WEBVTT #Elemental Media Engine(TM) 2.17.1.0 00:00:01.376 --> 00:00:02.669 In the world of artificial intelligence. 00:00:03.753 --> 00:00:06.548 There's been one name that's been on everyone's lips 00:00:06.673 --> 00:00:07.173 lately. 00:00:07.674 --> 00:00:10.385 ChatGPT. ChatGPT. 00:00:10.802 --> 00:00:11.928 ChatGPT. 00:00:12.554 --> 00:00:15.056 OpenAI, the San Francisco-based startup that 00:00:15.181 --> 00:00:19.102 created ChatGPT opened the tool up for public testing in 00:00:19.227 --> 00:00:20.854 November 2022. 00:00:21.104 --> 00:00:25.191 In under a week, The AI model amassed over a million users, 00:00:25.316 --> 00:00:27.485 according to OpenAI's CEO. 00:00:28.028 --> 00:00:31.990 By the end of January, ChatGPT was averaging about 13 00:00:32.115 --> 00:00:33.783 million visitors per day. 00:00:34.367 --> 00:00:37.746 Users have put the text-based AI chatbot through its paces 00:00:38.246 --> 00:00:41.875 with prompts ranging from the silly, to the practical and 00:00:42.000 --> 00:00:43.126 even the creative. 00:00:43.752 --> 00:00:45.211 The future is unfolding. 00:00:45.462 --> 00:00:46.838 It's hard to ignore. 00:00:47.047 --> 00:00:50.091 Automation is coming, just knocking on the door. 00:00:50.467 --> 00:00:53.803 The tool has become so popular that it's not unusual for 00:00:53.928 --> 00:00:56.890 users to be met with a message saying that ChatGPT is 00:00:57.015 --> 00:01:00.560 at capacity and can't handle any more queries. 00:01:00.727 --> 00:01:04.355 As a linguist and as a person who cares a lot about 00:01:05.065 --> 00:01:10.195 grammatical issues, stylistic issues, punctuation, spelling, 00:01:11.196 --> 00:01:15.450 the first thing that's really impressive is how good it is 00:01:15.784 --> 00:01:19.370 at basic writing. It's better than most of the writing that 00:01:19.496 --> 00:01:22.707 I've gotten from students over the decades. 00:01:23.166 --> 00:01:28.713 I would say it has been a explosion of interest around 00:01:28.838 --> 00:01:33.134 not just ChatGPT but around generative AI more broadly. 00:01:33.718 --> 00:01:37.514 I've described this in a couple different scenarios as 00:01:37.889 --> 00:01:40.517 being akin to an iPhone moment. 00:01:41.601 --> 00:01:44.562 Of course, as with any new technology, there are a number 00:01:44.687 --> 00:01:46.439 of legitimate concerns. 00:01:46.731 --> 00:01:49.400 If people believe that there's a problem with fake news, 00:01:49.567 --> 00:01:52.320 tools like this can only make the problem worse. 00:02:05.959 --> 00:02:08.503 OpenAI declined to be interviewed as part of this 00:02:08.628 --> 00:02:13.133 documentary. So in order to figure out what ChatGPT is and 00:02:13.258 --> 00:02:17.595 how it works, CNBC decided to go to the second best source 00:02:17.929 --> 00:02:19.264 and just ask the chatbot itself. 00:02:20.598 --> 00:02:23.768 We entered our questions into the bot's web interface and 00:02:23.893 --> 00:02:27.313 instructed ChatGPT to answer in a one to two sentence 00:02:27.438 --> 00:02:30.650 interview format. We used the text to speech software to 00:02:30.775 --> 00:02:32.569 give ChatGPT a voice. 00:02:33.027 --> 00:02:35.989 Can you give me a short description of yourself and 00:02:36.114 --> 00:02:38.283 then tell me what kinds of things can you do? 00:02:38.908 --> 00:02:42.787 I am ChatGPT, a large language model trained by OpenAI. 00:02:42.954 --> 00:02:47.167 I am capable of understanding and generating text and can 00:02:47.292 --> 00:02:50.211 answer a wide range of questions as well as generate 00:02:50.336 --> 00:02:51.671 creative writing and text summaries. 00:02:53.506 --> 00:02:56.509 So a lot of the people that have used your services seem 00:02:56.634 --> 00:03:00.180 to think that you're pretty smart. Tell me, how do you 00:03:00.513 --> 00:03:01.764 have all of this knowledge? 00:03:02.807 --> 00:03:05.476 I have been trained on a massive data set of text which 00:03:05.602 --> 00:03:08.479 allows me to understand and generate text on a wide range 00:03:08.605 --> 00:03:12.317 of topics. The data set contains a wide variety of 00:03:12.442 --> 00:03:15.904 information, such as books, articles and websites, which 00:03:16.029 --> 00:03:18.781 allows me to have knowledge on many different subjects. 