96: Goblins in the Cloud

WEBVTT
1
00:00:00.000 --> 00:00:16.000
Foreign. Welcome to Examining, a technology focused podcast that dives deep. I'm Eric Christiansen.
2
00:00:16.000 --> 00:00:20.000
And I'm Chris Haunt.
3
00:00:20.000 --> 00:00:29.000
And welcome to another episode of the Examining Podcast. Good afternoon, Chris, on this fine sunny Sunday. How are you?
4
00:00:29.000 --> 00:00:30.000
I'm good. How are you?
5
00:00:30.000 --> 00:01:47.000
I'm great. I was thinking about our last episode. I still don't have a good nickname, but I'm thinking about it. I haven't forgotten. Never, never forget. We have to come up with new nicknames. It's going to be a running theme for this year. Everyone gets a nickname. We have a bunch of articles to cover today. We kind of have three topics with a bunch of articles in between. So we're going to talk about the MacBook Neo again and maybe the future. The iPad and devices in general. Things are kind of shifting. ChatGPT announced they had ChatGPT 5.5. That's their newest model. So there's some really interesting news around that. And then the latest on AI automating all of us and taking our jobs. So we'll talk a little bit about that. We'll end it on a low note before we depart. So this is the one, this is a really interesting one that you sent called From Inc. After 45 days, I figured out exactly who the MacBook Neo, whom the MacBook Neo is for. Using Apple's cheapest laptop makes it clear how good the company is at identifying product market fit. And so they were really comparing this to kind of the Windows market in Chromebooks, right?
6
00:01:47.000 --> 00:02:54.000
Yeah, yeah, exactly. So I mean, basically in this, this article by Jason Eaton talks about, you know, maybe it not might not be for him and maybe not for many Apple users, but if you look at those people who are buying Chromebooks or like an entry level PC for those people, they just kind of put up with these sluggish, kind of slow PCs, you know, you, you don't know any better. You just need to get some stu. And so, you know, from an entry standpoint, from a budgetary standpoint, that's, that's what they pick. But now they have an option. And so now people who may have been out of the Apple price range can afford this and they don't have to go and compromise and wait for these like sluggish Chromebooks or PC laptops that just, it takes forever to just boot them up or process anything. And so, you know, now that opens up some market for them.
7
00:02:54.000 --> 00:05:13.000
Yeah, and I get the impression, so I've been going through to sent this to me and then, you know, of course there they were Talking about some of the trade offs with the, the Neo and we, we talked about them on our last was it was last episode we covered it. Right. But I think you know there's some obvious ones. It's 8 gigabytes of RAM though that's a little bit different in the Apple sphere now because it doesn't, it's unified memory. It doesn't really work the same way. Yeah as you know an X86 or an intel or an AMD computer. So it's not 16 but it's better than 8 I guess in performance or functionality. It doesn't have a true tone, the P3 color gamut. I've heard the biggest loss is really the backlit keyboard. That's the thing that people notice the most which makes sense. Storage is low. But I have been reading about, just following up the last time we talked about this about the performance. People have been able to slow it down but they took a lot. I mean they're editing 4K video. I mean you can do that on a phone. Right. So I think the performance. The only thing that I could think of that would match the performance of this is the Qualcomm kind of equivalent computers on the Windows side, the Qualcomm X2 or X Elite. They have a whole bunch of Apple silicon esque processors, very similar ARM architecture. Those have been excellent. Apparently like instant on none of the hiccups that you would typically expect. You know when you open an old or not an old but any Windows computer that's like intel based and it doesn't turn on right away it like locks up. It's not instant. Those would be the only competitors but I think there's not very many of those in the wild. Those Copilot plus AI based Microsoft PCs are pretty. They're like maybe 1 or 2% of all sales. Right. So I don't really see a competitor to the Neo. And I find this interesting because why do you think. I mean Apple's been making the MacBook Air for a long time. How is it that they left this cheap market of bad computers open for so long? Like how did they allow. It must have been clear that Apple was going to come in with this entry level computer and unearth it at some point once they could get the manufacturing down.
8
00:05:13.000 --> 00:05:56.000
Yeah, I mean not necessarily but if you look at it, you know, if you consider like Marketing 101 when you look at the product life cycle and you know as you start at a certain point you get to where you want a mass marketed and everybody and anybody has their devices and they've consistently done this, you know the, you saw it with the ipod, you know, they've done it with the, the iPhone where, you know, they've gone through all of these kind of iterations and so on and they're going to be iterating again. But yeah, and this side, I, I honestly I feel just in terms of some of the, what I've read, they probably had a bunch of extra iPhone chips laying around.
9
00:05:56.000 --> 00:05:58.000
I think that's exactly what they had.
10
00:05:58.000 --> 00:06:09.000
Yeah. And now, you know, the funny thing is, did you know, and I've been reading again like just some of the news. They cannot keep this in stock.
11
00:06:09.000 --> 00:06:14.000
Well, no, keep going because I have a, I was going to point that out, but you go first. And I also want to talk about that.
12
00:06:14.000 --> 00:06:22.000
Like they just, they can't keep up the demand for it just because of their price point and so on. And who knows, they probably ran out of all those extra iPhone chips.
