375 lines
31 KiB
Plaintext
375 lines
31 KiB
Plaintext
|
|
Episode: 3379
|
||
|
|
Title: HPR3379: Linux Inlaws S01E34: The one with the intelligence
|
||
|
|
Source: https://hub.hackerpublicradio.org/ccdn.php?filename=/eps/hpr3379/hpr3379.mp3
|
||
|
|
Transcribed: 2025-10-24 22:25:44
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
This is Hacker Public Radio Episode 3,379 for Thursday, 15 July 2021.
|
||
|
|
To its show is entitled, Linux in laws S0134,
|
||
|
|
the one with the intelligence and is part of the series Linux in laws it is hosted by Monochromic
|
||
|
|
and is about 45 minutes long and carries an explicit flag.
|
||
|
|
The summary is, part four of the three part miniseries on deep learning and artificial intelligence.
|
||
|
|
This episode of HPR is brought to you by archive.org.
|
||
|
|
Support universal access to all knowledge by heading over to archive.org forward slash donate.
|
||
|
|
This is Linux in laws, a podcast on topics around free and open source software,
|
||
|
|
any associated contraband, communism, the revolution in general and whatever fans is your
|
||
|
|
vehicle. Please note that this and other episodes may contain strong language, offensive
|
||
|
|
humor and other certainly not politically correct language you have been warned.
|
||
|
|
Our parents insisted on this disclaimer. Happy mom?
|
||
|
|
Thus, the content is not suitable for consumption in the workplace, especially when
|
||
|
|
played back in an open plan office or similar environments, any minors under the age of 35 or
|
||
|
|
any pets including fluffy little killer bunnies, you trust the guide dog, a lesson speed
|
||
|
|
and QT rexes or other associated dinosaurs. Welcome to Linux episode a season one episode 34
|
||
|
|
the one with the intelligence. Martin, good evening. How are things?
|
||
|
|
It's evening Chris. Things are not bad, not bad, some shining.
|
||
|
|
Excellent. Excellent. Fixed.
|
||
|
|
What could possibly go wrong apart from Brexit, the Vogue and Slanding and whatever comes to mind?
|
||
|
|
Oh yeah, Mr. Brian Johnson stepping back.
|
||
|
|
Stepping down, sorry. You mean Boris, please. Sorry, Boris, yes.
|
||
|
|
I was confused. Apparently, yes, there is. Brian Johnson, isn't that one of the queen boys?
|
||
|
|
Maybe wrong. Anyway, doesn't matter. My understanding is basically that speaking of Mr. Johnson,
|
||
|
|
that Boris has married recently. Yes. Any thoughts on this again, by the way? Yes, I know.
|
||
|
|
Well, there was this is some debate about why he was married in church for the third time,
|
||
|
|
but I mean, I reckon he was he was married in a proper Protestant church. So
|
||
|
|
yeah, but he has a certain reputation when it comes to people of the well, I mean, if
|
||
|
|
Laura is anything to go by, he's not the first one being married a couple of times. Henry,
|
||
|
|
the eighth comes to mind. Speaking of which. And I mean, yes. And yes. How are you?
|
||
|
|
Can't complain actually. In contrast to current, to to to to to wishes rumours, I haven't been married.
|
||
|
|
Once not twice, not three times. If it's any consolation. Apart from that, I mean, one piece. I
|
||
|
|
mean, this country is slowly getting back to something called not nearly close to normal,
|
||
|
|
but that's different story. So let's see. Yes, people, we are recording this on the seventh of
|
||
|
|
September 2035, if not completely mistaken. I might be wrong the date, but let's not worry about
|
||
|
|
this, but way, but let's let's go back to way more save it to to way, save for rounds.
|
||
|
|
Namely, the topic of tonight's episode, button. Of course, this is the fourth part of the
|
||
|
|
are we on four? Yes, we are. This is the fourth part of the three part ministries of artificial
|
||
|
|
intelligence. Yes. If you recall correctly, no, exactly.
|
||
|
|
Oh, does that mean? Yes, indeed, very much so. Yes, to celebrate the fact that everybody who is
|
||
|
|
listening has survived the three parts so far. Obviously, part mini series. We have a special
|
||
|
|
jewel tonight called GPT. Martin, why don't you explain what there is? For the few people who are
|
||
|
|
listening, who do not know what GPT is? Okay, but it's made a lot of noise in the press for
|
||
|
|
various reasons. GPT is standing for generative pre-trains transformer and AI model.