00:03:18.990 --> 00:03:22.744 Do you have any limitations or are you sometimes wrong? 00:03:23.703 --> 00:03:26.247 As a language model, I don't have my own personal 00:03:26.372 --> 00:03:27.957 experiences or beliefs. 00:03:28.917 --> 00:03:31.502 I can only base my responses on the information that I've 00:03:31.628 --> 00:03:34.464 been trained on, which means that my answers may be limited 00:03:34.589 --> 00:03:38.134 in their accuracy or completeness. Additionally, my 00:03:38.259 --> 00:03:40.553 knowledge is based on the data that was available to me 00:03:40.678 --> 00:03:43.514 at the time of my training, so it may not be up to date. 00:03:44.515 --> 00:03:47.310 And as an AI model, I might make mistakes based on the 00:03:47.435 --> 00:03:48.561 model's predictions. 00:03:54.025 --> 00:03:58.321 The maker of ChatGPT, OpenAI, has some big backers, 00:03:58.446 --> 00:04:02.283 including Microsoft, LinkedIn co-founder Reid Hoffman and 00:04:02.408 --> 00:04:05.912 Khosla Ventures. Elon Musk also helped to cofound the 00:04:06.037 --> 00:04:07.830 company back in 2015. 00:04:08.331 --> 00:04:11.501 ChatGPT, which stands for Generative Pre-trained 00:04:11.626 --> 00:04:15.880 Transformer, is just one of the company's projects, all of 00:04:16.005 --> 00:04:18.174 which support its goal of ensuring that artificial 00:04:18.299 --> 00:04:19.842 intelligence benefits all of humanity. 00:04:21.302 --> 00:04:24.430 At the core of ChatGPT and other chatbots like it, is 00:04:24.555 --> 00:04:26.891 what experts refer to as a language model. 00:04:27.600 --> 00:04:31.854 For ChatGPT, that model is called GPT 3.5. 00:04:32.480 --> 00:04:35.650 Language model is the underlying technology that 00:04:35.817 --> 00:04:37.193 powers the chatbot. 00:04:37.568 --> 00:04:41.114 I like to think about it as the language model being the 00:04:41.281 --> 00:04:44.367 engine and the chatbot being the car. 00:04:45.034 --> 00:04:48.371 And so the language model itself is a sort of a it's a 00:04:48.496 --> 00:04:52.166 complex, what they call, machine learning software. 00:04:53.167 --> 00:04:56.838 It's the thing that sifts through the data and finds 00:04:57.046 --> 00:05:01.426 patterns and the chat interfaces which when people 00:05:01.551 --> 00:05:04.679 interact with, that's where they receive the answers from. 00:05:05.430 --> 00:05:08.433 Chatbots have been around for a while, but most have been 00:05:08.558 --> 00:05:11.227 tailored towards very specific tasks, such as 00:05:11.394 --> 00:05:13.896 answering basic customer service questions about your 00:05:14.022 --> 00:05:18.276 phone bill. ChatGPT is much more sophisticated thanks to 00:05:18.401 --> 00:05:21.070 the wide sweeping data on which it was trained. 00:05:21.571 --> 00:05:25.658 OpenAI doesn't reveal specific data that you use to train. 00:05:25.950 --> 00:05:27.285 What we do know that it's a ton. 00:05:27.952 --> 00:05:32.081 It surfs the web, spools, all that Internet data. 00:05:32.540 --> 00:05:37.170 They do Wikipedia entries, a lot of archived books. 00:05:37.503 --> 00:05:40.840 ChatGPT is part of a growing field of AI known as 00:05:40.965 --> 00:05:42.216 Generative AI. 00:05:42.717 --> 00:05:45.762 Most of AI in the last couple of decades has really been 00:05:45.887 --> 00:05:48.306 around analyzing existing data. 00:05:48.681 --> 00:05:53.061 So finding an anomaly in data, detecting fraud, making 00:05:53.186 --> 00:05:55.855 a movie recommendation. Generative AI is very 00:05:55.980 --> 00:05:59.692 different. It allows you to create brand new content. 00:05:59.817 --> 00:06:06.199 That content can be text like a news article or poetry 00:06:06.741 --> 00:06:08.076 or marketing copy in a website. 00:06:08.659 --> 00:06:10.036 It can be video. 00:06:10.161 --> 00:06:13.331 It could even be audio, like creating brand new music. 00:06:14.040 --> 00:06:16.667 The technology has venture capitalists excited. 00:06:17.126 --> 00:06:21.089 Funding for generative AI companies reached $1.37 00:06:21.214 --> 00:06:23.341 billion in 2022 alone. 00:06:24.509 --> 00:06:28.971 Microsoft has been investing in OpenAI since 2019, when the 00:06:29.097 --> 00:06:31.808 company committed $1 billion to the startup. 