So, first place in the Sprint Deep Fake Detection Challenge.
Outcompeted about 50 other teams.
With lots of creativity, I have a very interesting technological foundation and that is...
Oh, Max, don't do it so exciting. Tell us who won.
No, that's just the end of the episode.
Great.
Well, my colleague Maximilien Brose is about to quit and if you don't want to do it short, I'll do it short.
My name is Karina Schröder and I'm happy that you're here with us at KI Verstehen.
And you basically expect every moment that you, Max, put me back in here.
Because last time, and I remember very well when we talked about Deepfakes,
so with KI created false content, you simply gave me false speech messages.
I haven't forgotten that, Max.
Yes, but I can promise you this episode that I won't make you excited.
If you want to hear again how I was in Perth, you can of course do that.
We link the episode in the show notes.
Today we're going to do something a bit different.
You're telling us about a competition.
The Federal Agency for Jump Innovation is running short for Deepfake Detection.
Exactly. And there I imagine three teams in front of two of the Deepfakes technology.
So really detecting and that was also part of the competition.
A third team, which thinks about the problem differently,
namely real content in the network wants to mark.
To be honest, there is currently almost no more important assignment for KI than to do that.
Because if you move online, you have the feeling that you don't know what's real and what's not real anymore.
Later, since Sora 2 exists, so such a video generator from OpenAI,
the end of September has appeared, I have the feeling that all these short clips that are now being posted on the Internet,
I have no idea anymore, I'm not sure anymore.
Totally, that was a really crazy moment for me, because I'm sitting in front of social media now
and really always think, "Is that real now? Is that KI generated?"
I really didn't have that before.
And yes, now a little newer Google, they just released a new image creation tool in the framework of Gemini.
Or rather, it's a further development of NanoBanana from Google, namely NanoBanana Pro.
I think the name sounds so ultra-funny.
But to be honest, it's not funny at all.
What annoys me the most is that I like to watch memes, such as pictures and videos that are satirical on the Internet about current events,
such as politics, culture and so on.
Somehow they talk, they comment.
And now I always have the feeling that I only see KI-generated content and don't do anything lovey-dovey anymore by people.
Yes, sure, but I also have to say that NanoBanana Pro really does this pretty well, generating these pictures.
I can do that for free, let me generate a few pictures.
And they get problems, the models used to have, for example, with letters.
Now totally great, that the letter also says exactly what it has written here and also looks real.
And the pictures, they just look really fake and have a total depth of detail and are also super easy to generate with a prompt.
But now, in addition to all this fascination, of course, there is also the shadow side.
Because these generated content, of course, can also be used for fake news and propaganda.
And with that, the question arises for our information space, is that what we actually see the reality?
That only says cost, which we have already briefly heard at the beginning and which led the Deep Fake Detection and Prevention Challenge sprint.
And for such fundamental challenges, which of course there are no solutions for these challenges, we want to find solutions.
And in these cases, we start with challenges, i.e. our innovation competition.
I think I'm completely confused by reality television, because I have a kind of a deathmatch battle in my head,
where all these detectors approach each other on one day and, so to speak, always choose the opposite side until there are still two left.
And then the final break is, do I really imagine that?
Yes, not quite. I can tell you a little bit of the challenge.
In fact, they already started in October 24.
There were more than 50 applications from all over Europe.
And there was a jury of 12 teams selected within two weeks, which then also sprinted directly with each of them up to 350,000 euros.
So that they can further develop their technology ideas into a product.
And after half of the time, there was the first between-challenge,
where the teams had to show how well their prototypes already worked and which customers they wanted to achieve with it.
And the seven best teams out there, they each got about 420,000 euros to develop their technology.
Yes, totally. The money, by the way, comes from federal funds.
So the federal fund is also 100% income, which is sprinting.
And in the end, these seven teams had to introduce their products to a jury again.
And in the jury, there were people who used to work at Open AI,
or from the civil society, who deal with deepfakes and information,
or people from public institutions such as the BSI.
So the Federal Office for Information Security.
