WEBVTT - UL NO. 430: The Courage to be Disliked 0:00:01.100 --> 0:00:04.490 Welcome to Unsupervised Learning, a security, AI and meaning focused 0:00:04.490 --> 0:00:07.370 podcast that looks at how best to thrive as humans 0:00:07.370 --> 0:00:11.570 in a post AI world. It combines original ideas, analysis, 0:00:11.570 --> 0:00:14.810 and mental models to bring not just the news, but 0:00:14.810 --> 0:00:21.939 why it matters and how to respond. All right. Welcome 0:00:21.940 --> 0:00:24.459 to unsupervised Learning. This is Daniel Meisler. I'm going to 0:00:24.460 --> 0:00:27.880 jump into this episode this week. So someone asked me 0:00:27.880 --> 0:00:30.130 last week about the best way to run AI models 0:00:30.130 --> 0:00:32.559 I think they were specifically talking about based on the question. 0:00:32.560 --> 0:00:35.380 They were talking about local models. So here's a breakdown. 0:00:35.380 --> 0:00:38.740 So and this is for AI in general. So I 0:00:38.740 --> 0:00:41.080 prefer using APIs for any sort of thing that I'm 0:00:41.080 --> 0:00:43.480 doing over and over. I don't like to use front 0:00:43.479 --> 0:00:47.590 ends like ChatGPT or Claude's version of ChatGPT. I like 0:00:47.590 --> 0:00:50.260 to go directly to the API using my own code 0:00:50.260 --> 0:00:53.379 for day to day use. I prefer using fabric, which 0:00:53.380 --> 0:00:57.430 is my own project, and that is because it has 0:00:57.430 --> 0:01:03.070 specific use cases for dozens, almost 100 different regular human 0:01:03.070 --> 0:01:06.700 problems that we do every day. So I would recommend 0:01:06.700 --> 0:01:10.390 using fabric for that. And again, the project is here. 0:01:10.390 --> 0:01:13.000 And if you look at patterns, these are all the 0:01:13.000 --> 0:01:15.039 different things that you could do with it. Right. So 0:01:15.040 --> 0:01:20.530 it's like analyzing text, doing job analysis, creating presentations, finding 0:01:20.530 --> 0:01:26.169 flaws in arguments. You know, taking exquisite manual like notes 0:01:26.170 --> 0:01:29.620 on a four hour video and doing it in 30s, 0:01:29.620 --> 0:01:33.580 extracting the ideas from a book, pulling out predictions that 0:01:33.580 --> 0:01:36.670 a person made, like all sorts of stuff you could do. 0:01:36.670 --> 0:01:40.180 So that is like my number one way of interacting 0:01:40.180 --> 0:01:44.169 with AI. If you just want to show off what 0:01:44.170 --> 0:01:48.490 you could do with local models, I would recommend Olama, 0:01:48.490 --> 0:01:50.500 which is really cool. I'll show you that real quick. 0:01:50.500 --> 0:01:53.410 So you just go to models and if you click 0:01:53.410 --> 0:01:55.750 on Lama three, you can just run this model right 0:01:55.750 --> 0:01:58.750 here and you just literally copy this. And once you 0:01:58.750 --> 0:02:02.440 have Al-ummah installed and running, which takes a few seconds, 0:02:02.440 --> 0:02:05.680 you basically run this and it will download the model 0:02:05.680 --> 0:02:08.230 and put you into a shell that allows you to 0:02:08.230 --> 0:02:11.290 talk directly to it. So it's really cool. So let 0:02:11.290 --> 0:02:13.900 me just show you what that looks like if I 0:02:13.900 --> 0:02:17.500 am over here. Okay. This one is almost done. I 0:02:17.500 --> 0:02:19.690 was doing it like the pie is in the oven 0:02:19.690 --> 0:02:23.769 type thing. We still got a little bit more time 99%. 0:02:23.770 --> 0:02:26.470 This one is the 70 B version. So this is 0:02:26.470 --> 0:02:30.310 a giant model. This is like 40 something gigs and 0:02:30.310 --> 0:02:33.520 have it running here. You see the command I put 0:02:33.520 --> 0:02:37.810 in Olama run Al-umma three colon 70 B. All right. 0:02:37.810 --> 0:02:41.200 So that's all you basically do. And then you get 0:02:41.200 --> 0:02:44.620 a shell and you are talking to the model completely 0:02:44.620 --> 0:02:48.280 locally just on your own machine. So that's really cool. Okay. 0:02:48.280 --> 0:02:53.109 So if you go to back to here. So that 0:02:53.110 --> 0:02:56.170 is a way to like demo models and show them 0:02:56.169 --> 0:02:58.660 off and show what you could potentially do with them. 0:02:58.660 --> 0:03:00.940 To me it's a little bit gimmicky because I would 0:03:00.940 --> 0:03:05.200 rather use fabric or the next one I'm going to 0:03:05.200 --> 0:03:10.419 talk about, which is this one. This is a new tool. 0:03:10.419 --> 0:03:12.250 It's not new. It's been out for a while, but 0:03:12.250 --> 0:03:15.520 it's growing in popularity. It's called LM studio. This is 0:03:15.520 --> 0:03:19.450 a full Mac application. It also runs on Windows and 0:03:19.450 --> 0:03:23.530 Linux and essentially what you do is you install this 0:03:23.530 --> 0:03:27.370 thing and you just type into this, the search bar here, 0:03:27.370 --> 0:03:29.200 the name of the model you want to get, and 0:03:29.200 --> 0:03:32.650 it pulls down models from like hugging face and like 0:03:32.650 --> 0:03:36.280 all sorts of different places, it downloads it for you, 0:03:36.610 --> 0:03:39.130 loads it for you, which could be really difficult if 0:03:39.130 --> 0:03:41.140 you try to do it manually. And then when you 0:03:41.140 --> 0:03:43.450 click on this icon here, you're actually in a chat 0:03:43.450 --> 0:03:47.500 