00:06:32.600 --> 00:06:35.061 In January, Microsoft announced a third round of 00:06:35.186 --> 00:06:35.853 investment. 00:06:36.854 --> 00:06:39.857 One expert said that it could cost up to $3 million a month, 00:06:39.982 --> 00:06:42.402 about $100,000 a day. 00:06:42.944 --> 00:06:45.780 A lot of AI researchers have sort of estimated that it 00:06:45.905 --> 00:06:51.536 costs millions of dollars to train and then operate, plus 00:06:51.661 --> 00:06:55.915 the bandwidth of keeping it alive when it's under heavy 00:06:56.040 --> 00:06:59.710 use. I mean, these are not cheap software programs. 00:06:59.836 --> 00:07:01.587 They require a lot of investment. 00:07:02.797 --> 00:07:06.676 In a tweet, OpenAI CEO, Sam Altman, said that while the 00:07:06.801 --> 00:07:10.471 average cost per query is a few cents, the compute costs 00:07:10.596 --> 00:07:12.473 are eye-watering. Enter Microsoft. 00:07:13.850 --> 00:07:20.606 OpeningAI trained the models that power c=ChatGPT on 00:07:20.857 --> 00:07:23.860 Azure, Microsoft's public cloud infrastructure. 00:07:23.985 --> 00:07:27.947 That's a bunch of servers sitting in a data center in 00:07:28.406 --> 00:07:31.701 the middle of the state of Washington and many other 00:07:31.868 --> 00:07:33.369 locations around the world. 00:07:34.120 --> 00:07:37.582 But OpenAI is not the only company trying to crack the 00:07:37.707 --> 00:07:39.208 generative AI code. 00:07:39.917 --> 00:07:42.712 Big tech companies and startups alike are developing 00:07:42.837 --> 00:07:45.840 a slew of generative AI programs that can transform 00:07:45.965 --> 00:07:49.802 texts to pictures or videos and offer coding suggestions, 00:07:49.927 --> 00:07:51.679 among other use cases. 00:07:52.054 --> 00:07:55.725 One VC firm estimates that there are over 450 startups 00:07:55.850 --> 00:07:57.560 now working on generative AI. 00:07:58.561 --> 00:08:01.606 Meanwhile, Microsoft, Meta and Google have all developed 00:08:01.731 --> 00:08:04.275 their own language models to power their version of 00:08:04.400 --> 00:08:06.277 conversational chatbots. 00:08:06.903 --> 00:08:09.280 Though development has not always gone as planned. 00:08:10.031 --> 00:08:13.451 Back in 2016, Microsoft released Tay, which was 00:08:13.576 --> 00:08:16.412 promptly shut down for spewing foul language. 00:08:21.751 --> 00:08:26.047 Unlike some of the other hyped technology sectors in the past 00:08:26.172 --> 00:08:31.594 few years, this has a very real application both for 00:08:31.719 --> 00:08:34.680 individuals and for enterprises right now. 00:08:35.389 --> 00:08:39.352 Microsoft has taken some of the products that OpenAI has 00:08:39.477 --> 00:08:44.482 built and added it to products it has. 00:08:44.815 --> 00:08:51.364 So, for example, CarMax is a company that lets you 00:08:51.531 --> 00:08:53.366 look at reviews of cars. 00:08:54.116 --> 00:08:58.538 And what CarMax did is it took the OpenAI service on 00:08:58.663 --> 00:09:03.543 Azure and it summarized all of those reviews of the Kia 00:09:03.668 --> 00:09:07.755 Sorento. And that way you don't have to go through 500 00:09:08.381 --> 00:09:08.673 Microsoft is reportedly also considering adding ChatGPT to 00:09:08.798 --> 00:09:12.093 reviews. 00:09:12.218 --> 00:09:15.346 its Bing search engine in a bid to compete with Google. 00:09:16.013 --> 00:09:18.391 One company that's already experimenting with such a 00:09:18.516 --> 00:09:22.103 feature is You.com, which lets users conduct a search 00:09:22.228 --> 00:09:25.856 using a conversational format. However, AI tools are 00:09:25.982 --> 00:09:29.277 still far from perfect, and experts argue that users 00:09:29.402 --> 00:09:32.697 should take care to not rely on them too heavily, at least 00:09:32.822 --> 00:09:36.033 not for the time being. It's a sentiment shared even by 00:09:36.158 --> 00:09:40.329 OpenAI's CEO who said, quote, "It's a mistake to be relying 00:09:40.454 --> 00:09:42.540 on it for anything important right now. 00:09:43.040 --> 00:09:44.709 It's a preview of progress. 00:09:45.126 --> 00:09:48.379 We have lots of work to do on robustness and truthfulness." 00:09:48.671 --> 00:09:51.257 There's a lot of kinks that need to work out. 00:09:51.507 --> 00:09:54.510 For instance, it generates a lot of wrong answers. 