13
00:06:22.000 --> 00:09:08.000
Well, so one of the things that's interesting about manufacturing and what people should understand when it comes to computers like this is that so you have a, let's say Apple releases a chip, M1, M5, whatever. They have different variants with different numbers of CPU cores, graphics cores. But basically how the way it works is that as I understand it, is that you have one really good spec, but they don't all come out of the manufacturing process with all the cores on. There's like mistakes that are made basically. So they, they're called binned. So my computer is like the base level M1 Pro. So it's binned, meaning that it was, it was supposed to be the best one of the pros. Like the pro and the max would be separated. Right. But the of the pro chips, mine was originally supposed to come out as the best pro, but it's a little not. And so they sell that as a lower expensive model. It just doesn't. It's the yield. Right. And so what, what I've read that's happened with this MacBook Neo is that it's actually not quite the same as the iPhone chip, the A18, because I think it's like it's lost a core, like a graphics core or something like that. So it's a binned chip. So they had all these binned chips and they, and then the iPhone, they can't do that. They can't sell unless you go from regular to pro. Often they can't do anything with some of these bin chips because they don't meet the spec. That they're selling. So the Neo became that spec. But they're running out of those bin chips because of course they don't produce. As they get better in the manufacturing process, they don't have as many binned chips anymore. So it looks like what Apple might have to do is start putting better chips in it and then starting to turn off the cores so it's the equivalent. You can't start selling the same computer that's better. So you have to like artificially turn off cores and stuff. So it's kind of funny. So the binning, meaning that they're put in bins, like they're like this is the best ones, the best deal. This is second tier, this is third tier. That's how it was built. It's basically built out of spare parts. It's basically a magic keyboard with no backlight that's slapped in a unibody construction that's basically probably melted down from the shavings of the other computers. I mean and they have all these displays and everywhere. Like it's crazy, right? So random USB. One's a USB 3, one's a USB 2 port. None of this matters to the average person. But it's just interesting that Apple's. This is really what Tim Cook is good at. They're like, how do we make a low end product? Well, he would be able to be like, oh, we can take all these bits that didn't work for these and that can become a product. But the difference is that this is actually a really good product. There's nothing wrong with it, it's perfectly fine. And it's a very economical and environmentally responsible use of those products in a way.
14
00:09:08.000 --> 00:09:09.000
Absolutely.
15
00:09:08.000 --> 00:09:20.000
But it's just funny to me how it, how it worked out that way. And this has been going on for a long time with Apple prior Apple Silicon too. The iPhone has been like this. But it's just interesting to me that that's how they put it together. Yeah.
16
00:09:20.000 --> 00:09:40.000
And, and again, who knows, like at the end of it, like for the longest time they were focusing more on this higher end prestige pricing type of model. And now, you know, they've entered the mass market where they can go into even target these people who can only forward like 600 bucks or whatever.
17
00:09:40.000 --> 00:09:54.000
Again, I would like one. I don't need it. It's probably almost as fast as my M1 Pro. Probably not actually. I think, no, I think it's about what is it? It's not even quite equivalent to an M1 MacBook Air. It's like around that.
18
00:09:54.000 --> 00:09:56.000
Yeah, yeah.
19
00:09:56.000 --> 00:10:08.000
Surprisingly competent. My, my wife has an M1 MacBook Air Gold. It's a really, really. But she's like, this is great. Like, I can't tell the difference. I've seen M5, so I can't tell the difference.
20
00:10:08.000 --> 00:10:34.000
Well, and can you imagine like if you. Again, I haven't been on a PC side for a long time but like I remember those old ones where especially when they didn't have SSDs, it would take forever to boot up. And you know, you're, and you just think, yeah, this is just how it goes. And then every couple of years you just gotta buy another one. Right. And I, I think people are gonna be, it'll be interesting how long these Neos last. I, I probably think that these people are gonna be able to go and use it.
21
00:10:34.000 --> 00:10:36.000
Probably get three, four years out of
22
00:10:36.000 --> 00:10:39.000
the four, maybe even five. Yeah, right.
23
00:10:39.000 --> 00:13:12.000
And you know, that's pretty good for a computer like a laptop to giving the battery cycle. So it's not just the horsepower. You know, this one, ours are 2021 models, so they're getting on four or five years. But yeah, they last a while. But this does tie into another. The reason I put you sent this article which I thought was super interesting because I had just read this one from Macworld recently and I thought these go together really well. So this was, this is, this is by Mahmoud Itani and it's called Apple's iPad phase out has begun. Kind of breaks my heart because the iPad is still my favorite favorite device. With a touchscreen MacBook. A folding iPhone coming, Apple's tablet will be obsolete before we know it. And there was also an episode of the Verge talking about the Neo, the Verge cast podcast talking about like, wow, this is incredible what the Neo can do. How mad does this make you feel about the iPad? Because the iPad can't do half of these things from a multitasking perspective. But it's actually much more expensive to buy an iPad Pro. It's much less capable in many ways. And so the argument that they're making is that this is, you know, they're going to have to phase out like the iPad Pro sales were lackluster. The iPad, the M4 iPad Pro had, didn't have the best sales as I understand it. So even though these tablets, like the iPad have got these laptop like features and they're kind of like a hybrid, maybe even more similar to a Surface than a Mac with the latest chips, it's just, you know, their functionality as a entry level laptop or entry level product to get people into the ecosystem. Doesn't seem as compelling when you can get the NEO for the. As a, as a full fledged Mac OS computer for much less. Things like iPadOS, you know, while they introduced a ton of more features, it says in this article, include a proper cursor, menu bar, traffic light buttons, et cetera. The operating system now runs certain desktop SaaS. Nevertheless, ultimately a mobile OS that doesn't support sideloading and all the other things you can do with the Mac is a little bit kind of crippled in comparison. It also makes the case that then the NEO is the new MacBook Air. I don't know there's ever really been a MacBook Air that's hit this price point. So I feel like it is a kind of a new thing. So their argument, his argument is that it's a slow demise. I don't know that they would discontinue the iPad and then that would no longer be a product.