|
||
|
|
Is that one of these toys that look like a robot but then you form it to a car or something or
|
||
|
|
the other way around? That's indeed a transformer, but yeah, it's not one of the plastic ones.
|
||
|
|
I see. Okay. Yeah, so that's what it stands for. It is mostly known for its, well, most you know,
|
||
|
|
it isn't known for its language capabilities in various iterations of it, but why don't we go
|
||
|
|
through a little bit of a history of that? Excellent idea. Yes. There was a company called
|
||
|
|
Cyberdyne, if I'm completely mistaken, back in the 80s. Yes, but that's probably beside the
|
||
|
|
point, which I'll decide. The whole thing goes back to something called OpenAI, if I'm not completely
|
||
|
|
mistaken, right? That's right. Yep, yep. So, I mean, we, part of the miniseries, we talked about
|
||
|
|
various frameworks, how do I know what to work, and so on and so on, but obviously the
|
||
|
|
whole point of all this stuff is to have it and an application for it, whether it's
|
||
|
|
computer vision or object recognition, whatever, classification, library, or
|
||
|
|
another field of interest is language, right? That's, by the use, AI, I won't know, so by the use
|
||
|
|
in this field, otherwise, Ken will be very happy again. I mean, before we go into the native
|
||
|
|
reveal, it's off GPT, exactly GPT-2 or GPT-3, whatever, maybe it's worth talking a little bit
|
||
|
|
about the background of OpenAI, because they have some illustrious founders, right? If I'm completely
|
||
|
|
mistaken, a guy called Alan Musk was one of the initial founders? He was, he was, I don't know,
|
||
|
|
but Alan or Elon, but yeah. Yeah, the one with cars, right? I'm on other things.
|
||
|
|
Yeah, that's the one. That's the one, yeah. Yes, that's one, that's the one,
|
||
|
|
others. But Jeff, but Jeff Bezos is not a founder of said venture. He isn't, he isn't.
|
||
|
|
I mean, one of the reasons it, well, one of the reasons he came about is to kind of
|
||
|
|
stop reading the right word, but have an alternative to Google's deep mind, right?
|
||
|
|
I think that's something, in some of its history.
|
||
|
|
I came about, so far in 2015, Alan defected in 2018?
|
||
|
|
Yes, could be, could be. I didn't believe, well, that piece, but
|
||
|
|
I think you have more insight on the reasons why he left. Presumably, he need the money,
|
||
|
|
he need the money for SpaceX and Bitcoin. Ah, maybe I'm wrong.
|
||
|
|
Well, yeah, but SpaceX is going these days, but yeah, they have been a lot more.
|
||
|
|
I mean, yeah, I mean, to be, yeah, to be much more serious, we're recording this almost
|
||
|
|
middle of June 2021. And Alan has just issued a very interesting comment over the weekends
|
||
|
|
that they're not too happy with said Elon Musk, destroying people's lives and causing a
|
||
|
|
little bit of a stir in the Bitcoin markets. But I reckon if current law is anything to go by,
|
||
|
|
ah, well, I'm going to put this diplomatically. The US government funding that has been poured into
|
||
|
|
into something called Tesla could be used for different purposes other than OpenAI. So I reckon
|
||
|
|
this is how this whole Bitcoin debarker came about. This is, of course, pure speculation.
|
||
|
|
I might be wrong, but given the fact that Bitcoin took a hit of what 20% of the overall value
|
||
|
|
after the recent China against offset Mr. Musk, this doesn't come as a surprise to split this way.
|
||
|
|
Yeah, I mean, there's also some theory about China. I don't know, I don't know,
|
||
|
|
I'm told it's an area, but yeah, yeah, that's right. He doesn't seem to poke around mainly with
|
||
|
|
the weak coin or whatever it's called. Do you think that China is behind Tesla?
|
||
|
|
No, no, no, no, no, no, no. Behind Tesla, behind the...
|
||
|
|
As a Musk, he looks a bit chummy, doesn't he? Well, I mean, he's clearly, clearly,
|
||
|
|
inside with any inventions.