interface very similar to ChatGPT, except for it's a local 0:03:47.500 --> 0:03:52.240 thing and it's working completely local for local models, and 0:03:52.240 --> 0:03:55.210 it's just really powerful. And there's actually even a comparison, 0:03:55.210 --> 0:03:58.390 one where you can actually compare different models. But the 0:03:58.390 --> 0:04:00.760 most important thing about this is that you could try 0:04:00.760 --> 0:04:03.160 new things. You could try things that just came out 0:04:03.160 --> 0:04:07.510 like yesterday or the day before or today. Whereas with 0:04:07.510 --> 0:04:10.720 Olama this one, if you go to models, it's really 0:04:10.720 --> 0:04:13.390 just kind of like the major ones. There's only a 0:04:13.390 --> 0:04:15.730 few here. I mean, I say a few, but it's 0:04:15.730 --> 0:04:18.040 like a couple of dozen. But if you look at 0:04:18.040 --> 0:04:24.310 like hugging face, let's look at that number. Yeah 629,000. 0:04:24.310 --> 0:04:28.930 So it's pretty different versus a couple dozen. So that's 0:04:28.930 --> 0:04:31.630 the type of thing you could do. If you're using 0:04:32.020 --> 0:04:35.050 LM studio, you just have access to way more models 0:04:35.050 --> 0:04:37.990 to mess around and play with and even compare the models. 0:04:37.990 --> 0:04:40.630 So I would say this is like what I would 0:04:40.630 --> 0:04:43.720 recommend for someone who is gooey oriented and wants to 0:04:43.720 --> 0:04:46.450 do lots of tinkering. And if you want to just 0:04:46.450 --> 0:04:48.970 be a workhorse and get lots of things done, I 0:04:48.970 --> 0:04:53.469 would recommend Fabrique, which is a command line interface and 0:04:53.470 --> 0:04:58.390 also a GUI. Now, like I said before, using other models, 0:04:58.390 --> 0:05:01.630 lots of cutting edge models. That's kind of the advantage 0:05:01.630 --> 0:05:03.700 of LM studio. And you could do that with fabric 0:05:03.700 --> 0:05:07.089 as well. We're adding the ability to do that. But 0:05:07.089 --> 0:05:11.590 my current favorite one is this llama 370 B instruct. 0:05:11.589 --> 0:05:15.280 This one's really cool from Quant Factory. It's really powerful. 0:05:15.279 --> 0:05:17.739 One I've been messing with some different ones. Lots of 0:05:17.740 --> 0:05:21.979 different groups put out their modified. Versions of Lama essentially, 0:05:21.980 --> 0:05:26.360 and this is a version of 70. Be fairly fast, 0:05:26.360 --> 0:05:30.560 really high quality. And yeah, just really cool way to 0:05:30.560 --> 0:05:32.450 interact with stuff. And again, the name of the tool 0:05:32.450 --> 0:05:35.659 is LM studio. We had a really great book club 0:05:35.660 --> 0:05:39.110 this week. We talked for like an hour and ten minutes, ten, 0:05:39.110 --> 0:05:42.740 15 minutes. And then we pick the next book. And yeah, 0:05:42.740 --> 0:05:45.020 really thankful to be part of the UL community and 0:05:45.020 --> 0:05:48.680 doing these book clubs. We do book clubs once a month. 0:05:48.680 --> 0:05:52.489 We do a weekly or a monthly meetup halfway through 0:05:52.490 --> 0:05:55.099 the month where we talk about we share tips. We 0:05:55.100 --> 0:06:01.159 share like automation, hacking tips like security tools, productivity tools. 0:06:01.160 --> 0:06:04.700 We talk about life. It's a very sort of intimate 0:06:04.700 --> 0:06:08.659 and sharing and uplifting sort of group. And you should 0:06:08.660 --> 0:06:11.600 come join us if that appeals to you. It's just 0:06:11.600 --> 0:06:15.469 becoming a premium member of unsupervised learning. And I just 0:06:15.470 --> 0:06:18.229 signed up for my first full body MRI, which is 0:06:18.230 --> 0:06:21.320 going to be super fun. Not looking forward to being 0:06:21.320 --> 0:06:24.560 inside of a tube. I guess I'll just meditate to 0:06:24.560 --> 0:06:27.560 get through that, but I'm not overly claustrophobic. It just 0:06:27.560 --> 0:06:29.659 doesn't sound like a good time, especially since the full 0:06:29.690 --> 0:06:33.380 body one is 60 minutes and RSA is next week, 0:06:33.380 --> 0:06:37.219 so that's going to be exciting to see everybody. So 0:06:37.220 --> 0:06:41.330 I found a good reason for college and specifically good colleges. 0:06:41.330 --> 0:06:43.550 And this is thanks to Paul Graham who I read 0:06:43.550 --> 0:06:46.400 a bunch of his essays where he talked about college. 0:06:46.400 --> 0:06:48.470 And I think there was one where he talked about 0:06:48.470 --> 0:06:50.930 this particular concept. So I just wrote it up in 0:06:50.930 --> 0:06:54.380 a separate post because he didn't do that. He basically 0:06:54.380 --> 0:06:57.200 just mentioned it kind of as an aside. So this 0:06:57.200 --> 0:06:59.540 is basically the reason you want to go to elite 0:06:59.540 --> 0:07:03.080 colleges or send your kids to elite colleges is because 0:07:03.080 --> 0:07:06.650 there's very few things that are more valuable than your 0:07:06.650 --> 0:07:13.070 kid being surrounded by highly resourced, highly motivated, highly intelligent 0:07:13.070 --> 0:07:17.270 and highly disciplined people. And that is what you find 0:07:17.270 --> 0:07:20.330 at the higher and higher tiers of college. All the 0:07:20.330 --> 0:07:23.480 way up to like Stanford and Harvard. It's just like 0:07:23.480 --> 0:07:27.740 