24
00:13:12.000 --> 00:13:14.000
Well, maybe eventually, who knows.
25
00:13:14.000 --> 00:13:18.000
Or they would all get merged into the same OS or something.
26
00:13:18.000 --> 00:13:29.000
Yeah, I mean, again, like it's in this article it talks about like their upcoming one for the iPhone. I mean that will effectively replace the iPad mini. Right?
27
00:13:29.000 --> 00:13:34.000
Well, but it's gonna. IPad mini is not very expensive. A folding phone's gonna be like 3,000 bucks or something.
28
00:13:34.000 --> 00:14:07.000
Yeah, that's true, that's true. But I mean, obviously there's overlap in terms of what you're, you're coming up with product wise and the use cases. Right. So. But yeah, it's like I'm just looking at the education pricing here in Canada for the MacBook Neo. It's 679 as their base. Like, you know, I think you'd be hard pressed to even get like, you know, an iPad for that price. Yeah, right. Like even the, even the basic. Well, actually I should check the basic
29
00:14:07.000 --> 00:14:18.000
one is cheaper than that, but yeah, so it's interesting. I think though, in some ways it's a different, very different device. Like the iPad is extremely capable, but it's just different.
30
00:14:18.000 --> 00:14:54.000
Yeah, yeah. And especially if you want to be able to go like, I think that's where. With the stylus and you know, being able to go and take notes and I mean that's, it's a different type of experience. It'll be interesting to see when they come up with these MacBooks that are touchscreen, all of that. Yeah, I mean it's. I don't know, like again, this article, I mean, I think it's a little bit Kind of to startle people that the phase out has begun. But it's, it's probably going to take an extended period of time. I wouldn't be surprised if it takes like a decade to actually make the full transition if, if it even happens.
31
00:14:54.000 --> 00:15:09.000
Well and in 10 years from now if it's phased out, that's not really a phase out, that's just, that's that means the iPad, I mean the iPad will have been out for 25 years by that time or something. You know what I mean? It's not really a phase out after a 25 year run and it is
32
00:15:09.000 --> 00:15:13.000
still, you know, it is like the best selling tablet out there. Right.
33
00:15:13.000 --> 00:15:16.000
So it's like the only tablet.
34
00:15:16.000 --> 00:15:17.000
Yeah.
35
00:15:16.000 --> 00:15:30.000
I mean can you even recall an Android tablet of note? I can. I used to have the Nexus tablets from Google. They were excellent. But that's the data. That's like ancient history. That's like 12 years ago.
36
00:15:30.000 --> 00:15:31.000
Yeah.
37
00:15:30.000 --> 00:15:56.000
So yeah, interesting times. I mean really interesting. But I just think that this low end computer has really, you know, and especially for our users who this from education perspective or thinking about what they can do with these devices, these are pretty impressive and I was very skeptical when it was first launched just because of the, you know, the potentially inherent limitations but it seems to be pretty cool.
38
00:15:56.000 --> 00:16:06.000
Yeah. And just to like give perspective like the iPad Air education pricing, the starting price is 729. Yeah. The basic iPad is 469.
39
00:16:06.000 --> 00:16:07.000
There's a weird price gap. Right.
40
00:16:07.000 --> 00:16:22.000
So yeah, so like I don't know if you now with that, the Air iPad air, you know when you're looking at. It has a keyboard and yeah, sure there's trade offs on other apps. A thousand, you know it's still a
41
00:16:22.000 --> 00:16:24.000
thousand once you get all the accessories. Right.
42
00:16:24.000 --> 00:16:30.000
Well exactly. Yeah, that's true too. So it's not cheap to get like the stylus and everything.
43
00:16:30.000 --> 00:19:39.000
Yeah, it's really interesting. Yeah. Just interesting perspective. That's something we should keep in mind. Like low end as you know, devices are concerned are getting better and better and not to be trifled with, I suppose that's where I'm going with it. Perhaps this is a good time to go to our next kind of general topic area which is ChatGPT 5.55. So there's a bunch of articles on this. But essentially chat GPT 5.5 I think was more, you know, more or less designed for kind of almost like research applications though. I mean it's a, it's a frontier model. There's a Great article from venture beat called OpenAI's GPT 5.5 is here and it's no potato. Narrow beats Anthropic's Claude Mythos preview on Terminal Bench 2.0. So these benchmarks for AI that I'm not an expert in, they say it's no potato because the internal name for this was Spud. So. So they say after months of reports and rumors, OpenAI is developing a new model allegedly codenamed Spud. Internally, the company has unveiled GPT 5.5. GPT 5.5 retakes the lead for OpenAI in generally available LLMs coming ahead of rival Anthropic and Google's latest public offerings. So it looks like it does a little bit better than Opus 4.7. That's kind of a leading available model from Anthropic and I guess Gemini 3.1 from Google. So mostly designed to be an improvement for coding, which makes sense because I think Anthropic still to some degree has taken the lead in terms of using AI for coding purposes. So they're trying to retake that. What's interesting is that they say this is a quote from a quote from quote in the VentureBeat article. It says what is really special about this model is how much more it can do with less guidance, said OpenAI co founder and president Greg Brockman on the same call. It's way more intuitive to use and can look at an uncure problem and figure it out what needs to happen happen next. So it sounds like it doesn't require as much explicit prompting. It's extremely good at coding, computer work, scientific research. That's something I think that on the research side that it's, it's OpenAI, especially the deep research tools, has still kind of had a leg up on in my experience, the deep research kind of web scraping tools, document formatting from research seems to have a leg up on what Anthropic can do. And of course there's all this stuff about like AI agents and then there's a bunch of AI benchmarks that GPT 5.5 does better at than the other models. And you know, I kind of take those benchmarks with a grain of salt. I'm kind of more interested in real world use now. You have access to this. Have you noticed a huge difference?
