hpr-knowledge-base/hpr_transcripts/hpr3339.txt

Episode: 3339
Title: HPR3339: Linux Inlaws S01E30: Politicians and artificial intelligence part 2
Source: https://hub.hackerpublicradio.org/ccdn.php?filename=/eps/hpr3339/hpr3339.mp3
Transcribed: 2025-10-24 21:07:21

---

This is Hacker Public Radio Episode 3339 for Thursday, the 20th of May 2021.
Today's show is entitled, Linux in laws S01 E30,
politicians in artificial intelligence part 2 and is part of the series Linux in laws
it is the 30th show of monochromic and is about 58 minutes long and carries an explicit flag.
The summary is, part 2 of the miniseries, on deep learning, politicians and other approaches
to intelligence or not. This episode of HPR is brought to you by AnanasThost.com.
Get 15% discount on all shared hosting with the offer code HPR15. That's HPR15.
Better web hosting that's honest and fair at AnanasThost.com.
This is Linux in laws, a podcast on topics around free and open source software
and is hosted a contraband, communism, the revolution in general and whatever fences 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 on less on speed and QT rexes or other
associated dinosaurs. At the beginning of the episode you will notice a reference to quantum
computing which may seem to be out of context for some of our listeners. For a resolution of this
mystery check out the outtakes at the end of the episode. This is Linux Quantum Physics in laws.
Season 1, episode 30. Tens of flow and such.
Ah, good evening, our things. Hey, Chris, things are great and wonderful.
Perfect. Always, how are you?
Can't complain, but know that the muleys are rolling in.
Is it? No, right. The cartels coughing up finally, I think.
No, but apparently the company that I'm working for is 5-1 IPO.
I think. I thought they just had some funding.
Well, no point in stopping now, right?
How does that work? How are you getting people to fund you and then
if he has the plan, I think, yes.
I'm not sure there's much logic in that, anyway. Fair enough, fair enough.
It's when features logic into IPO. Sorry, if you're listening offer and friends,
feel free to send me some more shares. Yeah. Oh, my.
Ah, ah, ah, ah, ah, ah, ah, ah.
Oh, for the off chance.
You're listening. Yes, set me, Shamsha.
Set me, Shamsha would listen to us too.
Yes, and where's the open source support for Redis in Redis Labs?
Okay, but Redis is not the subject of tonight's episode,
but rather something called neural networks as perhaps at least there's
a free hall specifically. So again, frameworks specifically.
Yes, as avid listeners will recall, there was a part one
where we basically tackled the foundation, and this is part two now,
of a 27 and a half part series on machine learning deep stuff.
Our friend Ken did ask for more contributions in this.
Yes, you also send feedback when we want to tackle this in a minute or two.
Oh, yes, no jokes aside, this second part is actually
our building on top of the first part.
Now we got to actually tackle some way more concrete stuff like actually frameworks
that implement neural networks. Indeed.
Like some Martin, why don't you give us a quick recap of episode one?
Episode one, oh, as ages ago, isn't it?
For those, for those few listeners who missed out on it.
Okay, episode one was only the basics of how neural networks work, right? So, which,
yeah, not everybody wants to dive into because that is a lot of work and why not reuse what other
people have built already in magnificent frameworks like PyTorch and TensorFlow.
Indeed. So, before we go into the nitty gritty details, maybe we should do a little bit more
of a detailed recap of what was epic, what episode one was all about. Essentially,
we spoke about the long history track. Yes, but for those people who do not want to spend another
three hours on the director's cut. Well, they have to really. Sorry, that's the prerequisite
for this episode. Yeah, but that's not very user friendly Martin, is it?
I mean, this is, this is what I told you during the marketing brief about a week ago,
you have to pay attention to the needs of something called the listeners.
I think you watched too many popular series, I mean, what I have to recap the pieces.
You were teeing it off the next one afterwards. Martin, quite a few people from America
are listening. Donald, if you're listening to this, we're still fans.
So you have to take, you have to take our listenerships, our listenership into account,
so very much. Yes, some people will have not the attention span that we do.
Some people will have even longer attention span. Grumpy old quarters, if you're listening,
you're looking to what category you belong.
We must do another episode with them. That was actually very, very good. I don't know,
I missed anyway. Doesn't matter. Let's do a very quick extended recap of the first
morning. Yeah, neural networks essentially are your networks in terms of your neurons,
which essentially can be represented by a mathematical function.