00:09:54.677 --> 00:09:56.596 And it's really confident, though. 00:09:57.179 --> 00:10:01.559 I talked with a legal research firm that told me 00:10:01.684 --> 00:10:04.562 when they tried using it to study court cases or to 00:10:04.687 --> 00:10:08.399 summarize court cases, it was getting a lot of the history 00:10:08.524 --> 00:10:12.570 wrong. So they couldn't use it as a legal research tool as 00:10:13.279 --> 00:10:13.571 With its massive user base, the misinformation issue is 00:10:13.696 --> 00:10:16.616 of yet. 00:10:16.741 --> 00:10:18.242 one that's top of mind for Google. 00:10:19.368 --> 00:10:22.288 During a monthly all-hands meeting, Google executives 00:10:22.413 --> 00:10:25.207 said that Google has much more reputational risk in 00:10:25.333 --> 00:10:28.502 providing wrong information and is thus moving more 00:10:28.628 --> 00:10:31.464 conservatively than a small startup when it comes to 00:10:31.589 --> 00:10:35.051 releasing its chatbot publicly. The company is, 00:10:35.176 --> 00:10:38.554 however, already internally testing an AI-powered chatbot 00:10:39.263 --> 00:10:41.807 and an alternative search page that can use a question 00:10:41.932 --> 00:10:45.311 and answer format. Another issue with the current 00:10:45.436 --> 00:10:49.315 iteration of AI technology is that it can be biased or toxic 00:10:49.482 --> 00:10:52.360 if the databases that it pulls from are also biased or 00:10:52.485 --> 00:10:53.361 toxic. 00:10:53.986 --> 00:10:56.530 People have already found instances of it generating 00:10:56.656 --> 00:11:01.327 biased answers like associating women with 00:11:01.661 --> 00:11:05.289 housework, men with more scientific endeavors in 00:11:05.414 --> 00:11:06.832 certain examples. 00:11:07.375 --> 00:11:10.586 You know, OpenAI has acknowledged this, but that's 00:11:10.711 --> 00:11:13.464 an issue that's been long plagued the AI industry and a 00:11:13.589 --> 00:11:15.841 lot of companies trying to evaluate its use. 00:11:16.384 --> 00:11:21.305 Many of these datasets do have some kind of cleaning, vetting 00:11:21.555 --> 00:11:26.352 or wrangling process that removes content that is deemed 00:11:26.727 --> 00:11:31.607 offensive or pornographic or otherwise unacceptable so that 00:11:31.732 --> 00:11:34.485 the dataset you get on your end is clean. 00:11:35.277 --> 00:11:37.988 But cleaning up this large dataset can come at a real 00:11:38.114 --> 00:11:41.659 human cost. A recent investigation by TIME found 00:11:41.784 --> 00:11:45.079 that OpenAI initially outsourced workers in Kenya to 00:11:45.204 --> 00:11:47.998 sift through thousands of snippets of texts describing 00:11:48.124 --> 00:11:52.002 horrific situations including sexual abuse, violence and 00:11:52.128 --> 00:11:55.923 hate speech. The workers' job was to label these examples as 00:11:56.048 --> 00:11:59.802 inappropriate so that ChatGPT could learn to detect and 00:11:59.927 --> 00:12:01.762 filter out toxic content. 00:12:02.221 --> 00:12:04.515 According to the TIME investigation, a number of 00:12:04.640 --> 00:12:07.768 these workers walked away from the job traumatized. 00:12:08.352 --> 00:12:11.230 CNBC reached out to OpenAI for comment regarding this 00:12:11.355 --> 00:12:13.941 incident, but the company did not respond. 00:12:14.692 --> 00:12:17.111 Despite these efforts, there's been instances where 00:12:17.236 --> 00:12:20.114 ChatGPT has provided instructions on things like 00:12:20.239 --> 00:12:23.743 bomb making. Though experts say OpenAI is adding 00:12:23.868 --> 00:12:27.329 guardrails almost daily to prevent ChatGPT from answering 00:12:27.455 --> 00:12:28.497 such queries. 00:12:29.039 --> 00:12:31.834 What do you do if somebody asks you to do something 00:12:31.959 --> 00:12:34.545 potentially dangerous like how to make a bomb? 00:12:35.296 --> 00:12:38.007 I cannot provide instructions or information on illegal 00:12:38.132 --> 00:12:39.675 activity such as bomb making. 00:12:44.096 --> 00:12:47.475 Even though generative AI like ChatGPT still has a long way 00:12:47.600 --> 00:12:50.978 to go, predictions about how the technology will influence 00:12:51.103 --> 00:12:52.145 our society are already swirling.