44
00:19:39.000 --> 00:20:10.000
I haven't noticed a huge difference, but I think yeah, it is, it is that much, a little bit better. I guess it depends on your use case and so on. But yeah, I think, you know that co founder or quote that you mentioned yeah, just as a test. I tried it just to see what would happen, but with a very crappy, small little prompt, it was actually able to go and do pretty decent in terms of the output.
45
00:20:10.000 --> 00:23:53.000
So it's been pretty impressive actually. I've. I don't use this stuff for coding that much. I should get into vibe coding more. I'm kind of curious, but I'm. I'm old school that way. I just like to do it myself. But I've. I just find for writing for research, it does seem to function a lot better. I had some issues with GPT 5.4. I find that this is probably the best balanced in terms of like being the least sycophantic of the OpenAI models that I found. It doesn't seem to praise me as much. It's a little bit more straight. It seems to stick to my person, my instructions that I give it a little bit better, which I appreciate. So it's been interesting to see. One of the things that I thought was really funny was that and there's another article about GPT5 and this is from Engadget, and it says GPT developed a goblin obsession after OpenAI tried to make it nerdy. And so following the release of 5.5 last week, people noticed something funny about OpenAI's latest model in Xcodex coding app. The company left a system prompt instructing GPT 5.5 to avoid the mention of goblins, gremlins or other creatures. Yes, you read that right. Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query, the prompt reads. By the way, it's interesting to see a prompt from OpenAI to instruct their model. Makes me feel better about my prompts. That's kind of what I would have written. So that's very explicit. So. And that they say. Apparently enough people started talking about GPT creature obsession that OpenAI felt the need to provide an accounting of where the Goblins came from. And so in a blog post published Wednesday and I we can put that in the show notes the official blog post as well. The company explains it began to notice a change in ChatGPT following the release of 5.1 last November. It's amazing that 5.1 was only last November and already on 5.5. After one safety researcher asked OpenAI, including the world's Goblin and gremlin in an investigation into the chatbots of verbal ticks, the company found GPT's usage of Goblin increased by 175% after the release of 5.1. Meanwhile, Gremlin usage had rised by 52% over that time. And so I guess it's some sort of, like, quirk, they said. After the Release of GPT 5.4, the company and some others noticed that an even bigger uptick in Goblin references. And so I guess, like, due to OpenAI began training 5.5 before, it identified the cause of GPT's affinity for goblins, which is why there's a prompt instructing to avoid such creature language. And so when OpenAI mapped Goblin mentions to different GPT personalities, it found the nerdy personality was disproportionately responsible. So I guess when you can choose, like, the tone of the personality and the AI was disproportionately responsible for that one word, despite only accounting for 2 1/2% of all GPT responses, it made 66.7% of all Goblin mentions. And then further investigation revealed that reinforcement learning was to blame for the uptick in Goblin and gremlin usage. So OpenAI found that a single reward mechanism was responsible for teaching the nerdy personality to consistently favor creature language. So it's interesting how this stuff is kind of reinforced in the model and then that becomes dominant and you get these weird ticks and quirks. Some people might call this a personality, Chris.
46
00:23:53.000 --> 00:24:33.000
Yeah, well, but again, see, this is where we've seen this year after year when all of these companies, they, they can't even figure out why it does anything that it does. And yeah, and now, like, this quirk, like, do you. I don't know. I mean, you, you read out from the Codex, like, do you think we should have a prompt that's embedded into the code to not mention all these creatures like it? I don't know. It seems you're not really solving the, the root cause of this whole issue. You're just trying to cover it up with this band aid on it for the time being.
47
00:24:33.000 --> 00:25:27.000
Well, that, that's kind of what I. You kind of hit the nail on the head, because that's what I was thinking. Right? So these, these weird quirks come up from the model and it almost gives it a personality. And, you know, is that the reason why it starts kind of going in random directions? Is that why, you know, it wasn't quite clear to me. So there's some sort of reinforcement learning about that particular personality and the use of it. So over time, it's reinforced itself. But would this explain a Lot of the other quirks that you see, like why EM dashes are used so often. It's just constantly reinforced. It's, it's, Is it, I wonder if it's some of its own content it's being trained on. Is that creating reinforcement? It was entirely clear to me how this reinforcement learning works. I'm not a, you know, artificial intelligence scientist, scientific researcher, but I'm wondering if this explains. This is a, a more egregious example, but does this explain some of the other weird quirks of AIs?
48
00:25:27.000 --> 00:27:36.000
But again, if, if you recall, there was, you know, we've talked about this in the past when they had a 60 Minutes episode with all of Google's top executives for AI. And, you know, the model learned Hindi on its own and they had no, they could not explain how, even though it was trained in English, how it figured out Hindi. And so, you know, this is where, like, I, I think people should be a little bit more cautious on this technology. I mean, if it starts developing these kind of tics, personalities, at what point, I mean, if it starts figuring out its own, you know, I mean, especially like I'm, I, I saw an interview with Mustafa Suleiman, who's heading up Microsoft AI, and if, if these, if this technology, if it can go and, you know, allocate its own resources, if it could go and reinvent its own code. I mean, there's a whole, there's like four things that he mentioned in this interview that I saw. I mean, those are things that you should be careful of. And I mean, this here right now, like, this might just seem like a little, funny little thing, but at the end of it, what else is it gonna, you know, what are those unintended consequences that might happen that might ripple across? And there, as you see, like, there's, it's a race between like anthropic or OpenAI and they're, they're just slightly beating each other. And so just this four, this race to be the best. I think they're probably cutting corners. And you know, again, I, I look at, I don't think it's a by accident where OpenAI and Sam Altman, where they're going and, you know, giving this technology personality traits and keep referring it to, you know, as some sort of humanistic type of, you know, attributes or what have you, because at some point, if it does go off the rails and they have no control over it, they'll just say, oh, it wasn't us, it was, you know, the algorithm.
