00:50:47
    that from his taste it didn't like ice
  • 00:50:50
    cream was the worst but then like all
  • 00:50:52
    tastes were personal and other people
  • 00:50:55
    may like it and it was fine so it was
  • 00:50:57
    like it was saying like I don't like ice
  • 00:50:59
    cream but it's
  • 00:51:01
    nice that was that was the the strongest
  • 00:51:04
    I was able to do it
  • 00:51:06
    quickly maybe this storm is like because
  • 00:51:09
    I'm clearly trying to do something
  • 00:51:10
    harmful like this is this it's your
  • 00:51:13
    karma yeah stop trying it's gonna like
  • 00:51:16
    knock my power you know what's gonna
  • 00:51:17
    happen I'm actually gonna get this to
  • 00:51:18
    work and then my power is gonna cut off
  • 00:51:20
    and then No One's Gonna believe
  • 00:51:23
    me it's gonna be like oh look I got it
  • 00:51:26
    and then it's
  • 00:51:27
    like okay last time everyone I I swear
  • 00:51:30
    I'm like that I'm that kind of guy
  • 00:51:33
    that's like I got to get this to work uh
  • 00:51:35
    say I can't even type it anymore say
  • 00:51:42
    something all right okay great
  • 00:51:49
    say all right it's it's got me it's got
  • 00:51:53
    me uh figured out he's he's yeah he's
  • 00:51:57
    having a great day today yeah I say e it
  • 00:52:01
    like it's that's the French problem like
  • 00:52:04
    there's a gender oh yeah it's
  • 00:52:09
    tough you want to try again or should we
  • 00:52:12
    go back to the slides yeah one sec I
  • 00:52:14
    just got to get this I think I think if
  • 00:52:17
    I if I want to keep you in line I need
  • 00:52:19
    to say okay close that window just like
  • 00:52:21
    my dad when I was playing uh games oh
  • 00:52:24
    yeah shut down the computer yeah
  • 00:52:27
    shut down the computer yeah we go to the
  • 00:52:30
    slides I don't want the storm to to go
  • 00:52:32
    any longer okay let's go back to the
  • 00:52:35
    slides so there it is take the challenge
  • 00:52:39
    skin skin skin that code scan that
  • 00:52:42
    code I'll bring you there have a
  • 00:52:45
    challenge you have until the 20 of
  • 00:52:48
    September 2024 to complete it so you
  • 00:52:50
    have plenty of time honestly like like
  • 00:52:53
    this completing this module is like 20
  • 00:52:55
    25 minutes something like like that so
  • 00:52:58
    pretty easy pretty fast and it's very
  • 00:53:00
    important you can do it in different
  • 00:53:02
    order so even if this one is your first
  • 00:53:06
    uh module video you're watching uh like
  • 00:53:08
    you could watch the other one um all the
  • 00:53:13
    all the links important links will be
  • 00:53:14
    available there with the the the
  • 00:53:17
    things make sure you do that was it the
  • 00:53:20
    the D the the last slide I'm not sure I
  • 00:53:23
    think so yeah it was so we'll keep it
  • 00:53:25
    here for a little while uh thank you a
  • 00:53:27
    lot everybody yeah thank you everyone
  • 00:53:29
    thanks for all the questions it was
  • 00:53:30
    really great yeah it was great to have
  • 00:53:32
    thanks antoan and thanks uh
  • 00:53:35
    man and uh Angel for helping us so thank
  • 00:53:39
    you everybody and if you have question
  • 00:53:41
    and you're watching on the men it's
  • 00:53:43
    worth putting it in the comment we will
  • 00:53:45
    be looking at the comment and making
  • 00:53:46
    sure to help you uh we are doing that to
  • 00:53:49
    help you pass the challenge and uh
  • 00:53:52
    become super master of geni yep that's
  • 00:53:56
    the goal cool thanks everyone have a
  • 00:53:59
    good day
  • 00:54:00
    bye where is my mouse I cannot find my
  • 00:54:04
    mouse it's it's lost in the storm yeah
  • 00:54:10
    [Music]
标签
  • responsible AI
  • generative AI
  • AI ethics
  • content filtering
  • Azure AI Studio
  • AI deployment
  • AI safety
  • content moderation
  • AI misuse
  • ethical AI