They take input values, they produce, they produce output values, and the magic actually happens
inside that neuron in terms of how you model the function. Hang on, hang on, there's no such magic,
and there's no such thing as magic. I've done a PhD.
Are you a computer? It's been a drill for a genius, perhaps.
Okay, Mark. Why don't you explain the magic then?
No, I just said there is no such thing as magic in computer science.
Well, there is, but subject for another episode. Okay, no jokes aside. The idea is essentially
to model the human brain. And as people tend to simplify things because how a
human brain works is much more complicated than simple artificial neural network is capable
to model. They simplify things. So essentially, the white works is that in neural,
that is simple, the actions of a neuron are represented by a model.
So it's a simple mathematical formula. It's some multiplications on the rest of it
that take input values, produce output values, and the idea behind such neural networks is that
you have a multitude of neurons being interconnected. And that multitude then is multiplied by
so-called, or categorized, rather, in so-called layers. The neural network functions as follows,
the idea is that you have a training phase where you essentially teach the neural network how
it's supposed to behave like given input values, the expected output values, and based on the output
of the multipliers of neurons, you modify the individual parameters that are attached to these
functions. Yeah, you have those connections really. Exactly. And you may also modify the connections
if you want to be really close to something called the human brain, because what the neuron brain
actually does, it modifies the interconnectivity between neurons all the time. And this is something
that we call human learning. Yes, to over simplify things just once again, after this training phase,
you should have a model in terms of you have neurons, and you have values, or parameters
associated with these neurons, that you can then use to essentially put the thing into production
in terms of, you give it some real input values, and then the neural network does its magic in terms
of recognizing speech, identifying patterns in an image, in all the rest of it, because at the end
of the day, a neural network is just about pattern matching, on the very sophisticated level,
grad is, but that's where the magic stops. Yeah, so then this phase is normally referred to as inference,
isn't it? And now the second part of this 27.5 apartment series is actually, now how to put this
into into into a bigger context in terms of, now that you have the theoretical foundations,
how you can actually put this into production in terms of how you can apply this to the reward,
because this is say, if you want to program a neural network, you don't have, you don't want to
sit down and just program the neurons themselves, program the functions running inside the neurons,
and feed them the input values, and then do the inference, face yourself in all the rest of it,
no, you want to use some model that is already out there working, and that's exactly what these
frameworks that mark measurements in PyTorch and in terms of flow are all about. And the hint
really is in the name of TensorFlow, because also PyTorch does it the same way, the idea behind
the interconnected neurons is passing around values that are then used to carry out the functions
of this neural network, and funny enough, these values are easily captured, efficiently implemented,
and easily modeled as tensors. Maybe you would like to explain to the uninitiated what a tensor is.
It's quite straightforward, actually, it's an algebraic mapping of end-to-me engines.
There you have it. Okay, panel. Of course, that's oversimplifying things.
Well, no, it's a multi-dimensional array, really, put it that way.
This is one of the expressions of a tensor, yes. In its simplest form, actually,
a tensor flow, sorry, a tensor, sorry, a tensor represents essentially a mapping from one algebra
extracted to another. Think of it like a vector. A vector being mapped onto a different vector,
and the mapping is actually the tensor itself. Now, this simple mapping is also called degenerate
tensor, because it simply takes one vector and maps it onto the other. The interesting bit,
basically, is if you have whole fields of algebraic structures, like matrices and so forth,
and this is the reason why tensors are actually quite good at capturing these inter,
these relationships, these are the connectivity issues. So, if you look at popular frameworks,
like tensor flow, and actually the end is in the name, and PyTorch, the basic algebraic abstraction
would be a tensor. In terms of, this is how you model your neural network on the very lowest
layer. Okay, yeah, thank you for that explanation. Where were we with this anyway?
We wanted to explain how tensors from PyTorch essentially do it.
Well, why don't we cover a little bit about why there is more than one, and what happened before?
Absolutely, Martin. Well, in PyTorch, I don't know where you are in your context,
I just want to find a thing, carry on. No, no, no, I just laid the foundation, Martin. Give it,
give it another shot about the non-mathematical background to this.
I mean, this is it, I don't want to see a thunder, so just go right ahead.
I am, it's lost. Martin, why don't you bootstrap?
Yes. Why don't you bootstrap the tensor flow that is you, and then you can start all over again.
I don't think people have a reboot function, you know?
Implementation, enough implementation flow, okay?
I think five, five, six, seven years ago, people started building various frameworks,
like I mean, I think Carras is one of your favourites now. I don't know how long that's been around,
but and in the last five years, TensorFlow came out first to try and get a bit more of a
standard to this. And then about a year later, you will like, do you know the background of PyTorch,
by the way? Probably not as well as you do, so give it a shot.