Yes, that's always a bit difficult to do with your lips, but you should always say that.
I think that when I think about it, I will still be a bit timid.
Because I think the work was going on all the time.
And then you have to think about it in front of the jury.
So the mood of the place was definitely very exciting.
Totally. And that's really a lot of work for all the teams in this year.
Almost 13 months have passed.
Also because you have evaluated how much they have improved since the first major evaluation,
well, six months ago.
And then a little teaser said that there were huge jumps.
And after the first good six months, the teams were more on the stand,
so that they could show that their technology works theoretically.
So practically a first implementation that was not yet practically production-ready
at many corners and ends.
That you could have given good knowledge to the customers.
But what is normal is that we are trying completely new things
and that takes a certain time until the development standard is reached.
That is, even if we can use the things to evaluate on a data set,
the performance was actually quite bad.
And today it looks completely different.
In fact, we actually have accuracy rates in the recognition of real images
in a data set that is mixed from real and synthetic generated images
that exceed 99 percent.
Okay, that sounds crazy impressive.
But I think maybe we should get a little more concrete.
You said you want to introduce three projects.
Then put it off, Max.
Exactly, then maybe we start with Gretchen AI.
This is a start-up, so around ten people that have been trained
from the German Research Center for Artificial Intelligence for the last few years.
Okay, moment. Gretchen AI.
So there it is about the question of Gretchen.
And what is the question of Gretchen now?
So whether the content is real or fake?
Yes, I think that's how you can understand it.
And briefly to both of you, Tim Pultzel and Anna Das.
These are both AI researchers at the German AI Research Center.
And there, the two have been researching deepfakes for almost four years.
In a large research project.
And what is exciting about it is that they work in this research project
with media, such as the German wave or the RBB together,
with images or videos that you see online, discover whether they are real or not.
We at the AI Research Center, of course, we have the AI architecture and the training
and everything that is associated with it, we had technical parameters in mind
and thought we had the best model.
It must have really been great now.
It took a whole week until we understood how to believe these models
and then hand over to journalists, so that trust is actually built on the other side.
Because with deepfake detectors, you have to admit,
no model can really recognize 100% of each deepfake.
Okay, but actually the models also have to be trained with deepfakes
and now we know that new generation of images are coming every day
or improved generation of images are coming out
and new deepfakes are coming again and again.
Exactly, and maybe you haven't trained with them yet.
And I think you've already described that very well.
That's the big problem or the big challenge, the generalization.
So the detectors have to work for all generators,
even if the models don't know the images yet.
Yes, and that's also the biggest challenge for Cretien AI,
we have told them both.
And a lot of deepfakes are coming too.
Someone may have just changed such a detail in the picture,
but with that the entire statement is wrong.
And that's why you get, in most cases, such a percentage value for detectors.
This picture is 85% a fake.
And then, of course, it helps you as a factor checker,
I would say really only necessarily.
Well, actually it has to be waterproof, right?
So if you want to publish it and give people security,
then you have to be sure yourself.
Exactly, and then you don't just look at the technical data of this picture.
And that's why Tim Pulsiel and Anab,
also received feedback from the editors.
What the colleagues of the factor checkers are doing,
they ask questions, they take the context factors,
they say, okay, where does the picture actually go?
Can that make sense at all?
Regardless of whether a deepfake detection AI model
really performs super well or only half way well,
the question is, does it make sense and can I believe it?
And we as AI developers have said,
and that was, I think, the outstanding moment,
wait a minute, why do we actually only do deepfake all the time?
Why don't we go into the process of questioning
and proving, verifying or falsifying a deepfake?
It's supposed to mean their detection technology
that goes through all the steps that would also go into real factor checkers
and do you support their work with it?
Okay, how do I have to imagine that?
Because as you describe it, I have a AI agent in my head
who is working through these steps and working independently.
Yes, with AI agents, you are in the poodle's core of an AI.
That's what I played for the good word.
I'm a great artist.
As a germanist, I think it's great art.
