1 00:00:04,720 --> 00:00:08,590 In this lesson we are going to talk about responsible AI. 2 00:00:15,560 --> 00:00:17,420 So responsible. 3 00:00:17,420 --> 00:00:26,060 AI is a very new and down, uh, discipline which, uh, tries to uh. 4 00:00:27,510 --> 00:00:38,640 Include the principles that you should follow in order to have an acceptable, socially acceptable, 5 00:00:38,760 --> 00:00:41,520 uh, AI application. 6 00:00:41,520 --> 00:00:51,750 So I think it may be interesting to review, at least very quickly, these main principles that right 7 00:00:51,750 --> 00:00:55,080 now are included in responsible AI. 8 00:00:55,110 --> 00:01:00,390 The first one is fairness and eliminating bias. 9 00:01:01,440 --> 00:01:06,720 This principle emphasizes the importance of developing AI systems. 10 00:01:06,750 --> 00:01:16,920 AI applications that treat all users fairly regardless of their race, gender, age, or other characteristics. 11 00:01:16,980 --> 00:01:26,790 It involves eliminating bias in data and algorithms to prevent discrimination and ensure that AI decisions 12 00:01:26,790 --> 00:01:29,250 are fair and equitable. 13 00:01:29,370 --> 00:01:39,270 So remember, after immediately after the launch of ChatGPT, we started to see a news about how different 14 00:01:39,270 --> 00:01:41,610 the ChatGPT responses were. 15 00:01:41,610 --> 00:01:46,260 If the user asks about a the. 16 00:01:47,470 --> 00:01:57,340 Perfect job for a man or for a woman, or the perfect candidate for a software engineer, a role in 17 00:01:57,340 --> 00:01:59,110 terms of race or whatever. 18 00:01:59,110 --> 00:02:12,070 So we saw that there was a huge polemic, you know, a and scandal, if you want, around the responses 19 00:02:12,070 --> 00:02:19,000 that, uh, you know, the most popular, uh, artificial intelligence application was providing and 20 00:02:19,000 --> 00:02:26,950 this, this, this talk to a lot of changes in the way ChatGPT was, uh, uh, managed and handled, 21 00:02:26,950 --> 00:02:27,490 etc.. 22 00:02:27,490 --> 00:02:31,120 So this first principle is, is very important. 23 00:02:31,120 --> 00:02:37,090 And as you can imagine, it can impact in the reputation of a company. 24 00:02:37,150 --> 00:02:42,250 Uh, and uh, even in the, in the legal future of, of one company. 25 00:02:42,250 --> 00:02:42,790 Right. 26 00:02:42,940 --> 00:02:47,050 Second principle, reliability and safety. 27 00:02:48,260 --> 00:02:56,660 Reliability refers to the ability of the LM applications to function consistently and predictably. 28 00:02:57,170 --> 00:02:58,490 Safety. 29 00:02:58,520 --> 00:03:05,900 Safety involves protecting AI applications from malicious manipulations and ensuring they do not harm 30 00:03:05,900 --> 00:03:08,210 users or the environment. 31 00:03:08,360 --> 00:03:14,270 This includes implementing robust measures to prevent errors and failures. 32 00:03:14,270 --> 00:03:21,680 So reliability, as we have been talking about before in the program, is one of the most important 33 00:03:21,680 --> 00:03:27,170 things, uh, to consider when you are preparing an LM application. 34 00:03:27,620 --> 00:03:34,250 Remember hallucinations and all these problems and how, for example, in the second sites project, 35 00:03:34,250 --> 00:03:39,200 we were, uh, working with this with this problem, right? 36 00:03:39,260 --> 00:03:46,760 Uh, you know, uh, adding the sources, a link to, to the sources we were using to provide any response, 37 00:03:46,760 --> 00:03:47,540 etc., etc.. 38 00:03:47,540 --> 00:03:47,960 Right. 39 00:03:47,960 --> 00:03:53,420 So reliability is very important and safety as well. 40 00:03:53,750 --> 00:04:01,400 And safety is talking about, you know, uh, malicious intents, uh, manipulations etc., etc.. 41 00:04:01,400 --> 00:04:08,630 So we need to check our LM application before we launch it. 42 00:04:08,660 --> 00:04:15,590 Uh, in search of problems like this one's reliability very important. 43 00:04:15,590 --> 00:04:17,810 And safety as well. 44 00:04:17,930 --> 00:04:21,769 Third principle privacy and data protection. 45 00:04:22,830 --> 00:04:32,730 This principle focuses on ensuring the confidentiality and security of personal data used by AI applications. 