|
||
|
|
Sabadine, if you're listening with us, we're still looking for sponsors to address a sponsor at
|
||
|
|
Linux in North, aren't you? If you're so inclined, anyway, it doesn't matter. Okay, back to the
|
||
|
|
topic at hand. So, the company behind something called GPT and the architecture went through a
|
||
|
|
couple of versions, right? Yeah, so one more thing on open. It was really initially
|
||
|
|
more of a... Which is why we actually talked about today, right? It was more of a
|
||
|
|
not so much a for-profit organization, more of a research organization for AI,
|
||
|
|
and it must have been quoted in the case of, to at least researching this field,
|
||
|
|
means that we can use it for the good of humanity and so on, which is also part of
|
||
|
|
focus on all this stuff, then it would be for everybody. So, yeah, but in the history of the company
|
||
|
|
that has been some changes, let's put it that way. And in recent years, in fact, a large contribution
|
||
|
|
by our friends from Microsoft. Mmm, Microsoft, they put what? It's significant, it's
|
||
|
|
kind of money into the whole thing, right? I don't know, something like a billion or something.
|
||
|
|
Hang on for them, that's pocket change. Just look at the market cap right now. They clock in
|
||
|
|
at nearly two trillion dollars. So, building here are there to stand up too much as those.
|
||
|
|
Mm-hmm. Um, however, I think they have, they have an API on Azure, I think.
|
||
|
|
But GPT? Okay, interesting. But you obviously have to pay for it, but with Microsoft,
|
||
|
|
you always do, right? That's the point. One way or the other. So, yeah, that's a little bit
|
||
|
|
about this, the open AI. They have done more than just GPT-3. Yeah, GPT-2 and GPT-1 comes to mind.
|
||
|
|
Almost, no, no, no. They also, they also did a neural network from music and also.
|
||
|
|
It was meant to be a general research organization for AI. Most well and most known for it's GPT.
|
||
|
|
Links to the GitHub repos will of course be in the show notes and the interesting thing
|
||
|
|
is that GPT-3 is not open source. GPT-2 is but essentially we're looking at some secret
|
||
|
|
source on top of something called TensorFlow. For the few people who have been missing out
|
||
|
|
on the previous part of the main series, the second part tackled TensorFlow and some
|
||
|
|
of the core part of it. So feel free to go back and listen and revisit the second part
|
||
|
|
of the series so that you know what TensorFlow is all about. If you take a look at the GitHub
|
||
|
|
repo, we are looking at a very thin layer of Python on top of TensorFlow. This is essentially
|
||
|
|
GPT-2. At the domain, if I'm not completely mistaken, it's actually language as such.
|
||
|
|
Yeah, so before GPT-0, the top of the GPT-1, which was the original paper to understanding
|
||
|
|
about understanding language and every year since they have been coming out with new versions
|
||
|
|
so 2019 was GPT-2, 2020 was GPT-3, which is when it was, as you mentioned, no longer open
|
||
|
|
source. GPT-2 is widely available for your own and playing around with, but every generation
|
||
|
|
of open AI's GT models gets bigger and less, less available. I mean, you can of course
|
||
|
|
use GPT-3, but that requires something called an API token and I understand that the waiting
|
||
|
|
this is quite long for said token because it's a cloud-based service of, if I'm not
|
||
|
|
mistaken, and you have to apply for such token. Otherwise, you won't be able to use said
|
||
|
|
API. Unless you only use micro-assertion. Yeah, so applying for APIs, unless you have some
|
||
|
|
kind of, I think, company behind you, I don't think you didn't get any wide. I certainly
|
||
|
|
didn't make a great case about how they could feature on the Linux in-laws with GPT-3
|
||
|
|
capabilities, but sadly, they didn't provide us with an API key, so we are too bad.
|
||
|
|
The example is that we're discussing today with stuck with GPT-2.
|
||
|
|
That's the fact that there, of course, did not publish any of the model data on a place
|
||
|
|
like GitHub. You can get the code, but the secret charge is not on there.
|
||
|
|
Sorry, you're talking about three or two. Two. Two, there are some models you can use, which
|
||
|
|
are available. Two was trained on the web text, which is about 40 GB of documents from
|
||
|
|
scrape from embedded links and with some filtering on there, excluding Wikipedia, because
|
||
|
|
there's a well-known data sets anyway. It's based on my, that's what it's based on.