So if you look up a picture on the platform,
then check the model for one thing,
what typical directors do.
There are certain artifacts, changes in the pixel transition and so on,
which are typical for certain deepfake generators.
But what about Hinaus?
Identify the AI agents.
What kind of people are actually in the picture?
What happened in the picture?
Where was it taken up and at what time?
And then our AI agents look up the entire internet
for similar information, say Anabdass,
for example the context or the person or the whole picture itself.
Yes, for example, if it was already in a different context before,
and all this information, almost Cretian AI,
then together so that they are good for the user-readable.
That reminds me a lot of it.
I still remember when we made the first episodes of AI understanding,
we always said to the people,
they should do this Google image search
in order to see if the pictures look different.
Exactly, the agents do that too,
and they are also looking for sources
to see if it can actually be true.
So now an example.
Here comes a photo in a editorial from 2008,
which shows how Olaf Scholz and Vladimir Putin meet
at an event in Beijing.
So, now look at the AI agent.
Is there any media report for this?
Or was there a completely different politician
to see the picture instead of Olaf Scholz?
And does he then provide a dashboard for the fact-checkers
with which they can then verify or falsify the picture themselves?
Even with translation, by the way,
if you now get media in Arabic or Japanese.
That sounds really useful and super exciting.
I would really like to try that out.
But Sherlock Carina still has a question,
and we media-makers know that there are serious sources.
Well, there are some serious sources,
and there are very serious sources on the Internet.
Yes, exactly. Of course, they also have that in mind
with the development.
So Tim Pultz, he told me that sources are different,
quality media and fact-check editorial
that have a relatively high journalistic standard
are of course much stronger.
But, and I find that really exciting,
do you know that for your own editorial,
yes, ultimately, your own weight, which media you find more out?
It makes sense, for example,
if you report on so multi-perspective topics,
such as climate change,
then maybe climate reporters report on other aspects,
such as the trade plate.
Okay, how far are we now away from the fact
that we can simply use Gretchen AI here in Germany?
So Gretchen AI says that they cut very well
by evaluation tests for the fact-check editors.
With which you can go to the media.
That must certainly continue to be developed
and especially continue to be trained.
But it is still impressive,
when they started sprint-challenging 13 months ago,
they actually only had a plan.
And now they have a product,
and they want to continue to develop it.
And then it is also said that
systems in newsrooms are being built, Tim Pultz.
But you also want to go into the insurance industry,
because there would be long-term damage pictures sent.
And of course, the insurers have to see,
if this is real or if this is fake.
Yes, or the defense sector has said, Tim Pultz.
So you really want to get ready there.
That gives me a bit of hope,
but I'm very curious now,
because that was number one now.
I hope that the other two can also offer so much.
Yes, I would say so.
And as soon as we get to that,
it will be a bit special,
because this team founded Christoph Biel.
And he is, I would say, a pretty funny guy.
Imagine you meet someone at a party,
and then you should explain to him,
what you are doing for an AI detector,
how will you explain your product?
To run for a party.
Yes, but that is not the special thing about Christoph Biel,
but he is not a typical AI researcher, you can say.
I come from the film, and I also make films.
And I have been working with technology for many years.
We have already worked in other projects before.
I have also lived in Argentina for a long time.
We have just moved to Germany.
Before, I worked with augmented reality, with virtual reality.
Okay, so I would say that someone who just came around a lot,
who has experienced a lot,
but in the core he is now a director?
Yes, exactly, director and film producer.
But he has already started to trend his own vision models in Argentina,
and has worked with Agustín Rousseau,
a data scientist from Argentina,
who studied data science at that time.
And then Agustín Rousseau has applied for a job
at Christoph Biel, to write chatbots.
Okay, he was my boss, and then we...
Then they changed the company into this project,
and then they started researching the area.
And then they used deepfake detection.
Okay, so now he is applying for a sprint competition again.
Yes, now he is applying for a sprint competition.
Exactly, and the project is originally called Cinematic Context ORAI.
And the brand is called ItsReal.Media today.