46 00:04:32,760 --> 00:04:40,650 It requires effective policies and technologies to protect personal information from unauthorized access 47 00:04:40,650 --> 00:04:46,620 and misuse, as well as compliance with data privacy laws and regulations. 48 00:04:46,620 --> 00:04:50,940 Remember a immediately after the launch of ChatGPT? 49 00:04:51,060 --> 00:04:58,230 It was forbidden in some European countries like Italy, for example, and this was the reason. 50 00:04:58,380 --> 00:05:04,680 So a data protection laws are especially a. 51 00:05:06,010 --> 00:05:14,590 Difficult or sophisticated in the European Union, as they will be in many other countries around the 52 00:05:14,590 --> 00:05:14,950 world. 53 00:05:14,950 --> 00:05:23,800 And this is one of the areas where, if not properly prepared, LLM applications can find a lot of trouble. 54 00:05:23,800 --> 00:05:33,190 So before launching your LLM application, consider carefully where your data come from and if you are 55 00:05:33,190 --> 00:05:34,270 a. 56 00:05:35,230 --> 00:05:45,160 Keeping this principle of privacy and data protection well cover fourth principle transparency and explainability. 57 00:05:46,600 --> 00:05:56,080 Transparency in LLM applications in AI applications means that processes and decisions should be open 58 00:05:56,080 --> 00:05:59,410 and understandable to users and stakeholders. 59 00:05:59,740 --> 00:06:06,310 Explainability refers to the ability of artificial intelligence systems to provide understandable and 60 00:06:06,310 --> 00:06:14,260 meaningful explanations of their decisions and actions, which is crucial for building trust and understanding. 61 00:06:14,620 --> 00:06:20,830 So this is one of the principles that, in my opinion, are not a. 62 00:06:22,160 --> 00:06:24,590 100% follow by everybody. 63 00:06:25,040 --> 00:06:34,940 So if you remember Elon Musk, the founder of Tesla, and right now the the owner of the previous Twitter 64 00:06:34,940 --> 00:06:43,550 that right now is called ex Elon Musk, as you know, uh, which is one person closely, very closely 65 00:06:43,550 --> 00:06:48,530 associated with artificial intelligence because he was the originator. 66 00:06:48,830 --> 00:06:55,610 Uh, he had the initial idea for OpenAI, the company behind ChatGPT, and he was the person, you know, 67 00:06:55,610 --> 00:07:01,280 building the group of people and and, you know, preparing the initial philosophy of the company, 68 00:07:01,280 --> 00:07:01,880 etc.. 69 00:07:01,880 --> 00:07:07,100 Right now, he's not there anymore, and he is not in good terms with the team anymore. 70 00:07:07,100 --> 00:07:09,680 And in a. 71 00:07:10,460 --> 00:07:17,480 One of the last interviews of Elon Musk talking about Elon Musk, talking about artificial intelligence, 72 00:07:17,480 --> 00:07:23,300 got the attention of many people because one of the things he said there is well. 73 00:07:24,900 --> 00:07:32,730 Most LLM foundational models foundation models have been trained with. 74 00:07:34,150 --> 00:07:41,950 Data that a in many cases a ha is copyrighted. 75 00:07:43,120 --> 00:07:49,150 So this is one of the areas where LM a. 76 00:07:50,120 --> 00:07:51,860 Foundation models. 77 00:07:51,860 --> 00:07:58,550 Uh, the companies behind them, especially the private ones, are not going to be very specific, are 78 00:07:58,550 --> 00:08:07,640 not going to provide a lot of details because they can get in trouble because it is, uh, supposed 79 00:08:07,640 --> 00:08:18,980 or we think we can guess that the initial data they used to train their models may not be 100% open 80 00:08:18,980 --> 00:08:19,700 source. 81 00:08:20,300 --> 00:08:28,850 So this is one of the principles where you will have to well, you will have to think about and prepare 82 00:08:28,850 --> 00:08:36,440 yourself, uh, but you will also need to realize that not everybody is in the same line here. 83 00:08:37,870 --> 00:08:43,150 The principal number five accountability and governance is. 84 00:08:43,970 --> 00:08:45,470 Also an interesting one. 85 00:08:45,470 --> 00:08:53,900 Accountability implies that developers and users of LM applications and artificial intelligence applications 86 00:08:53,900 --> 00:09:02,270 must be responsible for their decisions and actions, and governance refers to the need to establish 87 00:09:02,270 --> 00:09:09,530 legal, ethical, and policy frameworks to guide the development and use of artificial intelligence, 88 00:09:09,530 --> 00:09:13,730 ensuring compliance with ethical and legal standards. 