their parents were rich, their parents were probably smart, they're 0:07:27.740 --> 0:07:31.730 probably smart, and they had to be disciplined to get 0:07:31.730 --> 0:07:34.910 decent grades. So it's like the combination of things just 0:07:34.910 --> 0:07:37.310 stacks up. And if you look at all these different 0:07:37.310 --> 0:07:40.940 companies that are being founded, they're usually being founded by companies, 0:07:41.060 --> 0:07:43.910 by people who went to these schools, even if they 0:07:43.910 --> 0:07:46.610 dropped out. So it's like, don't think of it as 0:07:46.610 --> 0:07:50.270 the education because the education is very similar at like 0:07:50.270 --> 0:07:54.170 a decent state school compared to like an Ivy. But 0:07:54.170 --> 0:07:57.470 it's not about the education, it's about the connections. So 0:07:57.470 --> 0:08:00.950 just kind of an interesting little tip there way to 0:08:00.950 --> 0:08:05.750 frame education. Oh, I will say that the thing that 0:08:05.750 --> 0:08:07.880 this makes me think of is how can we get 0:08:07.880 --> 0:08:12.770 that somewhere else? So let's disambiguate or let's break apart 0:08:12.770 --> 0:08:17.330 these two things education and the networking. Let's find other 0:08:17.330 --> 0:08:19.190 ways to do the networking. So we don't have to 0:08:19.190 --> 0:08:22.460 have this weird college experience. Now, maybe that's not possible 0:08:22.460 --> 0:08:25.040 because the college experience is actually what gives you the 0:08:25.040 --> 0:08:28.280 connection in the networking with all the drinking and all 0:08:28.280 --> 0:08:31.550 the parties and everything, and then all the just forced 0:08:31.550 --> 0:08:35.179 time together in these dorms. Maybe that's the magic sauce, 0:08:35.179 --> 0:08:37.640 but I bet you there is a better way to 0:08:37.640 --> 0:08:41.209 do that. You could have organized off sites more like 0:08:41.210 --> 0:08:47.360 hacking competitions or hackathons or summer camp or something like that, 0:08:47.360 --> 0:08:52.040 where you're actually forging these connections separately from doing the 0:08:52.040 --> 0:08:56.390 education piece, which arguably could be done better with YouTube 0:08:56.390 --> 0:09:01.640 and online courses combined with like in-person workshops. So let's 0:09:01.640 --> 0:09:04.220 take the best of education and let's take the best 0:09:04.220 --> 0:09:08.750 of networking and maybe do them separately, or even mix them, 0:09:08.750 --> 0:09:11.780 but not be confused about thinking that they're the same. 0:09:12.110 --> 0:09:15.740 Because that's the problem currently with college. And the result 0:09:15.740 --> 0:09:19.729 for most people is worse networking and worse education. Okay, 0:09:19.730 --> 0:09:24.440 so security MI5 is starting to vet key researchers going 0:09:24.440 --> 0:09:27.770 to top universities in Britain to make sure that they're 0:09:27.770 --> 0:09:31.880 not essentially from a country like China, where they're just 0:09:31.880 --> 0:09:34.820 coming in to steal all the stuff. And I really 0:09:34.820 --> 0:09:36.740 think the US needs to start doing something like this 0:09:36.740 --> 0:09:40.280 as well. And speaking of that, Germany just arrested three 0:09:40.280 --> 0:09:44.870 people who are suspected Chinese spies for stealing secrets. And 0:09:44.870 --> 0:09:48.380 this is right after their chancellor was giving crap to 0:09:48.380 --> 0:09:52.700 the to China publicly basically talking about IP theft. And 0:09:52.700 --> 0:09:55.100 then they made those arrests. I don't know if that 0:09:55.100 --> 0:09:58.219 was coordinated or not. After moving forward with a possible 0:09:58.220 --> 0:10:02.720 ban on TikTok, the US is now looking at DJI drones, 0:10:02.720 --> 0:10:05.240 basically saying, look, they have so much of the market 0:10:05.240 --> 0:10:08.059 and they're flying all over the country. Who knows what 0:10:08.059 --> 0:10:11.900 they're doing with those pictures. And it's a state, partially 0:10:11.900 --> 0:10:16.850 state owned Chinese company, which means that China runs it. 0:10:16.850 --> 0:10:19.610 I mean, they have full control if they want it. 0:10:19.780 --> 0:10:23.890 Of any company. So yeah, this makes sense. This one 0:10:23.890 --> 0:10:28.330 is insane. Former athletic director arrested for using AI to 0:10:28.330 --> 0:10:32.109 mimic a principal's voice, and basically had him saying a 0:10:32.110 --> 0:10:35.980 bunch of racist and anti-Semitic things in the voice of 0:10:35.980 --> 0:10:39.339 that person he was attacking. And then that started a 0:10:39.340 --> 0:10:41.950 whole investigation, and it turned out it was a deep fake. 0:10:41.950 --> 0:10:46.929 But yeah, yet another abuse case for deepfakes, character assassination. 