Well, I mean, it's the original, the pre, pre, or what PyTorch is based on is Torch, right?
Which was... That's a lottery, yes. Which was written in Lua, which you may be familiar with.
If you're familiar with Lua, I'm afraid, yes.
What would you say the attributes of Lua are?
Open WRT uses it as a basis for a web framework called Lucy, which is essentially the web-based
GUI for a router management. And there's also this crazy, it's a surgeon, it's a
laboratory. If you're listening to all friends who had this idea of providing a script-based interpreter
on top of a very popular nose-equal database, what right is? Indeed.
Anyway, I mean, Lua is not really known for its wide adoption, I would say, above all.
Hang on, Martin Martin, does Procedede ring a bell? How much? Procedede.
Yes, it does. And this is actually written in Lua, believe it or not.
Yeah, but if you compare it to other program languages out there, then it's a little bit low-down,
the... Well, Procedede would be one of the... The most popular XMPP
service on the planet. Yes, but that's not a programming language, is it?
Just an XMPP server, that's correct. It just happens to be written in Lua, that's all.
Good, good, glad we can agree on that. Anyway, so yeah, Lua is... I don't know, I mean,
obviously I saw Lua for the first time, well, not obvious, but first time I came because Lua was
busy with it, it's done about you, but so for many people, Lua is pretty unknown, right? And
is not that well-known for its modularity or widespread adoption amongst...
The preserving fraternity. Unless you take a peek behind the scenes, because Lua,
like Redis is basically present in pretty many areas.
Yes, I came across it first, basically, when I installed my own, my first OpenWRT instance on a router.
Yes, but you didn't program in Lua, did you? Not that no, but later. Yes, that that would
been in your Redis days, I imagine. Just like before. Okay, no, fair enough. Anyway, so yeah,
some people, or many people, are not that familiar for Lua and they, and so the founders of PyTorch,
they're like, oh, hang on, everybody's using Python, why don't we wrap that around Torch and create
PyTorch? All right, Torch was born, yes. So, yeah, why don't you tell us where TensorFlow came from?
Google. Uh-huh. No, the idea was essentially to provide a low-level infrastructure
for back propagation networks, because as probably the majority of our listeners would imagine,
Google was really the company that kick-started at artificial intelligence all over again,
even way before companies like Facebook and so on, coped on. Okay, this is what we didn't
just, this is what we didn't cover in the first episode. Well, not in great detail anyway.
Artificial intelligence goes back easily as a computer-based artificial intelligence,
goes back in, goes back easily to the 50s and 60s. Float artificial intelligence goes back
even way further. Who's flawed? Like, yeah, like the, like the attempt to breed certain
politicians who had strange ideas and the rest is pretty much history. But we won't go into that
because this is not a history podcast. It's not a political one, well, I don't know, it's the
least one. There was a certain amount of communism support here. Anyway, carry on. Absolutely,
anyway. So in the 50s and 60s, not especially North American scientists said the idea of,
well, we have computers now. So modeling a human brain can't be that difficult. So speech
recognition and image recognition and such things are just from the corner in the 50s.
So they develop a beautiful program along which is like list and other functional approaches
with the idea actually to build the foundation for something called artificial intelligence.
The trouble is that the existing hardware at the time didn't quite live up to the expectations.
So this is the reason if it takes a close look at the history of artificial intelligence,
the whole thing entered a hiatus or a hibernation period or whatever you want to call it
in the early to mid 70s. And of course, there were kind of wake up calls in between.
90s tried to revert the whole thing in terms of there was a company called Run called Nuance
that tried to use artificial intelligence for speech recognition. But these were kind of
isolated blips because even then the hardware wasn't really up to scratch to cope with the
with the computational demands of artificial intelligence. Things really changed
when a little unknown search company called Google and at the stage in the mid 2000s.
Because as probably we all know, Google is not necessarily about hardware but rather the
intelligence running on top of this hardware. As in doing software that is able to massively
parallelize algorithms running on cheap hardware. That's exactly how they did their first
search engine implementation. So you'll see this actually. If you take a look at their base
infrastructure technology like big table, like big file system, it's all published. You'll find
to show the links on the show notes. The idea behind this is actually that they come up with
software that took into account that a hardware is not perfect, hardware can fail, but the
real intelligence is actually in software. So I'll say that, but Google did develop the TPU,
which is that came later. Yeah, but the notion of cheap commodity hardware.