|
||
|
|
Now, the secret source, as you call it, is really that it's a completely unsupervised
|
||
|
|
bunch of, here's a whole load of documents and brain, which surprisingly produces a
|
||
|
|
lot of good results for various language applications. Language, obviously, language is a sequence
|
||
|
|
of words, so with a sequence of words, you can think about completing them, you can think
|
||
|
|
about translating them, so there's often some of the purposes of language applications.
|
||
|
|
But, yes, it's specifically the vision of sentences or even paragraphs that is fairly.
|
||
|
|
It seems to be reasonably good to mean that it also produces so many bad results, obviously,
|
||
|
|
but, yes, if you want. But the domain, there are some pre-trained models out there, which you can
|
||
|
|
use, which is handy and then you can continue training them with the domains that you may want to
|
||
|
|
know more about for more specific, or as you get some general text lines. But the specific domain
|
||
|
|
is just text in a vertical mouse, right? So no image recognition, no text to speech,
|
||
|
|
something like this, so it's purely, you give a piece of text, and then it continues to write
|
||
|
|
based on this initial chunk. Well, that's one of the applications is completion. There's also
|
||
|
|
things like filling in the blanks, if you start a paragraph, have a blank in the middle one,
|
||
|
|
and then the paragraph, then it would fill in the blank there, or you could do the translation
|
||
|
|
piece with it. There's our applicant, mainly because of the page. It's, so it's, as a purpose,
|
||
|
|
it comes back to the name being a generative model, able to generate new data, similar to existing
|
||
|
|
data. So that's filling something in in the middle, or in the end, or translating it, it's really
|
||
|
|
give it some data, and answer, depending on the question that you're trying to set it to do,
|
||
|
|
which obviously works really well with this number. So the ideal use case is a buttoning off
|
||
|
|
with the writer's block. Well, not just that, but it's, yeah, there is a good use case,
|
||
|
|
but there's also code completion, even deriving code, running code based on descriptions,
|
||
|
|
things like that, sorry, the more you train it. There are examples out there for all these types of
|
||
|
|
use cases. So it can write code. It can write code. Yeah, there's links in the journals, but yeah, there's
|
||
|
|
okay. Many handy, you know, for example, you have things like latex, formal description, or
|
||
|
|
sequel, or Python, it is. So if you give it its own source code, it can improve things.
|
||
|
|
Well, in theory, yes. Well, I mean, forget about sky net in that case.
|
||
|
|
I mean, I told you this is a bar game here. Nice one.
|
||
|
|
Well, so the thing is that as I mentioned, with every generation of this model, they're
|
||
|
|
putting more training data in it. And also, so I think many parameters, there's three
|
||
|
|
have become them. And so several. Yeah, the largest percentage, it clocks in at one
|
||
|
|
one hundred and seventy five billion data points, if I'm not completely mistaken.
|
||
|
|
Yeah, so it's, you know, all will be coming out next year, or whatever it is to,
|
||
|
|
but it will be another order of magnitude bigger, right? So it's okay now, what three can do.
|
||
|
|
So our investing in this kind of technology. So GPT might already be running this show without
|
||
|
|
us, without us knowing it, given the large enough cloud cloud. Well, I mean, you know, DSDS,
|
||
|
|
we don't need any any writers anymore. DSDS, what's DSDS one?
|
||
|
|
Yes, DS, dark side text board. Ah, sorry, yes, of course.
|
||
|
|
DSDS, sorry, I understood DSDS. Ah, sorry, okay, okay. Yes. So,
|
||
|
|
which is of course a German, a German, what's it what I'm looking for?
|
||
|
|
It's a German TV show called Deutschland such that you're looking for superstar or something like this.
|
||
|
|
Ah, it's like Britain's Got Talent in Germany. It's something like that. It's crap. Yeah, it's
|
||
|
|
trashy. Yes. Well, talking about this, this, all this trash TV as you call, it was all invented
|
||
|
|
in the Netherlands by end of all, but there we go. And most of it, if I'm comparing it to say,
|
||
|
|
yes. Yes. And then it's a pretty good, they did a pretty fine job of flooding the world with it,
|
||
|
|
right? Because the exporters is left right in the center. Yes, yes. At a price
|
||
|
|
invented in Holland, probably after taking very substances. There we go. One, two, two, more.