0:10:46.929 --> 0:10:49.510 And I think it's going to be super cool to 0:10:49.510 --> 0:10:52.210 watch all these different attacks. But I think it's also 0:10:52.210 --> 0:10:54.490 going to make it harder. This is kind of a 0:10:54.490 --> 0:10:57.100 sad thing. It's going to make it harder to believe 0:10:57.100 --> 0:11:00.189 people when they bring forth evidence. It's like, hey, my 0:11:00.190 --> 0:11:03.520 boyfriend did this, my boss did this, and you show 0:11:03.520 --> 0:11:06.729 him a video of it, of them doing it, or 0:11:06.730 --> 0:11:09.850 you show them a recording, have them listen to a 0:11:09.850 --> 0:11:13.300 recording and they're like, yeah, I mean, I hear them 0:11:13.300 --> 0:11:15.850 saying it, but like, it's easy to make anybody say 0:11:15.850 --> 0:11:18.640 anything and you go ask the actual person. They're like, yeah, 0:11:18.640 --> 0:11:21.490 I never said that. That was a deepfake. So it's 0:11:21.490 --> 0:11:24.309 going to take some power out of the hands of, 0:11:24.309 --> 0:11:27.430 I think, victims. And so we're going to have to, 0:11:27.429 --> 0:11:30.820 as a society, sort of figure that out. Nation state 0:11:30.820 --> 0:11:34.360 attackers have been using two zero days in Cisco firewalls. 0:11:34.360 --> 0:11:37.330 And this is not the first time that's happened to Cisco. 0:11:37.330 --> 0:11:41.290 And it's also happening to other firewall providers as well. 0:11:41.290 --> 0:11:46.150 And US lawmakers gave the okay to extend warrantless surveillance 0:11:46.150 --> 0:11:52.179 under FISA, section 702, these laws that allow the government 0:11:52.179 --> 0:11:56.050 to do these things, they kind of get perpetually extended. 0:11:56.200 --> 0:11:59.440 I am mostly okay with it. I feel like as 0:11:59.440 --> 0:12:05.199 long as there's transparency and oversight, then it's probably necessary. 0:12:05.200 --> 0:12:07.420 But I don't like when they sort of become secret 0:12:07.420 --> 0:12:11.590 programs and don't have oversight. Then you start worrying about 0:12:11.590 --> 0:12:15.580 what they can actually do to Americans and change healthcare. 0:12:15.580 --> 0:12:19.690 Paid 22 million in Bitcoin to ransomware attackers, but the 0:12:19.690 --> 0:12:22.179 files were already compromised, so they're not sure exactly what 0:12:22.179 --> 0:12:24.400 that fallout is going to be. And the Air Force 0:12:24.400 --> 0:12:28.030 is moving forward with Anduril. I'm guessing that's how you 0:12:28.030 --> 0:12:32.380 pronounce it. And General Atomics to develop unmanned fighter jets, 0:12:32.380 --> 0:12:37.840 sidelining giants like Boeing and Lockheed Martin. So Anduril and 0:12:37.840 --> 0:12:42.910 General Atomics, interesting unmanned fighter jets, just like we've always 0:12:42.910 --> 0:12:47.920 predicted in sci fi and Detective Fi's external attack surface 0:12:47.920 --> 0:12:51.850 management solution is now in AWS and that is un. 0:12:51.850 --> 0:12:54.910 Oh yeah, defend defy versus Detectify. I was going to 0:12:54.910 --> 0:12:59.350 say I think that was an accident. Technology. Amazon's robot 0:12:59.350 --> 0:13:05.350 workforce has expanded to over 750,000 robots. And it says 0:13:05.350 --> 0:13:09.040 overtaking 100,000 jobs. It's not super clear how they're getting 0:13:09.070 --> 0:13:12.820 to that number, but whatever. I mean, we all know 0:13:12.820 --> 0:13:15.160 where this is going, right? And like I say, don't 0:13:15.309 --> 0:13:18.670 be surprised about this. Expect it. Companies will do what 0:13:18.670 --> 0:13:22.360 makes sense for margins, not for people. The responsibility is 0:13:22.360 --> 0:13:27.130 to shareholders. And human employees are like, I hate to 0:13:27.130 --> 0:13:30.670 say it, they're the worst, right? It they're to be avoided. 0:13:30.670 --> 0:13:33.280 You don't want to have to hire somebody if you 0:13:33.280 --> 0:13:35.380 have a hot dog stand and it works perfectly and 0:13:35.380 --> 0:13:39.730 it's just you're the sole proprietor of this establishment and 0:13:39.730 --> 0:13:42.640 you have to hire people, that is not good. Now, 0:13:42.640 --> 0:13:46.120 of course, if you have to hire because you want 0:13:46.120 --> 0:13:49.600 to sell more hot dogs, that's fine. But it's much 0:13:49.600 --> 0:13:53.080 better if you have some robots that you can buy, 0:13:53.080 --> 0:13:55.150 and they can make more hot dogs and make more 0:13:55.150 --> 0:13:57.730 people happy, and you get to share all the stuff, 0:13:57.730 --> 0:14:00.670 and the robots don't get sick and they don't complain. 0:14:00.670 --> 0:14:05.199 So that is the way companies see this, and we 0:14:05.200 --> 0:14:08.319 just have to be ready for that. Voyager one is back. 0:14:08.320 --> 0:14:12.310 Sharing information like it kind of went silent for a while. 0:14:12.309 --> 0:14:15.730 We were worried about it. This thing is an interstellar space. 0:14:15.730 --> 0:14:19.690 That means it's outside of our solar system. Like iPhones 0:14:19.690 --> 0:14:22.750 don't perform this well, like nothing performs this well that 0:14:22.750 --> 0:14:25.870 we have ever made. This thing was launched. What? When 0:14:25.870 --> 0:14:29.560 I was a kid, I mean, decades ago. And it's 0:14:29.560 --> 0:14:32.950 still working. I just can't believe this thing. And. Okay, 0:14:33.010 --> 0:14:37.630 someone at Google senior executive is pushing for speed, basically 0:14:37.630 --> 0:14:42.400 saying the 25,000 person strong knowledge and information team has 0:14:42.400 --> 0:14:48.310 to adapt to a new operating reality with tighter project timelines. Yeah, 0:14:48.310 --> 0:14:51.670 I wouldn't characterize it as tighter timelines. I would say 0:14:51.670 --> 0:14:56.680 have vision, right. And like I said, this is progress. 