But these are commercial implications, which we're going to touch in a minute or two.
Anyway, to kind of long story short, the idea behind the initial Google version
0.9 if you will searches and Brian's invention was actually to do it with intelligent software.
You'll see this in the initial implementations. They had the foresight to imagine file systems that
are able to cope with petabyte of data. Because the file system at the time didn't measure up
to these expectations. So they simply sat down, not the two of them alone, but rather
these software engineers and device software that is able to power such large scale systems
and do it in a very reliable way. And that's exactly how they arrive at the technology that they
have in production right now. And along the way, they discover that there's this ancient
technology called artificial intelligence that has been on a sleeping for the last 30 years.
Because simply the hardware wasn't ready. So in addition to something called TPUs,
like terms of processing units, we want to tackle this in a minute. They come up with the idea,
okay, we produce software that is able to execute on standard hardware, but masterfully parallel.
The idea behind the initial Google was, and I'm just using Google as an example.
You buy cheap kit, but you expect this kit to fail. So you do a software that is able to cope with
failure. This is common law. Google hasn't confirmed this, but they were able to cram more
motherboards into a rack than the then, and I'm talking about early 2000s now, then the then
industry stands are kind of specified in terms of thermal distribution TDP.
So because they expected some of these CPUs to fail, if that was the case, the software would
notice that, and then some other parts would take over. Same thing for software, because
sorry, same thing for artificial intelligence, rather, because the real intelligence in artificial
is actually in the software. So this is the reason why they took a close heart look at the
foundational research that had been done in the 50s and 60s, for example, with regards to
backpapagration networks, which weren't new about 20 years ago, but rather go back to the 50s and
60s because that's exactly the part of time when people came up with this.
Indeed.
So because GPUs and CPUs only go so far, some of the Google engineers had this broad idea of,
okay, similar to GPUs, actually, we can do specialist hardware, especially if you want to
deploy this in the cloud context, that is just able to tackle the tens of flow algorithms that
are implemented by tens of flow, as in the library that implements backpapagration networks.
Hence, the idea of a GPU of a tens of flow processing unit was born that simply takes
or that it simply implements the algorithms that are implemented in the library on the
framework itself and executes that on hardware. That's the overall idea, achieving a massive
speed up in addition, and this is where the speed spot is from a commercial perspective,
in addition to these GPUs being able to deploy in cloud environments, because this is where
the money is. You can go into a hyperscaler, in the Google in that case, and simply deploy a
farm of GPUs that then execute your tens of flow algorithms. Indeed. Other cloud vendors are
available, but without GPUs. Indeed. Well, Microsoft has something similar, right?
Well, most of the people stick with GPUs, right, because there are more
easily available, well, having said that, no, not so easily available these days, but
in fact, they are quite slow. Gamers, if you're listening, stop buying them.
Well, I think the gamers would like to buy them, but they can't.
Yeah, probably people from your cartel are buying them instead.
I can't really comment on this, actually.
Okay. Good. You see crypto mining on GPUs has its limitations.
Indeed. Indeed. So yeah, that was a good point. I mean, the TPU is really,
an ASIC in short, isn't it? But just like where specific GPU mining is also programmed in hardware
through ASIC miners are being and should be more efficient than GPU miners.
But going back to these frameworks, is there anything else in the question?
PyTorch and TensorFlow these days, because are there any other frameworks that come to mind,
a platform from PyTorch and TensorFlow, because these would be the two prominent ones now?
Yeah, well, before both TensorFlow and PyTorch became so popular, there was a whole variety
of the right. I mean, Cara's being one of them, and Kathy, and MXNet, and God knows how many
there were. But they're all kind of very specialized, and the good thing about the adoption of
TensorFlow and PyTorch is that, you know, with any open source projects, you get the benefit of
many people improving these frameworks.
I mean, the beauty is with these two frameworks, they just say they're open source.
And depending on your on your on your hardware specification, you can pull them down,
and you can run them, you can actually, if you want to do this, you can run them on an SOC GPU.
But then it would say, a dedicated GPU, you can run them on a GPU as well, right?
Yes, you can do that too, but don't expect miracles.
No, but the beauty is for the majority of something called the main specific frameworks,
pre-ten models are available. Indeed, what is the pre-ten model?
pre-ten model is where someone has done the training for you.
You mentioned the training phase earlier, and why is that important?
Well, because it takes a hell of a long time or a hell of a long, a large amount of processing
power to do the training phase, because it's a very simple process or simple.