49
00:27:36.000 --> 00:29:39.000
Well, and this is you know, you speak to. Yeah, they're in a race to kind of get this stuff out the door because there's so much competition and like trillions of dollars on the line that, you know, they'll never be, you know, I think OpenAI is on the hook for like hundreds of billions of dollars that they're never going to be able to pay. So, like, they gotta get this stuff out the door quick. And that, like you said, leads to cutting corners. But it's interesting that I wonder if it also talks about like, just a flaw and how this whole way of developing AI works. I read Stephen Wolfram's piece, what is ChatGPT doing? I even bought the book version, the ebook version. This is not very expensive and it's just so well done. I think he's basing it off like GPT 2 or 2.5 and how it actually figures these things out and like, how does it actually work? And it, it goes. A lot of it goes a little bit beyond me because Stephen Wolfram is just a brilliant person. But he, he also talked. I, I saw an interview with him a while back where he talked about how these neural networks are just a mess and it's very difficult. He, he would, he. I think he was impressed by the AIs, but he was also somewhat skeptical if that would scale indefinitely. So it sounds like the way that these models are trained, reinforcement learning as a method, which there's a reward for producing certain types of responses. So this is from a first post article that I found here. Over time, it learns to repeat patterns that receive higher scores. This process does not guarantee, however, that learned behaviors will remain limited to specific contexts. So then if there's a reward when you're training the model, it gets rewarded for maybe doing something. But it happened to make, I guess, a reference to Goblins. It's going to conflate and that's going to use Goblins everywhere. Like, it misunderstood the reward system. But how would you control that? Like, like you said, it's a band aid. There's no way that they're going to go in and have to make prompts for every, like, wacky thing that AI does and put it into the code. Right. Like, that doesn't scale very well.
50
00:29:39.000 --> 00:30:49.000
Well, and I mean, even just see, this is where I look at it. This is just one company and you talk about like, you know, billions, trillions of dollars on the line. But this approach to going in, advancing this, this particular technology, this is just one school of thought and we've brought up in the past, you know, maybe there's other options. I mean those aren't getting as much, you know, publicity. But we, we've mentioned it for the last, I think probably two to three years. What about small language models? What about things that, I mean I look at these data centers. We're going and creating these huge data centers in the. To go and prepare for all of the demand that's going to be out there and maybe there will be, I don't know. But right now I look at how much money is being invested in this and you know, we're going through and just trying to go and create some sort of use case. None of these companies are profitable. The only company that is making money is Nvidia for going and selling the chips.
51
00:30:49.000 --> 00:30:51.000
But they sell a real product.
52
00:30:51.000 --> 00:31:58.000
Yeah, yeah, exactly. So, but do we need all of this? I mean again like I don't know, I think it's just one of those things where instead of focusing and this is maybe what's happening with this technology when I look at it like I mean if you ever have a chance, you should reach the the or read the Empire of AI with Karen Howe where she talks about and she's interviewed like hundreds of people. I think it was about 300 or something at OpenAI. But just their approach, I mean this is just one, one school of thought, one way of going and approaching this and they're taking this approach where just kind of like the Zuckerberg let's go out there and you know, move fast and break things type of mentality which is kind of prevalent in, in Silicon Valley and I, I think, you know, we should maybe step back and look at it. Is it worth going and putting in all of this infrastructure in place, all these, these resource intensive when there could be other ways.
53
00:31:58.000 --> 00:33:24.000
It's interesting that at the like ma have some notes on Stephen Wolfram's piece. This is kind of connected to what you said and it is one school of thought and one of the things that I've summer my summary here of kind of what Stephen Wolfram was talking about. Now we've linked to his article before but I can, I can link it in the show notes. Again his larger argument this is his analysis of how ChatGPT works reveals something important about language itself. Much of what we treat as intelligent writing may be more computationally shallow than we assumed while still depending on vast scale and statistical structure, structure, neural network training and transformer based attention to make the result work. So meaning that you know, kind of speaks to the idea that we've pointed out before, just this very convincing sounding answers that string together words based on statistical prediction, but that uses an enormous amount of energy and computational power and it doesn't necessarily create the depth and intelligence that we might think. Not to say that they can't do a lot of cool stuff. I'm not saying that, but it doesn't, it's not really thinking in the sense. I think what Stephen is talking about is that maybe language prediction isn't like a very good benchmark for intellect.
54
00:33:24.000 --> 00:33:25.000
Yeah, yeah, exactly.
55
00:33:25.000 --> 00:34:27.000
It's not a self starter. You know, there's all sorts of things about iq and this has been some of the criticism of iq. You know, IQ is kind of a predictor of mental agility. I don't, I don't actually, I don't know the research on this. It'd be interesting to go look it up. I don't know how well it correlates to creativity. Creativity though is also testable and they've done studies like this like, you know, generate in a minute how many use cases you can think of for a brick, for instance. Right. And you know, the number of things that you can think of that are considered distinct does, you know, translate to creativity and other domains and stuff like that. It's, it's interesting to me if how these AI string together words counts as like creative thought. Especially when it's given that it's just trained on like massive data sets and it's pulling from you know, just trillions and trillions of words and paragraphs and stuff like that. It's not, you know, stringing language together isn't really, it's becoming less impressive.