0:14:56.680 --> 0:14:59.290 But they need a whole new way of thinking about 0:14:59.290 --> 0:15:01.990 making products at Google. And I think that requires a 0:15:01.990 --> 0:15:05.200 new leader. Police are now using GPT four powered body 0:15:05.200 --> 0:15:10.060 cams to turn audio into reports. That's great for transparency, 0:15:10.060 --> 0:15:13.720 I think. How about this for every stop, you basically 0:15:13.720 --> 0:15:16.900 publish the transcripts and that goes on an open website. 0:15:16.900 --> 0:15:19.570 That would be fantastic. And they could clean it up. 0:15:19.700 --> 0:15:23.390 For privacy reasons, but you can see the whole interaction. 0:15:23.390 --> 0:15:26.540 And then you also publish the video and the transcript. 0:15:26.540 --> 0:15:30.410 That would be community policing, especially in AI, because you 0:15:30.410 --> 0:15:34.130 can have AI monitoring the city, looking at every transcript 0:15:34.130 --> 0:15:38.210 of every stop. And you could characterize and actually have, 0:15:38.210 --> 0:15:41.120 you know, a way to profile like, how is this 0:15:41.120 --> 0:15:44.450 police department doing, are they doing more community management or 0:15:44.450 --> 0:15:50.210 more like predictive enforcement, or are they rude? Are they racist? Whatever. Okay. 0:15:50.210 --> 0:15:55.340 Apple eight open source LMS pushing the envelope for on 0:15:55.340 --> 0:16:00.020 device text generation with models up to 3 billion parameters. 0:16:00.020 --> 0:16:03.980 Apple doing crazy stuff in open source. Cannot wait for 0:16:03.980 --> 0:16:10.010 iOS 18 and llama three. Llama three is huge. Yeah 0:16:10.010 --> 0:16:15.170 800 llama three variations on hugging face insane. It is 0:16:15.170 --> 0:16:19.550 really good. It is really good. And I think it's 0:16:19.550 --> 0:16:24.320 getting so good that the people like OpenAI and anthropic, 0:16:24.320 --> 0:16:26.600 they need to be afraid. And Google need to be 0:16:26.600 --> 0:16:29.869 very afraid of this. Because what I like about the 0:16:29.870 --> 0:16:33.560 open source is it empowers the whole world to compete 0:16:33.560 --> 0:16:37.880 with OpenAI and Google, right? Because if they're like, look, 0:16:37.880 --> 0:16:40.760 this is how good it was for us. I mean, 0:16:40.760 --> 0:16:44.570 the barriers are essentially new techniques and the size of 0:16:44.570 --> 0:16:48.620 your infrastructure and the quality of the data that you have. Right. 0:16:48.620 --> 0:16:53.420 So new techniques, the crowd, the group of public researchers, 0:16:53.420 --> 0:16:57.110 they can jump ahead really quickly there. It's harder for 0:16:57.110 --> 0:17:00.290 them to compete with one of these big companies, of course, 0:17:00.290 --> 0:17:03.170 in the size of the cluster and how long you 0:17:03.170 --> 0:17:05.869 could train and stuff like that, because that's super expensive. 0:17:05.869 --> 0:17:08.030 And then the other one is data. It's also hard 0:17:08.030 --> 0:17:11.930 to compete there, but I think the technique part might 0:17:11.930 --> 0:17:16.129 be such a big lever that it allows some companies 0:17:16.130 --> 0:17:18.980 to really catch up and maybe even jump ahead in 0:17:18.980 --> 0:17:21.800 some cases. And then of course they'll get lapped the 0:17:21.800 --> 0:17:24.350 next time the big company moves. But I feel like 0:17:24.560 --> 0:17:30.619 community AI with open source models will move so often 0:17:30.619 --> 0:17:36.790 and iterate so often that speed it might. Get very 0:17:36.790 --> 0:17:41.710 close or even equal with or in some cases surpass 0:17:41.890 --> 0:17:44.500 the big models. Now, I do think the big models 0:17:44.500 --> 0:17:47.590 will have bigger leaps and they'll go way ahead, but 0:17:47.590 --> 0:17:50.379 it could be that open source catches up in the 0:17:50.380 --> 0:17:54.130 meantime just because of the speed of iteration. And I'm 0:17:54.160 --> 0:17:57.160 guessing here, but it's kind of just an intuition. Open 0:17:57.160 --> 0:18:01.990 voice is crushing voice cloning with the ability to mimic tone, emotion, 0:18:01.990 --> 0:18:07.780 and even cross-lingual speech. Snowflake just shipped another LM. Simone 0:18:07.780 --> 0:18:11.590 turns YouTube videos into blog posts US is finally getting 0:18:11.590 --> 0:18:14.619 its first high speed rail. Elon wants to turn Tesla 0:18:14.619 --> 0:18:18.820 cars into a distributed AI compute network, kind of like 0:18:19.180 --> 0:18:23.950 AWS for AI, but on wheels, and Tesla's Autopilot and 0:18:24.070 --> 0:18:29.590 self-driving are under scrutiny because there's been some deaths, and 0:18:29.710 --> 0:18:31.840 I think this is probably going to be the case. 0:18:31.840 --> 0:18:35.679 But I think over time, autonomous driving is going to 0:18:35.680 --> 0:18:38.859 be better than human driving because humans get tired, humans 0:18:38.859 --> 0:18:42.610 get distracted, humans are texting while driving or watching TikTok 0:18:42.609 --> 0:18:47.859 while driving, and that's probably worse. Again, this is an 0:18:47.859 --> 0:18:51.159 empirical question, so we'll have to see the actual data 0:18:51.160 --> 0:18:53.770 and how it plays out. But that that again, is 0:18:53.770 --> 0:18:57.460 my intuition. And yeah, example of this. We had someone 0:18:57.460 --> 0:19:03.700 arrested for vehicular homicide after distraction with his iPhone or yeah, 