We could we went through it in the in episode one, right? It's it's, yeah,
a sheer horsepower kind of scenario, rather than anything intelligent training model.
Basically, adjusting your your weights and biases until you're going in the right direction,
and reach your your optimum for the model, which is done by doing many, many steps and
you know, doing gradient descent and so on as we discussed previously, but
yeah, so pre-trained models are great because they give you, you know, a lot of pre-processed
training in the box. And going back to the GPU or kind of power of processing general,
any idea why tensors were chosen as the main instructions for these networks?
Well, tensors aren't instructions, they're more the this data.
Sorry, not instruction, but abstraction.
Ah, abstraction, right? Well, because it maps, if you look at your neural network layers,
they are arrays of numbers, right? So, um, which is essentially a tensor.
Indeed, but the beauty with tennis is actually that you can decompose them into simple
arithmetic instructions, which you then can parallelize on different course in terms of GPU
course. Because if you take a look at how a matrix is multiplied with another matrix or a vector,
most of these computations can be done parallel. Only if you are at the reduction phase to use
the simple example of a map-produced algorithm, you actually have to consolidate this using a
single core, but leading up to this, you can parallelize this easily on different course.
And this is the whole idea behind frameworks such as TensorFlow as well as PyTorch to decompose
these sensors into independent algorithmic, sorry, into independent arithmetic rather,
and parallelizable instructions, so that you can use the full computational power,
if you're disposal, to give it a crack. Yeah, so do you know why these two frameworks are not
the most popular ones? Correctness evolution, right? Many people saw the advantages and simply
used them in their projects. Hmm. Lua or not? Well, this is the, yeah, I mean, I think this is
why the rise of PyTorch has come, you know, it was obviously, it came out after TensorFlow,
but it's pretty much equal in popularity these days because the, you know, the language of choice
for most, well, data scientists, whatever you want to call these people, is Python, right? So
it is a very natural fit, therefore, I mean, you can obviously also use the Python interface,
TensorFlow. Yes. Indeed. Full disclosure, of course, there's a whole episode on Python.
Yeah, in last year's back, I'm back at the log of something called Linus and Lars.
Okay, so you're wondering about the sips or problem with language called Python?
Yes, indeed. For those of you who are not using it.
Yeah, so other question I have for you is, why is there a tens of, what's the difference between
TensorFlow 1 and 2? Hmm, check out for Tracer, I'm tempted to say.
No, Mark, now wouldn't come as an expert on the end of flow. So if you have the answer,
go for it. Sure. Sure. Well, there is, there was, it's quite a significant difference between
TensorFlow 1 and 2, and it's really that TensorFlow 1 was imperative programming. So,
sorry, symbolic programming. So the, whereas Python has always been from the start,
build that way. So they're both converging to the same thing, but there's a big difference.
So if you ever starting with trying to build your own applications with TensorFlow,
and you come across TensorFlow 1 and 2 is quite a significant difference in the way the two work.
So, so it's worth noting that, you know, one is pretty ancient, but there's a lot of
examples, implementations out there on one. Why, why don't you explain the difference between
imperative and symbolic programming? Trish, imperative program would be like basic,
a cova, you tell the computer exactly what to do and how to do it. And symbolic program would
be much more like functional programming, i.e. a program, a function program language would
understand this, atoms and the relations between them. Yeah, what about the execution element of
these two different ways of programming? I think CPUs play a major role in this.
Well, it's, it's commonly familiar with the term eager execution. I'm not.
This is the reason why we have highly-paid experts on the show that mark this show.
That sounds like it.
Give me a crack mark.
I thought you were quite into your programming power lines, but there we go.
Okay, fair enough, fair enough. I thought you would probably be happy explaining this one,
but I can do it for you. I think I've done enough talking.
Is that so far? Give it a crack.
Okay, so I mean, you know,
that I can't correct you if you're wrong. I see.
Okay, so if we think about any kind of program, right, we have lines of code and
in an imperative programming language, they are executed in sequence as and then,
whereas in an symbolic programming language, we have a compilation phase where the most
optimum execution of that representation of the functionality is built, right? Makes sense?
It does so far. So that was the, but it's quite a, so TensorFlow started with the symbolic
approach and saying, because obviously there's advantages to this, right? Like,
yes? No? I'm listening, Martin. I'm just wondering if you wanted to fill in the blanks.
No, no, that's okay.
I've done enough filling the blanks for one day anyway, so I know where it's.
Oh, okay. I just wanted to make a discussion, but that's okay.