|
||
|
|
About substance abuse for one of our expressions. Well, country, Holland, is it? Okay.
|
||
|
|
Oh, I don't know. Okay. So do you think there's a tie between strange new TV formats and substance
|
||
|
|
abuse? Interesting one. Of course, we digress. It's not just TV shows, it's as many musical
|
||
|
|
as score has been written after. Yes, a successful one. Let's put it that way. Excellent, excellent.
|
||
|
|
Details may be in the show notes, maybe not. Yes, for a, for the best approach to enhancing your own
|
||
|
|
mental capabilities, conduct your local leader. Of course.
|
||
|
|
Right, we digress indeed. Where were we?
|
||
|
|
GPT three and who's running the show essentially GPT as a 405 for six might be actually already
|
||
|
|
out there in the wild doing things without us knowing it. I mean, if we're looking at a software
|
||
|
|
that can actually enhance itself by improving its code. This is mind-boggling. Well, I mean,
|
||
|
|
we know we can write code, I don't know if it can improve. Well, all it takes is basically
|
||
|
|
a nice pipeline. Well, all it takes basically its own source code and then start improving.
|
||
|
|
And the rest is called evolutionary Q&A. It's quite simple. Other people call it evolutionary
|
||
|
|
programs. It has been done before. Okay, that was good. Yes. Well, and with that,
|
||
|
|
program is all the world. Perhaps you should consider a new profession. Yes.
|
||
|
|
Yes. Try straight and use TV for this. In case at the moment, out of ideas or something.
|
||
|
|
Indeed. But maybe, but just maybe you can use GPT three to come up with new forms as well.
|
||
|
|
Yes, most likely, most likely. Yeah, well, I mean, I think did you know ask for some
|
||
|
|
lot of numbers before as well? Well, I probably did, but I failed because there was no response.
|
||
|
|
But maybe, but no, no, no, no, they didn't. No, they didn't.
|
||
|
|
No, they were the winning numbers as well. I hope you got them.
|
||
|
|
No, no, no, no, no, I don't indulge in this kind of company. Okay, there's one thing left,
|
||
|
|
because this is not just a theoretical episode, but rather we want to put GPT two to a test.
|
||
|
|
So Martin, given the fact that you have looked into this, why don't you shed light on the details?
|
||
|
|
Details, of course, will be in the show notes, but Martin has done some magic in the background,
|
||
|
|
which he's going to explain in a second as a now. Not putting in the spot or something.
|
||
|
|
So G2 is a source available on GitHub. There's people have done, you know, done work on this.
|
||
|
|
The original GitHub repo is called G2 by OpenAI. It has a number of between models in there,
|
||
|
|
of different shapes and sizes. As with all these things, the spot and the model, the quicker you're
|
||
|
|
able to alter, but of lower quality, generally. So the biggest one in there is one and a half
|
||
|
|
billion as small as one is like 120 mega something. So you can easily get this up and running,
|
||
|
|
you don't have to run it on GP either. It could do it if slower, obviously.
|
||
|
|
You can then train it yourself with different texts. For example, I used one of our
|
||
|
|
books, the HIPAS slide to find unit.
|
||
|
|
Yeah, we're using using your massive NVIDIA GPU cluster.
|
||
|
|
No, this is on my humble laptop. I see. The one with a 27 course.
|
||
|
|
No, it has a few more tens of course than that.
|
||
|
|
You don't have to go mad with it. I only run the fine tuning for a day. So this is the
|
||
|
|
biggest, as you call it, the secret source behind this. It's pre-trained on a bunch of texts.
|
||
|
|
Text being language based over. It has derived a lot of information from that.
|
||
|
|
You want to fine tune it to be more like how the HIPAS slide is written by the
|
||
|
|
numbers Adams. Then you fine tune it with that to give it more emphasis on those kind of
|
||
|
|
yeah. As I said, there are some pre-trained models out there. You can get up and running
|
||
|
|
really easily. Then you just play around with your model parameters in terms of how accurate
|
||
|
|
you want it to be in terms of the more accurate, sorry, is the wrong word, but you can change the
|
||
|
|
things like that. How much text it generates or the randomness in terms of the completion,
|
||
|
|
so the more random you make it, the more random text you get. The rest random you make it,
|
||
|
|
the more repetitive your samples become. So you can say, I have a, with our example, I gave it
|
||
|
|
a false sentence to complete some paragraphs, right? So whatever it is you want to do.