89 00:09:14,090 --> 00:09:15,200 So. 90 00:09:16,070 --> 00:09:19,070 In this second part, the governance. 91 00:09:19,070 --> 00:09:28,130 We have seen a lot of movements in the recent months, so we have seen a regulation comments coming 92 00:09:28,130 --> 00:09:33,770 in the European Union, United States and other parts of the world. 93 00:09:33,770 --> 00:09:41,660 So this is one of the areas where we are going to have a lot of movements in the coming months and years. 94 00:09:41,660 --> 00:09:52,850 There is a debate, especially in the startup world, regarding the a convenience or not convenience 95 00:09:52,850 --> 00:10:02,840 of a include a lot of regulations in a moment where we are start very, very very we are still very, 96 00:10:02,840 --> 00:10:03,470 very green. 97 00:10:03,470 --> 00:10:03,770 Right. 98 00:10:03,770 --> 00:10:10,670 So this is a very early moment for the a artificial intelligence generative artificial intelligence 99 00:10:10,670 --> 00:10:11,600 industry. 100 00:10:11,600 --> 00:10:20,210 And maybe it is not the right moment to start limiting the possibilities of the industry, but. 101 00:10:20,210 --> 00:10:25,190 Well, this is one very, very I would say a. 102 00:10:26,740 --> 00:10:29,740 This debate is right now over the table. 103 00:10:29,740 --> 00:10:30,190 Right. 104 00:10:30,190 --> 00:10:35,260 So let's see what happens with the regulations associated with artificial intelligence. 105 00:10:35,260 --> 00:10:43,840 But this is one of the most, uh, I would say, important principles in responsible AIS. 106 00:10:43,840 --> 00:10:52,210 And the last one that we are going to, uh, focus on in this brief introduction about responsible AI 107 00:10:52,210 --> 00:10:55,510 is inclusivity and accessibility. 108 00:10:55,630 --> 00:11:02,710 So this principle promotes the development of artificial intelligence applications and LM applications 109 00:11:02,710 --> 00:11:10,960 that are accessible and useful to a broad range of users, including those with disabilities or from 110 00:11:10,960 --> 00:11:12,490 marginalized groups. 111 00:11:12,730 --> 00:11:19,750 Inclusivity in artificial intelligence seeks to ensure that technology benefits the whole society, 112 00:11:19,750 --> 00:11:22,360 not just privileged sectors. 113 00:11:24,010 --> 00:11:32,350 So we have been talking about some problems that you, uh, may remember, uh, regarding the initial, 114 00:11:32,440 --> 00:11:35,980 uh, days of ChatGPT and this, uh, principle. 115 00:11:35,980 --> 00:11:36,460 Right. 116 00:11:36,460 --> 00:11:44,290 So it is important to have responsible AI in mind when you are preparing the launch of a profession 117 00:11:44,320 --> 00:11:46,990 of a professional LLM application. 118 00:11:46,990 --> 00:11:53,470 Sometimes this is a matter that is not going to be in the hands of the engineer. 119 00:11:53,920 --> 00:11:59,620 But so sometimes there it will be another department in your company handling handling this. 120 00:11:59,620 --> 00:12:07,450 But in many cases you would be as as the AI engineer in the company or one of them, you would be one 121 00:12:07,450 --> 00:12:16,810 of the people that should warn your company about possible problems in this area and how to prepare 122 00:12:16,810 --> 00:12:20,770 or prevent, uh, this, this, these problems to happen. 123 00:12:21,920 --> 00:12:30,380 In the next lesson, we are going to talk about a real world LM ops solutions you may find in the market, 124 00:12:30,380 --> 00:12:40,250 because this is an area LM ops where you probably are going to use external tools in order to handle 125 00:12:40,460 --> 00:12:42,050 this, this problem. 126 00:12:42,050 --> 00:12:47,840 So let's see some real, uh, solutions in the next lesson.