Or, or a two-way explanation, but it's fine. I can do it, I can do it.
So I'm having some Martin, I just listening, or that's okay, always.
That makes a change.
It does indeed.
Give the crack, Mr. Mr. Mr.
Give the crack.
Right, where were we? So yeah, okay.
So yeah, yeah, so, right, so if you think about a, I don't know, you, some people consider a,
the execution of a program to be a graph, right?
And so you basically built with the symbolic programming language, the graph is built
compilation time and is fixed, right? So whereas in an imperative program, I mean,
it's done, every step is done at the time as described by the programming language.
So the advantage of this model is that it's very efficient, right?
Your symbolic implementation, because the compiler is trying to find your most efficient way,
and you can reuse your memory space and all sorts of excellent stuff that you benefit
you get from that. Which is all well and good. If you have a, if it's bringing it back to,
neural networks, or if we think about models, then if our model is fixed, then that's great,
because you can run it as many times you want to, it's optimized to the N's degree and so on,
and you can paralyze it the however many ways you want to. However, if you have a model that's
dynamic or self-adjusting, then that doesn't work with a static kind of execution graph, right?
So that's, so Pythos from the start did that dynamic approach, and with TensorFlow 2,
they have adopted that as well. So yeah, so if you're looking at, or if you're using the methods
that are available in these frameworks, then you're probably not too worried. But if you are going
to develop your own models or even adapt the models that come with some of these frameworks,
then you may want to be aware of this. I mean, it's something that I came across with,
with TensorFlow that, you know, TensorFlow 1 and 2 are a huge need, different, not that I'm an
expert at any means, but I have certainly played with it a little bit, and it's similar to some of
my colleagues have been saying as well. So it's, but yeah, now both are, I have adopted the same
method. So they both do the same thing. It's just TensorFlow started somewhere else. So yeah,
that was the end of that episode. But yeah. No, interesting. Because I thought that
TensorFlow was mostly written in C++. I don't know. I don't go into the low level.
So it's interesting because C++ is a compiled language. Yes. Yeah. Yeah, but if it's more the,
okay. Well, maybe we should do an episode of the different levels of the different steps between
implementing TensorFlow and Python. Actually, it doesn't matter anymore, because both
are, yeah, I have adopted the same, but your execution now. So from that point. Yeah.
Because that this is why they changed TensorFlow, right? Because they couldn't do dynamic models,
because everything, once it was compiled, it was fixed. And so you couldn't adapt itself to,
or you couldn't be built to be adapted to ball. So TensorFlow would be the first AI framework that does
what's what I'm looking for. Just in time compilation of self-modifying code.
You heard it here first people.
If you're listening, the email address is sponsor at linuxilost.eu.
Well, you say this, but you know about, we have a future episode on
Google, certainly we do. I thought the current episode was on Google.
No, but you're wrong.
I think that's an idea. It's an appropriation that works.
I always say that. Okay. I thought that was the episode one. Anyway,
what we say. Yeah. So the one of the upcoming episodes, we're going to talk about a
quite popular, well, popular, an implementation of a model, which is
raised some waves in the community and press, because of its
abilities in terms of language. No, I'm listening, or I'm Martin as usual.
But I mean, if you look at adoption, okay, Google is the biggest user of TensorFlow, right?
But when they come up with stuff, well, you would hope they were the biggest users.
But if you look at the adoption of PyTorch, are you familiar with company called Tesla?
Yeah, they make cars, whatever things.
Cars and batteries and all sorts of stuff, and lock it.
But I think they just borrowed the name from some really famous physicist.
Yeah, it's long gotten, of course, but that's a different story.
Well, yeah, family Tesla, if you're listening, you've got a really good call case.
Tesla, don't try. Just don't forget about it.
Don't you reach out to Martin for for for Martin consulting? That would be way
money down the drain. To use the technical term now.
Let's not discuss marketing. That's not going to go well, is it?
Where were we? Yeah, so a lot more companies outside of Google are not outside of Google,
are contributing, supporting PyTorch, right? And its popularity along the research community is,
well, I think from me, eco, if not greater than 10s from these days.
It's a silly competition between Tesla and PyTorch.
Well, I find it a bit curious that why do we need to, right?
It seems a little bit of a wasted effort. So there must be some differences.
Otherwise, one would become more prevalent than the other.
If it's not possible, if current laws and it's going to go by, there are more frameworks
than just these two dominating ones.
Yes, but that's kind of like, yeah, why bother, right?
If you have these two that are seven years in the making, then why do you need another one?