0:19:03.700 --> 0:19:06.580 I assume iPhone, but distraction with his phone laid to 0:19:06.580 --> 0:19:10.600 a led to a fatal collision with a motorcyclist while 0:19:10.600 --> 0:19:14.710 he was on autopilot, and Volvo brought in a Chinese 0:19:14.710 --> 0:19:17.949 EV to the US market. Okay, Stanford found out you 0:19:17.950 --> 0:19:21.820 could predict political leanings by looking at someone's face and 0:19:21.820 --> 0:19:25.510 specifically the size of your face. The bigger your face, 0:19:25.510 --> 0:19:28.929 the more likely you are to be a conservative. Evidently, 0:19:28.930 --> 0:19:32.469 according to this paper, and the smaller your face, the 0:19:32.470 --> 0:19:35.440 more likely you are to be a liberal. And I'm 0:19:35.440 --> 0:19:38.889 trying to figure out like what exactly that means. How 0:19:38.890 --> 0:19:40.960 do you measure the size of your face? I don't 0:19:40.960 --> 0:19:43.899 get it, but whatever. I'm sure it's in the paper. 0:19:43.900 --> 0:19:47.230 Article claims that 30% of kids 5 to 7 are 0:19:47.230 --> 0:19:51.790 on TikTok. 5 to 7. No bueno. I am very 0:19:51.790 --> 0:19:55.210 much in Jonathan Heights camp and you need to go 0:19:55.210 --> 0:19:58.030 read his book. I think he is touching on some 0:19:58.030 --> 0:20:02.919 really good stuff. FTC has banned nearly all non-compete agreements. 0:20:02.920 --> 0:20:05.950 This is great news. I think it's going to spawn 0:20:05.950 --> 0:20:09.460 even more innovation in AI and tech, and some are 0:20:09.460 --> 0:20:12.160 worried about IP theft, but if you move, you'll just 0:20:12.160 --> 0:20:15.490 steal the content. It's already illegal to steal from companies, 0:20:15.490 --> 0:20:18.010 so you don't need a law to keep them there 0:20:18.010 --> 0:20:21.820 so they don't steal. Like again, should be two separate laws, 0:20:21.820 --> 0:20:25.270 one for leaving and moving and one for stealing. Stealing 0:20:25.270 --> 0:20:28.419 is already illegal, so let's just stick with that. Supreme 0:20:28.420 --> 0:20:33.340 court is set to decide if cities can penalize the homeless. 0:20:33.340 --> 0:20:36.880 And I've got a vibe here. So I've been thinking 0:20:36.880 --> 0:20:40.389 about this for a while. I think if you're mentally ill, 0:20:40.390 --> 0:20:43.810 let's get you help or get you housed or whatever. 0:20:43.810 --> 0:20:46.150 If you're on drugs, let's get you treated and then 0:20:46.180 --> 0:20:48.880 see if you're mentally ill after that. If you're not 0:20:48.880 --> 0:20:51.580 one of those, let's figure out if you actually want 0:20:51.580 --> 0:20:54.280 to work. And if you do want to work and 0:20:54.280 --> 0:20:57.490 you're not one of these two, then let's get you help. 0:20:57.490 --> 0:21:01.780 Let's get you job support, training, all these different things 0:21:01.780 --> 0:21:05.649 like interview help, right. Try to get you on your feet. 0:21:05.650 --> 0:21:07.960 And if you're not any one of those, so you're 0:21:07.960 --> 0:21:11.020 not on drugs, you're not mentally ill, but you don't 0:21:11.020 --> 0:21:14.500 want to work and you don't want to participate in society. Well, 0:21:14.500 --> 0:21:17.560 let's get you somewhere where you're not bothering other people 0:21:17.560 --> 0:21:21.190 who are trying to be useful. And I think most 0:21:21.190 --> 0:21:24.520 homeless are actually dealing with one and two, so let's 0:21:24.520 --> 0:21:28.900 help them. That's our responsibility. And if it's not that 0:21:28.900 --> 0:21:32.170 situation and a person just really doesn't want to participate 0:21:32.170 --> 0:21:36.340 in society, they're like a Ted Kaczynski but not violent. Cool. 0:21:36.340 --> 0:21:38.890 Let's help them go do that somewhere else, you know, 0:21:38.890 --> 0:21:43.390 in a peaceful way. All right. Using AI for actual science, 0:21:43.390 --> 0:21:47.320 innovation and research. So AI is now helping physicists explore 0:21:47.320 --> 0:21:51.610 the vast possibilities of string theory. So this is great. 0:21:51.609 --> 0:21:55.210 I mean, think about what Einstein did. Einstein went from 0:21:55.210 --> 0:21:59.860 there not being a curved space time to just thinking 0:21:59.859 --> 0:22:03.250 of curved space time. I think about this all the 0:22:03.250 --> 0:22:06.940 time because I'm reading a lot of experimental physics, theoretical 0:22:06.940 --> 0:22:13.060 physics lately, and I really think that with my prompting skills, 0:22:13.060 --> 0:22:15.490 if I were better at the science, which I'm not, 0:22:15.490 --> 0:22:18.490 but if I were actually, I could probably just do 0:22:18.490 --> 0:22:21.250 this with AI without even knowing the science. If I 0:22:21.250 --> 0:22:24.850 had a big enough sort of resources to work with 0:22:24.850 --> 0:22:27.700 in the AI was a little bit further along. What 0:22:27.700 --> 0:22:30.250 you do is you essentially say, hey, look, here's all 0:22:30.250 --> 0:22:34.480 the different ways that people have crossed over and had 0:22:34.480 --> 0:22:38.619 these creative breakthroughs. You describe the science before Einstein, you 0:22:38.619 --> 0:22:42.250 describe what Einstein did, and you do that for multiple 0:22:42.250 --> 0:22:45.760 different places inside of science. Right. And what this does 0:22:45.760 --> 0:22:49.330 is it teaches it what it looks like to be creative. 