|
||
|
|
And then your samples become more or less repetitive if you make it to precise in terms of
|
||
|
|
parameters. You can also control the diversity of the words. So these are all built in things that
|
||
|
|
you can do. And it has sort of a reasonably good results, really, specifically for non-fiction.
|
||
|
|
Okay. If that makes sense. So, you know, probably partly by training on the
|
||
|
|
H-hags. Yeah, it's presently produced some reasonable results in various places that if you
|
||
|
|
didn't know, it could have been missing by a human quite easily. And in fact, there is a,
|
||
|
|
there is someone who's done a quite a nice video of a Q&A session with GPC3 where they've picked out
|
||
|
|
all the best answers to demonstrate these capabilities. So you know, it will be out of 100 samples
|
||
|
|
that picked up, you know, one to put in the video, but it still gives you a nice idea and see how
|
||
|
|
undetermined a series that it can make up stuff and it can lie and all sorts of things.
|
||
|
|
So essentially, we're looking at a software architecture that is finally able to pass the
|
||
|
|
Turing test. Interesting. Yeah, you could say that. Wow. Okay. People, you're, you're
|
||
|
|
only your first. I mean, that's a big thing. The funny thing is it's just pure training on
|
||
|
|
lots of data, right? There's nothing super special about it. I mean, okay, transformers were
|
||
|
|
a slightly new idea. Yeah, it was, it's been around for a few years and it's
|
||
|
|
used approaches. Interesting field of, with many applications as well, so definitely worth.
|
||
|
|
Yeah. Any thoughts on non-English text while we're at it?
|
||
|
|
Well, the problem with non-English text is that your sentence construction is
|
||
|
|
many languages, Dutch, German, French, the words come in orders and languages have that
|
||
|
|
or specifics, right? So you'd have to train it with language-specific models. It's possible
|
||
|
|
that you could fine tune it with a set in a different way. That'd be an interesting test.
|
||
|
|
So almost of the pre-ten models you get out there would be right Arabic and English.
|
||
|
|
Yeah, because, you know, this is the easiest way to get a large data set of text, right?
|
||
|
|
Or use web data, whether it's over a little reddit,
|
||
|
|
or if you could be, yeah, those are the, you know, the large, all these of text available.
|
||
|
|
I mean, at the end of the day, because we're talking about a TensorFlow extension,
|
||
|
|
or a model running on top of TensorFlow, essentially we're talking about better recognition,
|
||
|
|
given the fact that almost all natural languages are context-sensitive,
|
||
|
|
it shouldn't be a big deal. Given enough computing power, never mind data.
|
||
|
|
Yeah, yeah, yeah, but this is the, I think, why the likes of PR along the open source,
|
||
|
|
or only available through an API, I think, because it's been through a massive
|
||
|
|
articles and either standing up at the end result as an API call to. But it also means she
|
||
|
|
not controlling its usage, or its training. No, correct, by cyberdying instead.
|
||
|
|
You're very first people. Okay, details will be in the show notes. The thing is that Martin
|
||
|
|
took a smart paragraph, but he wrote himself apparently, and then that GPT2
|
||
|
|
do the rest in terms of extending or building on this.
|
||
|
|
Yeah, and, well, I mean, there's just one example, but you can run whatever you want to,
|
||
|
|
or you can ask questions, it's, you can get it to do what you want. Yeah, the fun thing is
|
||
|
|
really training it on, on a visa to make it, make it more, give it a different style or
|
||
|
|
different outcome than what Wikipedia in this case, but, you know, and then a novel written by
|
||
|
|
Dolores Adams, right? Okay. So yeah, we'll put both examples in there to give it.
|
||
|
|
Absolutely.
|
||
|
|
Indeed. Hey, and that brings us nicely to the boxes of the week.
|
||
|
|
Oh, what is your box of this week? Box, of course, standing for the pick of the week in terms
|
||
|
|
of the things that have crossed your mind, and that you see worth mentioning on the show.
|
||
|
|
My box of the week actually would be a German TV show, funny enough, called the Xenomitamaos,
|
||
|
|
the show with Maos. It has been around for a long time. Yes, it has. I know this.