I mean, the functionality that they offer are obviously in the fields of computer vision,
NLP, etc, etc. So what would you be missing from those two frameworks to start?
Frameworks are pretty much like cars or women, right? Or men for that matter before the PC,
police gets to us. Cars get you from A to B and women make you happy. Or men for that matter,
they don't get me wrong. But at the end of the day, they have four wheels. Sorry, cars now.
The PC police will probably shoot me for this remark, don't worry about it.
And women do have hairs, arms or the rest of it. There are, of course, differences.
Cars, for example, have different engines, men and women have other difference. We won't go into
DD test, but you get my drift. Same goes for Operation Networks. At the very end of the day,
they are pretty similar. On the surface, they may differ a lot, but at the end of the day,
they just are about mostly pattern recognition, and about comes afterwards.
Yeah, sure. I mean, we were going to have someone on the show on these things with me,
but follow this. The reasons that didn't happen, though. Yes. But it's just, I mean, if you think about it,
why do you start contributing to one or the other, which would be an interesting question for me?
You have a problem solved. Find a tense flow, hasn't solved it or doesn't provide options to solve it.
It's better doesn't have options to solve it. So why do you choose one or the other to start contributing to it?
Because you like lure?
Maybe, maybe. Where does lure rank on the popularity of programming languages?
Um, definitely, 76 points, something on the CHOBs index or whatever. CHOB, of course,
you'll find the links in the show notes out there. Of course, the Danish company,
I think it's called the importance of being earnest or something.
That's the stack of those surveys, no? No, no, it's here with something different.
No, no, that's not what I was thinking of, but is it not online?
No, no, the stack of actually ranks lure at position number 17.
17? 75 points, yes. Oh, yeah. Of course, the routers and OW,
and open WRC helped a lot. Yes, yes.
Okay. Right, sorry, to, to a diverge. Um, yes, what else would you like to cover?
Oh, that's pretty much it, all right. Maybe teaser for the next part of the
Oh, the second part, mini series, episode mini series, whatever. Is this, is this,
is this specifically for Bob Boris? Is this, is this going following on from the current modeling?
No, I mean, just, just what we do want to cover during the next, for the next episode.
But now that we have, now that we have a very good grasp on the basics, anyway.
Can we do another recap? What's done? Yeah, podcast episodes cannot only go so far, people.
You, I mean, this is say, we only scratch the service here. And it is to say, we did not, we did
not go into the mathematics of this because this is not what Linux allows us all about because that
will get very quickly, very complicated. We leave. Yes, blatant cross, cross promotion,
we leave the mathematical details to a podcast called the grumpy old quarters. We have them on
the show. Yes, we had them on the show quite a few episodes back. There are legacy people,
mostly concert with windows with windows, sorry, with windows. What's the legacy people?
People that are old, the name is actually in the name.
Correct. Please, if you're listening, you, you won't come back on the show at any time because we
just, I thought it was just a hilarious episode and we really like you. But anyway, it doesn't matter.
Okay, jokes aside. Yeah, the next episode will be about the practical applications of these
kind of two foundational frameworks like pytorch and TensorFlow. We will go into, we will cover the
domain specific implicate, the domain specific implementations like Harrison, so forth based
on, based on pytorch. And as I said, this is say it's this episode is not about the
runner details like mathematics and so forth because that would fill another two or three
episodes of this mini series because it's just very complex. And that's the reason why I'm
I just need to go with that blackboard algorithm. Yes. And that's the reason why actually most people
would just use these frameworks rather than kind of re-implementing these frameworks themselves.
Deep. Because all the other fancy math stuff is abstractly away from you.
Mm-hmm. That's the useful line. Compuces.
And levels of programming languages. Yes. And before we close out this episode, of course.
Ah, yes. We do have. Yes. We do have to do feedback. Yes. Yes.
Yes. Yeah. As certain Mr. Kan Fallon of hate our fame. Yes. Indeed.
Mm-hmm. Right. And to say I always thought that artificial intelligence is misleading.
Ken, you're absolutely right. Artificial programming would better describe what's going on.
Yes to a certain extent. Artificial intelligence. Yes. Of course. Can brutal teaser Ken will be
be on the show very shortly.
Ooh, couple of weeks time, yes.
And of course, Ken is one of the beautiful people behind.
Perhaps I'm going to go like a barbecue radio,
that we still use as our main hosting platform
for all the episodes, okay?
It's out there.
So yes, at the fish intelligence
as a term is somewhat misleading, I grant you that
because this is where it comes down
to people's definitions or interpretation of those words, right?