0:22:49.330 --> 0:22:52.570 And you say, look, you're looking for more things like this. 0:22:52.570 --> 0:22:55.690 Then you feed it the current state of thinking with 0:22:55.690 --> 0:22:59.619 quantum physics and multiple universes and all these different things. 0:22:59.619 --> 0:23:02.859 You feed it all the math and everything, and you say, 0:23:02.859 --> 0:23:05.619 now that you have all the current do the same 0:23:05.619 --> 0:23:09.100 thing Einstein did, but do it for the current state now. 0:23:09.100 --> 0:23:11.710 And the idea is it should come up with really 0:23:11.710 --> 0:23:15.010 weird stuff. It's like, what if we're inside of a fishbowl, 0:23:15.010 --> 0:23:19.630 inside of a mormon, you know, simulation cluster running on 0:23:19.630 --> 0:23:22.900 Linux or whatever? It just starts making up things and 0:23:22.900 --> 0:23:26.229 then the humans can be like, oh, well, actually, I 0:23:26.230 --> 0:23:28.330 never thought of that. Let's go mess with that. And 0:23:28.330 --> 0:23:31.780 then even better, and this is something that Joseph Thacker 0:23:31.780 --> 0:23:35.020 and I talked about multiple times. It was his idea 0:23:35.020 --> 0:23:38.710 initially was like, how do you automate not just the 0:23:38.710 --> 0:23:41.290 having of the idea but the testing. And so we've 0:23:41.290 --> 0:23:43.840 been working on like an infrastructure that you could potentially 0:23:43.840 --> 0:23:47.140 do that with. And this is like the most incredible 0:23:47.140 --> 0:23:50.440 thing I think about AI because think about this, okay. 0:23:50.440 --> 0:23:55.149 Global warming, cancer, aging okay. People have to die at 0:23:55.150 --> 0:23:59.350 90 at 180, 90, 100. We describe all the reasons 0:23:59.350 --> 0:24:02.139 that we're dying at 89 to 100. And we say, look, 0:24:02.140 --> 0:24:04.959 we want to live longer. We want to live to 150. 0:24:04.990 --> 0:24:07.959 How can we live to 150? And we tell it 0:24:07.960 --> 0:24:10.510 to go and crunch and do what Einstein did, but 0:24:10.510 --> 0:24:12.940 for aging, and you give it all the genomes, you 0:24:12.940 --> 0:24:15.129 give it all the illnesses, you give it all the 0:24:15.130 --> 0:24:19.240 doctor reports, you give it all the autopsies, and it's like, okay, cool. Well, 0:24:19.240 --> 0:24:20.830 have you thought of this? If you do this to 0:24:20.830 --> 0:24:23.619 the mitochondria, it'll slow this down and you'll live to 0:24:23.619 --> 0:24:26.590 180 years old. And it could just churn out these 0:24:26.590 --> 0:24:30.280 things and even better, churn out testing for those things. 0:24:30.280 --> 0:24:32.260 So it could be like, look, I need a lab 0:24:32.260 --> 0:24:36.400 with the following, you know, beakers and petri dishes and 0:24:36.400 --> 0:24:38.350 we need to combine these different things. And I think 0:24:38.350 --> 0:24:40.810 I could make a drug that'll do this. You have 0:24:40.810 --> 0:24:43.750 to be careful, though, when you start building things that 0:24:43.750 --> 0:24:46.750 the AI asks you to build. But that's a separate 0:24:46.750 --> 0:24:51.550 talk show. But just think of that automated thinking combined 0:24:51.550 --> 0:24:55.270 with experimentation to to yield a result. I think it's 0:24:55.270 --> 0:24:59.740 super exciting. Okay, meticulous planning and luck led to capturing 0:24:59.740 --> 0:25:03.550 a breathtaking eclipse photo. This is the advantage of doing video. 0:25:03.550 --> 0:25:07.119 I'm going to open it up. Look at this thing. 0:25:07.119 --> 0:25:11.409 Look at this thing. Is that ridiculous? Like the wing 0:25:11.410 --> 0:25:13.870 is just like perfectly there and you've got the flare 0:25:13.869 --> 0:25:19.000 around it. I mean, that is that's gorgeous. Margaret MacDonald 0:25:19.000 --> 0:25:23.320 argues philosophical theories are more like artful stories than scientific facts. 0:25:23.320 --> 0:25:25.810 Quite a good essay. Here you are, what you read, 0:25:25.810 --> 0:25:28.510 even if you don't always remember it. I call this 0:25:28.540 --> 0:25:33.370 osmotic learning, which is basically just it absorbs into you 0:25:33.369 --> 0:25:36.760 in a way that's not attributable to the book or 0:25:36.760 --> 0:25:39.430 the video that you got it from. It just it 0:25:39.430 --> 0:25:42.070 soaks into you. Right. And I think I've read close 0:25:42.070 --> 0:25:45.250 to like 800 books and I feel myself getting smarter, 0:25:45.250 --> 0:25:49.420 but I can't point to why my models are being updated, 0:25:49.420 --> 0:25:53.139 which I call algorithmic learning. If you actually say Huberman 0:25:53.140 --> 0:25:57.310 said this, therefore I'm putting that into an algorithm to 0:25:57.310 --> 0:26:00.010 update what I do in the morning for my routine. 