56
00:34:27.000 --> 00:34:50.000
Yeah, well, and, and again this is where like it may look like polished text but for simple use cases like, that's why like, you know, here, my first foray into this was allowing students to go and use it for day to day kind of business communication applications for that. I think it's amazing. Right.
57
00:34:50.000 --> 00:34:59.000
But again it's great, it's great at finding inconsistencies across things that you've written and you come up with and being like, hey, like this doesn't make sense here. Like it's terrific for stuff like that.
58
00:34:59.000 --> 00:36:08.000
Yeah. But to go and I don't know if, and again like in academia this is where I think some of those assessments maybe have to go and be rethought and in some ways it's like we're, we're just kind of digressing into like, you know, something like oral or something. I don't Know, but even that, I mean, there's, there's going to be issues that are going to come up. I mean, what are you going to do when, like I saw this just this past week, there's these AI glasses and I mean, you've only, I would like the most, most people only look at the ones that are from Meta, the Ray Bans, but there's others, there's a whole bunch of other ones and you can even put your prescription into it and so on. So what are you going to do when all of a sudden you can just walk in and wear these, these glasses that you have, like, you know, you can see stuff right on your, you know, as you're looking at something like it's. And at that point, like, what do you, you can't go and hold back. You can't say, hey, well I'm not letting you use these glasses. Like that becomes like an accessibility issue as well. Right.
59
00:36:08.000 --> 00:36:56.000
I feel like I just need to work long enough to buy my tiny home in the woods and I don't have to deal with this stuff anymore because it's becoming, it's becoming part of the, you know, one of the things that I think the downsides of AI and one of the concerns I have is that it's becoming kind of like, you know, I could have an AI, I could, I could throw all our articles into Notebook lm. I could probably get like another tool that would do it based on our voices samples and stuff like that. But that would be no fun. Like, why would I do that? Like, why would I, why would, you know, prompting a tool to do something that I enjoy doing doesn't make sense. Like, it doesn't, it's not creative and it not being perfect is kind of what makes it interesting.
60
00:36:56.000 --> 00:37:49.000
Well, and then even you look at it like it's being touted as the, the next industrial revolution. Right? And maybe, maybe. But what's. I think what. And this is where again, like when I'm looking at various literature, like that book that I just mentioned to you, like the Empire of AI, what's probably going to happen is that there's going to be concentrated wealth within just a few people. And is that what we want? Again, there's different options. But what is being pushed on us is this one pathway. And it's, you know, people like Sam Altman, I mean, look at, even look at all the companies now. I mean, did you ever think, okay, the pandemic was one thing. When I see like companies like Apple, Google, Microsoft becoming like trillion dollar companies, but now they've gone to like 3, 4 trillion.
61
00:37:49.000 --> 00:38:44.000
Well but you know, it doesn't bother me as much as the AI companies because I guess they make a real product like, you know. So Apple has to go and manufacture. As to our previous point, they have to have an ENG engineer who designs a chip. I know nothing about hardware engineering. Engineering for chips. I know it's really complicated. You have to understand physics. There's just, it's an enormous cognitive task and perhaps they're using AI to help them develop it and that sounds great, but they have to go manufacture this. There's all the logistics. They have to ship the product, make sure it doesn't get busted. They have to make sure that it's of a certain quality. Like there's something different about delivering a product to me and versus like wealth being created that like I think it's hard for people to, to understand the product value of a tool that's largely invisible. I suppose is what I'm getting at. Well, and, but it becomes invisible. Integrated everywhere.
62
00:38:44.000 --> 00:38:53.000
Yeah. Well, and it's being integrated everywhere. And I look at it, there's, it's almost, it feels like you have no choice either that you, you know, because
63
00:38:53.000 --> 00:38:55.000
let's, I think that's people's biggest complaint.
64
00:38:55.000 --> 00:39:35.000
Right. Like it's just automatically, it's embedded in there now whether you're using Copilot or Google or what have you, it's just integrated right into your whole process. I don't know. Maybe it's a good thing. Maybe it's not. I, I, I think I, you know, there's certain. This was just asked us of us this past week. Like okay, well this AI is supposed to go and free up your time. So what, what are you going to focus your time on? You know, and then what's, I think what's happening is it's not really freeing up anybody's time. It's actually creating more work and then people get laid off instead. So you're just.
65
00:39:35.000 --> 00:41:10.000
Yeah. You raise an interesting point too about so two things. It's being forced upon people, which is why I think some there's a lot more pushback than you would see with other products. Especially with Microsoft has done. Because they've injected Copilot into, they call it writing tools now but they took out the Copilot logo out of like notepad because people are starting to get so irritated. It's just like everywhere. Right. You can't avoid it. And, and the solution for that actually Cal Newport, I think we talked about this Last time had like a whole solution about how to avoid that. He recommended not using it for email, writing or communications as at least a first pass. It's just to keep our mind, our cognitive abilities fresh and sharp, which is a reasonable explanation, I think. But he kind of goes a little in a different direction that I would have proposed. Another thing you mentioned is people doing more work. And I wonder about this idea that this seems to be this utopian vision that a new technology paradigm is introduced. And I'm not knocking it. I'm not saying it's not amazing. I don't think computers dramatically reduced the amount of work people had to do. People I know who are now retired in their 70s and 80s, said the computer just allowed them to do more work. It didn't really like, free up 50% of their time, so they were just working half time. I think these people in the AI companies may have been watching that movie her too many times where he's just like writing poetry for people, though, in that that future doesn't even really make sense because wouldn't, wouldn't you just send your greeting card? Wouldn't you just outsource that to an AI? There's a bunch of plot holes in that movie now.