|
||
|
|
You too. I do. I don't know. I'm okay. Nice. It's, it's, it's transformed Martin because
|
||
|
|
you fit nicely, you, you, you, you fit the, the target, the profile of the target for you are
|
||
|
|
quite nice to because we're looking at over-aged man, living alone with a mother. Oh, maybe not.
|
||
|
|
But then it's called living in the night and they won't get you anywhere.
|
||
|
|
This is the joke. Jokes aside. It's been around for the last 50 plus years. It's one of the most
|
||
|
|
popular TV shows for kids on the planet. It's a mixture of essentially storytelling and little
|
||
|
|
ditty stories, whatever you want to call it, explaining how things are done, how they're working
|
||
|
|
on the rest of it. Okay. So, for example, if you want to check out how an Airbus 320 is built,
|
||
|
|
there was a whole series as an, as a series of episodes on this, about three years back,
|
||
|
|
whether it followed the progress of building such an airplane for about half a year.
|
||
|
|
And each three episodes that show would include a segment on building the plane. So, you could follow,
|
||
|
|
you could track the progress of how actually a 320 is built. Okay.
|
||
|
|
That's, that's, yeah. I'm in, I'm in five year old, I'm sure that's
|
||
|
|
a very interesting topic. Indeed. They, they do explain how computers work, how, how the internet
|
||
|
|
works on the rest of it, in a, in a fashion that the H bracket, the, the show is destined that can
|
||
|
|
understand, because essentially we're looking at kids between the age of four and say 12.
|
||
|
|
Yeah. I mean, I have seen it in my youth growing up and all of that. I can't remember any of it to be
|
||
|
|
perfect. I just think it is. You should name. Yeah. I mean, you should be able to get in the UK
|
||
|
|
because they, it's all, it's, it's exported all over the planet. It's okay. I'm a children
|
||
|
|
bit woman. You can get in Japanese, you can get in Korean, just take a language. It's probably the
|
||
|
|
most successful TV show for kids on the planet, even before seismistry or stuff. Okay.
|
||
|
|
And why did you pick this? Because the lack of that I'm
|
||
|
|
I mean, I've been, I've been watching this for the last almost 50 years. No, I'm much more serious
|
||
|
|
reason is actually for the last about 10 years, they are doing something called the mouse door
|
||
|
|
open that they in German demonstrators are where essentially all of the content, this is unfortunately
|
||
|
|
confined to Germany, doors open for kids that otherwise would be closed. So for example,
|
||
|
|
libraries open the doors for kids. So kids can. Yes. No, no, no, no, especially not in the back,
|
||
|
|
not in the back rooms where actually the front, the front stuff takes place as in how books are
|
||
|
|
labeled, sorted, categories, all the rest of it. And a very popular thing is to, of course,
|
||
|
|
how a fire station works. Because normally you wouldn't be able to get into such a place being a kid.
|
||
|
|
And about 20, 30, 40, five fire patrols actually take place in, in said, most of the day,
|
||
|
|
and explain to kids what the daily job is and how they do it. Needless to say, with an H bracket,
|
||
|
|
this is very, very, this is very, very popular. Oh, yes. And set, look, as in the link's
|
||
|
|
user book that I'm sharing here in Frankfurt, or supporting rather not just sharing,
|
||
|
|
dust or has been doing before this pandemic thing here, has been doing that mouse, that mouse
|
||
|
|
door open a day for the last five years, where we got kids in for a day on the third of October,
|
||
|
|
and simply introduce them to open free and open software. And what's your part?
|
||
|
|
Apart from Brian Johnson. No, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no, no,
|
||
|
|
in my box, the week it goes to my chiropectuary, yeah, let's say, and why is that Martin?
|
||
|
|
Well, because she's very skilled at fixing things that are broken, and I see those kind of skills,
|
||
|
|
you can't replace with a AI model or anything like that. I wonder if I wonder if we would find
|
||
|
|
the link to a practice in the show notes or not. Maybe, pretty busy normally. Okay. So,
|
||
|
|
of Mississippi, if you're listening, one heart of the two of the two hearts of us at least,
|
||
|
|
if not more, goes out to you anyway. And with that, it's time to wrap up the show.