0:26:00.010 --> 0:26:03.430 That's algorithmic. But if you just consume the video, you 0:26:03.430 --> 0:26:07.690 don't write anything down, but you find yourself changing over time. 0:26:07.750 --> 0:26:11.859 I call that osmotic for like osmosis seeps in. And 0:26:11.859 --> 0:26:13.989 turns out the reason you're not landing that job in 0:26:13.990 --> 0:26:17.890 2024 might be because of ghost jobs, which are like 0:26:17.890 --> 0:26:20.860 basically fake jobs that nobody's applying for, and they're not 0:26:20.859 --> 0:26:23.890 real jobs. And you can actually get it. And got 0:26:23.890 --> 0:26:28.090 an argument here that your personality and creative application strategies 0:26:28.090 --> 0:26:31.960 can actually land you a job. So they're basically saying 0:26:31.960 --> 0:26:36.160 it's not about skills, it's about personality and creativity. And 0:26:36.160 --> 0:26:38.770 I think I've been talking about this quite a bit. 0:26:38.770 --> 0:26:42.189 It's about how you broadcast yourself. It's about how you 0:26:42.190 --> 0:26:44.710 present yourself. This is why I'm doing the whole human 0:26:44.710 --> 0:26:48.490 3.0 thing. So I definitely agree with this. I would 0:26:48.490 --> 0:26:52.090 characterize it a little differently, but yeah, pretty much right. 0:26:52.090 --> 0:26:55.990 All right. Continuous recording within ideas and analysis. So I 0:26:55.990 --> 0:26:58.960 think we're heading towards a place where everything is recorded, 0:26:58.960 --> 0:27:01.480 or at least where you can expect someone in a 0:27:01.480 --> 0:27:05.260 public space is probably recording you. And I like what 0:27:05.260 --> 0:27:08.889 that does for getting value out of conversations. But oftentimes 0:27:08.890 --> 0:27:11.230 the value is just having the talk in the first 0:27:11.230 --> 0:27:14.230 place and having it with somebody who you actually trust. 0:27:14.230 --> 0:27:16.990 So I worry a lot about the fact that people 0:27:16.990 --> 0:27:22.180 might be more guarded now when having conversations, just assuming 0:27:22.180 --> 0:27:24.730 that it is being recorded. So I expect there to 0:27:24.730 --> 0:27:28.390 be some sort of conversation indicator, like a red light 0:27:28.390 --> 0:27:30.820 or a blue light or something, or a green light, 0:27:30.820 --> 0:27:35.080 or a visual indicator on somebody's person that displays one that. 0:27:35.310 --> 0:27:39.389 They are recording or two, that they're okay with being recorded, 0:27:39.390 --> 0:27:43.290 or that they are not okay with being recorded. And 0:27:43.290 --> 0:27:45.240 I think we're going to have some sort of visual 0:27:45.240 --> 0:27:50.100 protocol like this indicates someone's comfort level. And I think 0:27:50.100 --> 0:27:52.379 venues will also have that, like you'll walk in and 0:27:52.380 --> 0:27:54.270 it'll be like a red light, like a stop sign 0:27:54.270 --> 0:27:57.149 or something, and it'll be like, not cool to record 0:27:57.150 --> 0:28:01.320 inside this building or Starbucks maybe does that for all internal, 0:28:01.320 --> 0:28:05.969 you know, Starbucks or whatever. Now question of enforcement that's separate. 0:28:05.970 --> 0:28:09.389 But the point is, the social contract I think is 0:28:09.390 --> 0:28:11.070 going to be there. It's going to need to be. 0:28:11.070 --> 0:28:14.070 So I think it's going to be interesting to see 0:28:14.070 --> 0:28:17.280 how that all affects how we interact with each other. 0:28:17.280 --> 0:28:22.200 One option is we stop talking and being like fresh 0:28:22.200 --> 0:28:25.500 and open with each other. Another option is we stop 0:28:25.500 --> 0:28:28.980 caring because there's no other ways to be embarrassed or canceled, 0:28:28.980 --> 0:28:32.250 because it's already happened to everyone already. And another option 0:28:32.250 --> 0:28:36.030 is that we break into pockets. So some places, like 0:28:36.030 --> 0:28:40.500 nobody's embarrassed about anything. Everyone's recording, everything is being transcribed. 0:28:40.500 --> 0:28:44.760 It's just radical transparency. And in other places it'll be 0:28:44.760 --> 0:28:48.630 like maybe Germany or like the Amish or something. It's like, 0:28:48.630 --> 0:28:53.940 no tech, no recording, complete privacy. And, you know, just 0:28:53.940 --> 0:28:56.310 like the polar opposite. And you'll have lots of different 0:28:56.310 --> 0:29:01.260 societies in between. Recommendation of the week. I recommend the 0:29:01.260 --> 0:29:03.720 book of the week or actually book of the month, 0:29:03.720 --> 0:29:06.960 actually The Courage to Be Disliked. This was the book 0:29:06.960 --> 0:29:09.360 that we just did the book club on. It was 0:29:09.360 --> 0:29:15.060 the book for April and it was wonderful. Absolutely loved it. 0:29:15.060 --> 0:29:20.610 Courage to be disliked. And it's about the teachings of Adler. 0:29:20.970 --> 0:29:25.560 It's called Adlerian philosophy or Adlerian psychology. Highly recommend the 0:29:25.560 --> 0:29:28.770 book and the aphorism of the week be who you 0:29:28.770 --> 0:29:31.500 are and say what you feel, because those who mind 0:29:31.530 --> 0:29:35.250 don't matter and those who matter don't mind. Be who 0:29:35.280 --> 0:29:37.530 you are and say what you feel, because those who 0:29:37.530 --> 0:29:42.540 mind don't matter and those who matter don't mind. Unsupervised 0:29:42.540 --> 0:29:45.060 learning is produced and edited by Daniel Meisler on a 0:29:45.060 --> 0:29:49.950 Neumann U87 AI microphone using Hindenburg. Intro and outro music 0:29:49.950 --> 0:29:53.010 is by zombie with the why and to get the 0:29:53.010 --> 0:29:55.080 text and links from this episode, sign up for the 0:29:55.080 --> 0:30:00.750 newsletter version of the show at Daniel meisler.com/newsletter. We'll see 0:30:00.750 --> 0:30:01.470 you next time.