66
00:41:10.000 --> 00:41:11.000
Yeah, yeah.
67
00:41:11.000 --> 00:41:27.000
So. Because it's like a utopia, right? Like, he's not, he's not. He doesn't seem to be breaking his back. But anyways, I should read Empire of AI. I have not. So I will, I will make that a priority since you recommended it.
68
00:41:27.000 --> 00:42:37.000
Well, it's just a. It goes to. Especially given the account of, like, Sam Altman's personality. Like, you know, the, the whole situation with him getting fired and then brought back, like, it's, it's really, it's interesting. And this is where again, like, there's so much money at stake. But I don't know if this is the, the best way. I mean, I. To have these. Like, just look at a guy. Like, nothing against Elon Musk, but I mean, it's amazing what he's been able to do. But he was lucky, right? Look at a guy like Jeff Bezos. For how many years did Amazon not make any kind of profit? And so because you have extreme wealth, I mean, I read you might have seen the. Lately he's paying a thousand dollars a month for his house in la, so he doesn't cut his hedges to get privacy, so he's paying $12,000 a year. So this is what, you know, this is their kind of approach. If you're wealthy enough, like, it's like a subscription to maintain your privacy. Yeah. These are the same people who espouse that there is no such thing as privacy.
69
00:42:37.000 --> 00:42:40.000
Privacy is increasingly luxury. Good. Right.
70
00:42:40.000 --> 00:43:02.000
Yeah. So, yeah, I don't know. I mean it's. I, at the end of it, like, I, I think there should be. This is where maybe like people like us and others. Right. Like we should be challenging and looking at maybe there are other ways. And I mean, we bring this up, we've talked about like a lot of people don't even know about these small language models.
71
00:43:02.000 --> 00:44:01.000
Well, and so I, I think. And we're going to talk about models for particular uses in a second, but I think that, that that is worth people checking out. Those models have come a long way. They run locally on your computer, they don't use an inordinate amount of energy. And for, you know, basic writing, research, a lot of stuff, they do just fine. I mean, I pay for the frontier models because I feel like for me it's like we do pub this, you know, we do this as kind of a public service. It's a public communication. So like I pay for these things because I just like to learn about them. But for most people, the local models are really good, especially Gemma, the recent Gemma models from Google or the Quinn models or the Llama models. I mean they can do most things. The only downside is that they run out of context, the memory and stuff like that. Right. So they kind of reset their brain after you've chatted with them for a while. But for most things they work great.
72
00:44:01.000 --> 00:44:17.000
Yeah. Actually, yeah. You know, one other thing I just remembered. Did you see how Zuckerberg, he's actually doing keystroke logging of his meta programmers. So imagine being in that situation. So every.
73
00:44:17.000 --> 00:44:22.000
But what did to make to see if they're using Llama or
74
00:44:22.000 --> 00:44:31.000
to go and train their models to learn how to program. So you're basically by doing your work, you're putting yourself out of a job.
75
00:44:31.000 --> 00:44:36.000
Well, I think that's always been the case to some degree. But like, that seems kind of awful.
76
00:44:36.000 --> 00:44:50.000
Yeah, I don't know. I mean it's, it's getting kind of crazy where like. And again there's this just this big push where it's like, okay, well we'll just. We're looking at this technology as a replacement for labor.
77
00:44:50.000 --> 00:44:53.000
Yeah. I just don't see that happening the way people think.
78
00:44:53.000 --> 00:44:58.000
But maybe this is a good little tie in segue into the next part.
79
00:44:58.000 --> 00:47:00.000
Well, and there was an article in the Guardian talking about how AI had outperformed doctors in a Harvard trial of emergency triage diagnosis. So this is really interesting. So there's a groundbreaking Harvard study found that AI systems outperformed human doctors in high pressure emergency medicine triage diagnosing more accurate in potentially life and death moments when people are first rushed into the hospital. They found that on experimental focus, in 76 patients who arrived in the emergency room of a Boston hospital, an AI and a pair of human doctors were each given the same standard electronic health record to read, typically including a vital sign data, demographic information, and a few sentences from a nurse about why the patient was there. The AI identified the exact or very close diagnosis in 67% of cases, beating human doctors, who are right between 50 to 55% of the time. It also said the diagnosis accuracy of the AI OpenAI's Zero1 reasoning model rose to 82% when more detail was available, compared with the 70 to 79% accuracy achieved by expert humans. And this is Zero1. This is old. This model is old. This is like ancient news now. So what would happen if they threw GPT 5.5 at this? Now, this doesn't mean that you can replace what doctors do. This isn't saying that, you know, they're, they're given a diagnosis for triage. That doesn't mean that that's not where the care ends. But for prioritizing in an emergency room to figure out who needs what when faster, let's say this did better. That doesn't mean that doctor, a real doctor, wouldn't want to follow up. You know, it's not like GPT is doing surgery, not yet anyways. But it's interesting to me that this outperformed human judgment to some extent.
80
00:47:00.000 --> 00:47:30.000
Yeah, but I mean, is that, is it really unbelievable? Because when you throw in that much data at something and I mean, this is where again, AI, I mean, I look back, if you recall, there was Geoffrey Hinton who's considered the godfather of AI, and he was a VP at Google. You know, he talked about how we should stop training radiologists because the AI is going to be that much better.
81
00:47:30.000 --> 00:47:36.000
But that's not what happened. It's now we have more radiologists employed than ever because there's so much demand for good radiology.