Perhaps.
Well, at the end of the day, the way we tackle
artificial intelligence these days
is more or less like programming
in terms of algorithms being implemented by a computer.
Yeah, but that, okay, so, but yes, okay,
the two are fairly synonymous at the moment,
but it doesn't mean our position in terms of
against can't be implemented in a different way, right?
Exactly.
If you would decouple artificial intelligence
from computers, fine, indeed.
But we're not.
No, because we are that kind of podcasting.
No, we are that kind of people in terms of people.
That's true, that's true.
In terms of people simply using computers
to do artificial intelligence programming.
And so, can't is actually spot on.
Yes, yeah.
Mm-hmm.
Yeah, I'm sure someone will come up
with a biological artificial intelligence solution
that's important.
No, actually, I actually put in some other non-front,
non-mind-heartware to begin or something
beyond computers, yes.
Like molecular biology or something,
like not necessarily brains, but the thing that comes out to it.
It's another topic for another episode.
Here's a question for you.
Is intelligence limited to people?
No, certainly, it's not.
Well, there you go, then.
You have it into very limited extent,
also in politicians, for example.
Ah, yes, of course.
Mm-hmm.
The day is very interesting, isn't it?
But it's a decouple.
Sorry, and other beans, I didn't go by.
Ha-ha-ha.
Well, like, people in marketing or no.
I wouldn't go that far over.
It's a different story.
And with that, we have not only reached the end of the tether.
In these?
The end of the show, I suppose.
Yeah, I think so.
So Martin, see you on the other side, for another episode
on artificial whatever.
Ha-ha-ha-ha.
And of course, for the hipsters out there,
that would be artificial space, full-stop star.
Sorry, plus being a complete regular expression now.
Excellent.
Martin, and with that, we're going to call it a day,
and see you soon.
This is the Linux In-Law.
You come for the knowledge.
But stay for the madness.
Thank you for listening.
A little voice making you forget that SkyNet was once
this evil empire trying to change the world.
If you can change, so can we.
This podcast is license under the latest version
of the Creative Commons license, type
attribution share like.
Credits for the entry music go to bluesy roosters
for the song Salut Margot, to twin flames
for their peace call The Flow, used for the second intros,
and finally, to the lesser ground for their songs we just
is, used by the dark side.
You find these and other duties license
under Creative Commons at Tremendo.
The website dedicated to liberate the music industry
from choking corporate legislation and other crap concepts.
You are currently the only person in this conference.
The only person in this conference, but it's this.
Oh, I'm going to end the session and start a new one.
Hello.
Hello.
Hello.
Hello.
OK.
OK.
OK.
Yes, that apparently works.
Uh-huh.
Do you know what a quantum torus recourse of neural network is?
Well, that what a torus is.
I was trying to work out what a quantum torus would be.
Hmm.
You deviate from standard quantum architectures and you throw
in qubits and then you have pretty much have it.
Like the neural networks basically are able to tackle and space,
peace based, and peace based.
Instantaneously, more or less.
Yeah, yeah.
Well, you've got to start somewhere.
This is about the programs for when this stuff actually is working
sometime in YouTube or not.
That's this delorean car.
Yes.
Well, that wasn't the success, was it, really?
Apart from the music.
It was probably before its times.
Well, no, it was made in Ireland by dodgy crook.
This was the issue with it.
If you're listening, whatever your name was, delorean guy.
You are crook.
Well, I think it produced at least six of them or something.
Well, it was all funded by the government or by the EU, even possibly.
Yeah.
That was Northern Ireland.
Because it was built in, yeah, you need.
Yeah, I mean you.
Okay.
I thank you so much time and see you, don't you?
I was like, other people.
Yeah.
There was a question from a certain Bob.
He was asking, when you refer into models, is that similar to the pictures
you're sending around usually or do you refer to?
No, no, no.
Quite different models actually.
Okay.
Well, there you go, Bob.
Thanks for answering.
Bob, if you listen, there are quite a few different categories of models.
And the model you are probably looking for, unfortunately, is not part of the show.
I thought you were working on that.
No, I'm not.
Sorry.
Sorry.
You've been listening to Hacker Public Radio at Hacker Public Radio.
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 HPR listener like yourself.
If you ever thought of recording a podcast, then click on our contribute link to find out
how easy it really is.
Hacker Public Radio was founded by the Digital Dove Pound and the Infonomicom Computer Club
and is 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 stated, today's show is released on the Creative Commons'
Introduction, ShareLight 3.0 license.