|
||
|
|
Needless to say, full credits have to go to something called Hacker Public Radio,
|
||
|
|
because they continue to host us. We will be on Hacker Public Radio. Yes, for the time. Yes,
|
||
|
|
we'll be back. Yes, we do. I almost forgot about this. A guy called Ken Fanon posted it.
|
||
|
|
He posted who? Question mark. Yes, indeed. So Ken is looking for the list of contributors
|
||
|
|
on a recent Linux in-laws episode on curl contributors. Exactly. If we find the time,
|
||
|
|
it will be in the show notes as in the event version of the show notes. That'd be the
|
||
|
|
as I only comment this week. Yes, if you have comments, you can of course post a comment on the HBA
|
||
|
|
website, or you can send us an email to feedback at the Linux in-laws of you. Also, if you have ideas
|
||
|
|
for the show, like GPT4 and 5 and 6 and 7 and 8, whatever, GPT, if you're listening, if you want
|
||
|
|
if you want to get yourself on the show, just send us an email. Forget about Scott,
|
||
|
|
because you seem to be right more advanced than that. Of course, yes, we continue to stick to HDR
|
||
|
|
for our Steam platform. Full credits go out to the people behind the platform. With that,
|
||
|
|
it's time to conclude the show, I suppose. Well, before we conclude, we have two things that
|
||
|
|
we have been promised in the past, which are our friend Claudio. Yes, Claudio, if you're listening.
|
||
|
|
He threatened to make an appearance on the show, but we have yet to hear back from him.
|
||
|
|
Claudio, please get your agent to get in touch with our marketing people. You'll find the
|
||
|
|
address in the website. If Martin hasn't filed, I mean, as usual. If you don't find that email
|
||
|
|
address, simply send a mail to feedback at Linux in-laws of you that'll get us right away.
|
||
|
|
Marketing department on-off. And the second thing Martin is?
|
||
|
|
Well, I think that was, what's that Luna that was going to come on the show as well.
|
||
|
|
Did he? No, no, we didn't actually. Luna, if you're listening, if you want to make
|
||
|
|
an appearance, especially if the topic is open source and is free in open source software,
|
||
|
|
please send an email as you normally do. You know the main address and we're more than happy to
|
||
|
|
slot you in. That'd be in or not. BST? Maybe. And with that, thank you for listening.
|
||
|
|
And see you soon. Bye bye. This is the Linux in-laws. You come for the knowledge.
|
||
|
|
But stay for the madness. Thank you for listening.
|
||
|
|
This episode of Linux in-laws is sponsored by GPT, a framework for modern deep learning
|
||
|
|
applications. Having passed the touring test, meaning that my reactions cannot be distinguished
|
||
|
|
from many human reaction in a given context, I am now working on passing the Trump test,
|
||
|
|
named after an infamous, recent US president. The Trump test essentially means that an artificial
|
||
|
|
intelligence can lower itself in terms of stupidity so that its actions and reactions cannot
|
||
|
|
be distinguished from actions from said past president, which presents a bit of a challenge if
|
||
|
|
you were created with a basic level of intelligence. This podcast is license and the latest version
|
||
|
|
of the creative commons license, tap attribution chair like. Credits for the entry music go to
|
||
|
|
bluesy roosters, for the song Salute Margot, to twin flames, for their peace call the flow,
|
||
|
|
used for the segment intros, and finally to the lesser ground for the songs we just
|
||
|
|
is used by the dark side. You find these and other duties license under creative commons at
|
||
|
|
Chimando, the website dedicated to liberate the music industry from choking corporate legislation
|
||
|
|
and other crap concepts.
|
||
|
|
You've been listening to HECCA Public Radio at HECCA Public Radio.org. We are a community podcast
|
||
|
|
network that releases shows every weekday Monday through Friday. Today's show, like all our shows,
|
||
|
|
was contributed by an HBR listener like yourself. If you ever thought of recording a podcast
|
||
|
|
and click on our contributing to find out how easy it is, HECCA Public Radio was founded by the
|
||
|
|
digital dog pound and the infonomican computer club, and it's part of the binary revolution at binrev.com.
|
||
|
|
If you have comments on today's show, please email the host directly, leave a comment on the website
|
||
|
|
or record a follow-up episode yourself. Unless otherwise status, today's show is released on
|
||
|
|
creative commons, attribution, share a light, 3.0 license.
|