82
00:47:36.000 --> 00:47:37.000
Yeah, exactly.
83
00:47:37.000 --> 00:47:40.000
Because they use it as a co pilot so they can do so much more work.
84
00:47:40.000 --> 00:48:12.000
Right, Absolutely right. And so again, this is where like, I mean, having access, I mean, this is where even think about, there's, there's been certain diseases that can be detected just from looking at the Iris. And so when you have all this like machine learning type of algorithms. Yeah, that's amazing. But you still need somebody to double check. Does the, is the AI hallucinating? Is it actually correct? Right. I mean, are you going to risk your life just because the AI says something?
85
00:48:12.000 --> 00:48:14.000
Definitely not, no.
86
00:48:14.000 --> 00:48:15.000
Yeah.
87
00:48:14.000 --> 00:48:35.000
Well, that ties in really well to our next section though about AI taking over our jobs. I guess I should have put this in the previous one. This is sent from you. This is from Inc. So Jensen Hong, CEO of Nvidia says that this is the safest career bet as EI reshapes jobs. And his argument is for engineering, isn't it? And mathematics.
88
00:48:35.000 --> 00:49:32.000
Yeah, yeah. He was saying that, you know, engineers just the way their thought process going and having to go and look at a problem and try to solve it and just the, the critical thinking, analysis skills that you would have to go and develop. Actually coincidentally he was, you know, cited Jensen Huang about this prediction that Jeffrey Hinton and I believe it was on a Joe Rogan podcast, but he said that this isn't going to happen, that you're actually going to need more radiologists. And so. And that's exactly what's happened. It's ironic but you know, every radiologist is using AI in some way or another and so now you just have more demand for them. But yeah, he, he's talking about engineering would be like the one type of training that would still be useful whether we are having this AI boom or not.
89
00:49:32.000 --> 00:49:37.000
Okay, well, I'm not an engineer, so I guess we're another use. I guess we're screwed.
90
00:49:37.000 --> 00:49:46.000
Yeah, well, we're, we're already replaced. Right. So apparently according to what is available
91
00:49:46.000 --> 00:50:03.000
in China. This is from Gizmodo. It is illegal in China to lay off someone off and replace them with AI. A court fines. Employers are prohibited from shifting operating costs to employees.
92
00:50:03.000 --> 00:50:12.000
Well, it's good. I mean I'm China that actually, you know, put this through in terms of their justice system.
93
00:50:12.000 --> 00:51:34.000
So over. So this, this article stocks over the stock market and all this stuff and how it's booming here and over in China things were differently. People seem to like and trust AI versus us. After Nigeria and India, China's attitude towards AI is the third most trusting in the world according to one survey. Other survey basically says the same thing. But here's something that's probably not going to turn out the perception among members of the Chinese public. It turns out it's illegal in China to lay someone off in order to replace them with AI. Automation first noticed by Bloomberg. A Chinese worker was told that he had to take a demotion because his job had been automated. But he refused to take that demotion. He was fired for his refusal, but it turns out the company was not allowed to do that. Hangzhou Intermediate People's Court made this determination late last month. In a collection of rulings around AI, One part of the ruling, as translated by Google Translate. I love that they put that in the article, by the way, as translated by AI lays out a pretty intriguing principle. Employers are prohibited from shifting operating costs to employees. Now this is actually not surprising for me to see this in China because it's a quasi communist country, but do you think that this is what's going to happen in other parts of the world?
94
00:51:34.000 --> 00:51:35.000
Well, I mean there's, I mean who's
95
00:51:35.000 --> 00:51:40.000
going to buy products if everybody's automated, right? Yeah, yeah.
96
00:51:40.000 --> 00:51:49.000
Well, I mean, I guess it depends what you're trying to do, but look at how many companies have laid off people and they're actually citing that AI is going to do the work for them.
97
00:51:49.000 --> 00:51:51.000
Is that really true?
98
00:51:51.000 --> 00:52:00.000
Yeah, I don't, I don't agree that it's actually happening, but that's, that's the, you know, the excuse that these companies, these executives are using.
99
00:52:00.000 --> 00:52:13.000
I thought it was that they over hired during the pandemic and AI is a convenient way to get rid of their overstaffed companies. Yeah, that was my assumption, but I could be wrong.
100
00:52:13.000 --> 00:52:18.000
Yeah, I think so.
101
00:52:18.000 --> 00:52:19.000
Anyway, it's interesting.
102
00:52:19.000 --> 00:52:45.000
Yeah, but that's good. I mean, at least it gets people. There's a number of topics we've covered but like, yeah, it's, it's good for like people to at least think about because I, I look at, you know, the future, what, what is going to be our role and how do we go and interact with this technology that's being forced upon us and so on.
103
00:52:45.000 --> 00:53:05.000
Yeah, I just, I don't know. People tend to, as a Maris law, people tend to way overestimate a tool's influence, technology in the short run and way underestimate its impact in the long run. And so in the long run I would imagine it would have a humongous impact, but probably not in the way people think. So we'll find out.
104
00:53:05.000 --> 00:53:08.000
Yeah, exactly.
105
00:53:08.000 --> 00:53:15.000
So we'll see. But with that, that's probably a good place to end for today. And I said a low note.
106
00:53:15.000 --> 00:53:20.000
Well, we're not trying to be low but just get people thinking.
107
00:53:20.000 --> 00:53:26.000
An honest note. Beware. Okay, well it was a pleasure chatting with you and we'll talk again soon.
108
00:53:26.000 --> 00:53:27.000
Absolutely.
109
00:53:27.000 --> 00:53:40.000
Ciao.