WEBVTT 00:30:51.440 --> 00:30:55.040 Gen AI, can it think like human? 00:30:55.040 --> 00:30:56.440 Can it reason? 00:31:01.080 --> 00:31:03.200 The golden test for intelligence is a Turing test. 00:31:03.200 --> 00:31:04.760 It's an idea that has been introduced 00:31:04.760 --> 00:31:07.880 more than 70 years ago, which is quite simple. 00:31:07.880 --> 00:31:10.080 Can a machine mislead the investigator 00:31:10.080 --> 00:31:13.040 into believing that he is talking to a human and not to a machine? 00:31:13.040 --> 00:31:17.680 Well, today's large language models can easily beat that test. 00:31:17.680 --> 00:31:21.360 Now, the question is if that makes them intelligent or not. 00:31:21.360 --> 00:31:23.160 There are very many other benchmarks out there, 00:31:23.160 --> 00:31:26.280 and some of them those machine easily beat. 00:31:26.280 --> 00:31:29.480 For example, it can take USMLE test, 00:31:29.480 --> 00:31:32.760 which is a test for becoming a doctor in the U.S. 00:31:32.760 --> 00:31:35.160 It can score very highly on GRE test. 00:31:35.160 --> 00:31:39.560 At the same time, it can make very, very simple and silly mistakes. 00:31:39.560 --> 00:31:42.880 And we all have seen that if we tried ChatGPT. 00:31:42.880 --> 00:31:47.280 So the jury’s still out there. But there is a common belief that 00:31:47.280 --> 00:31:52.080 these are really strong statistical models that have a lot of information embedded. 00:31:52.080 --> 00:31:57.560 In fact, famously, Yann LeCun said that the amount of memory 00:31:57.560 --> 00:32:00.840 compensates for the weak logic abilities of these models. 00:32:00.840 --> 00:32:05.160 Some critics of Gen AI just say it's a stochastic parrot. 00:32:05.160 --> 00:32:07.040 What do they mean exactly? 00:32:07.040 --> 00:32:08.560 It came from a paper published 00:32:08.560 --> 00:32:14.040 several years ago by one of the Google researchers. This term is kind of a stick. 00:32:14.040 --> 00:32:17.680 But what they're trying to say is it's stochastic, and it is true. 00:32:17.680 --> 00:32:18.880 The model, in fact, 00:32:18.880 --> 00:32:24.280 is just trying to predict the next word using for that distribution of the words 00:32:24.280 --> 00:32:27.320 that it learned from the large corpora of data. 00:32:27.320 --> 00:32:29.800 And so it's a random process, 00:32:29.800 --> 00:32:31.920 sampling from the distribution. And so it is stochastic. 00:32:31.920 --> 00:32:33.960 Why they call it parrot is to just emphasize 00:32:33.960 --> 00:32:37.800 that the model is missing any logical abilities, and, again, 00:32:37.800 --> 00:32:43.080 just generating tokens one at a time and out of a regressive manner. 00:32:43.080 --> 00:32:45.040 So it is, in some sense, the stochastic parrot. 00:32:45.040 --> 00:32:48.440 But at the same time, we are observing a lot of very interesting 00:32:48.440 --> 00:32:51.040 abilities of these models, quite unexpected. 00:32:51.040 --> 00:32:56.160 I think it's a bit of misleading to call it just a stochastic parrot. 00:32:56.160 --> 00:32:58.640 It's an interesting debate among the experts, 00:32:58.640 --> 00:33:02.880 but stochastic parrot or not, what's the impact for business? 00:33:02.880 --> 00:33:04.920 For business, it’s really doesn’t matter 00:33:04.920 --> 00:33:09.040 as long as algorithms can do their job, as long as we can, 00:33:09.040 --> 00:33:13.120 from somewhat unreliable parts, unreliable algorithms, 00:33:13.120 --> 00:33:15.440 build reliable systems that solve business problems, 00:33:15.440 --> 00:33:20.640 it really doesn’t matter if those algorithms can actually think or not. 00:33:20.640 --> 00:33:22.720 Now, of course, it's important for the future, 00:33:22.720 --> 00:33:23.840 important for the futurists, 00:33:23.840 --> 00:33:25.360 important for the future applications, 00:33:25.360 --> 00:33:28.960 but as of today, if they work, they work.