And the detector he is based on
knows that Christoph Biel has gathered as a film director.
In my head, there is a door,
he knows how to insinuate things,
but does it really help with AI detection?
Yes, above all, he says it helps,
because he understands pictures very well
and always looks at the whole picture with it.
It's like a reverse engineering.
When you make films, you take pictures,
and when you take pictures in advance,
they are completely apart.
You just look at each part of the picture.
How intensive is the shadow?
What direction do the textures have?
How is the skin formed?
What positions do the body have in the picture?
What is the difference with the rest of the picture?
How is the noise in the background?
What frequency does the picture have?
How do the different channels look
between red, green and blue?
Yes, and that's why he and Agustin Rousseau
and the rest of the small team
can really look at each other for two months at the beginning.
AI images and videos.
Oh my God, it looks very exhausting.
Yes, I think so too.
Let's compare what is actually different from real images.
And these findings are then flowing into the detector.
Okay, that's very exciting.
Because we've already hinted at it before.
Actually, it's the fact that deepfake detectors
look at many images first.
So learn about the images, what is deepfake, what is real
and then recognize patterns that we humans
usually can't understand at all.
As often with AI, you can't look into it.
And the idea is now that you make visible, or what?
Yes, well, above all,
now you have found yourself a category
where people have decided that this is real, this is deepfake.
Of course, this is totally useful for the clarity of the models,
which is always an important point.
And this knowledge about what is different from a real image
than in a deepfake is then flowed into different filters.
There are about 15 of them in the detector
and according to this filter, different AI models
are trained with huge data sets on images.
And three channels of this filter network
run at the same time and the whole model
then compares these three channels together.
And such a filter can, for example, look at how coincidental
is, for example, a pixel for the next one.
Then we have very simple technical analysis,
which also works, for example,
over color, over shadows, and so on, over textures.
And then we really haven't analyzed patterns yet,
where we really look at what is repeated in the picture.
When different people are in a picture,
how different is their movement?
Because the team just found out,
when they looked at the whole picture,
that movements generated by the person in AI
contain much more repetition than with real content.
And the advantage of this is really,
because with the background knowledge from the film,
you can get such a future-oriented,
effective selection of deepfakes.
And it seems to work quite well, too.
So the two said that when evaluating
such directors, where they are tested on large data,
they actually cut off quite well.
Yes, I hope so. But I'm just wondering,
when we think of "it's real.media",
they certainly have the same problem with generalization.
So, in fact, their generators
probably have to train constantly with deepfakes.
Yes, totally. All these deepfake directors
who have said from "it's real.media"
that they had to train every six to eight weeks.
Wow, that's a lot of work, in any case.
Totally. And through the film knowledge,
they may still have an advantage,
because they may have other and maybe even better categories
after you evaluate the pictures
and after you train the directors.
By the way, another problem,
what if you have an original picture
and then there was only a small detail
that changes afterwards?
"All the deepfake directors are difficult,
even for those from "it's real.media","
says Christoph Biel.
What modified pictures are,
it always went very badly.
To be honest, with our model,
we always had problems.
And we decided a second model
only for modified pictures
with other forensic techniques.
And we're doing that right now.
And we promised today,
that it will be finished next week.
So, David will be proud of that.
We're working together with
an image rights organization,
who want to introduce labels
for films, which are 100% real,
which are fictions, for example,
and then also generated with AI.
And Christoph Biel also says,
that it's important for him
not to forbid AI-generated things,
but to make it clear,
clearly, what is real and what is fake.
In addition, you can let yourself
brand from "it's real.media."
So they scan your website
and you label it as "it's real".
There are also some small platforms,
which have that.
And for personal reasons,
Christoph Biel wants to make
that for NGOs in the future.
And they also want to bring
a free export to the state.
Then you can ask for X.
Is this picture real or not?
And he evaluates this based on the detector.
And Christoph Biel says,
that he wants to make this technology
accessible to as many people as possible,
so that we can see deepfakes.
But also interesting,
he is not a single enemy of this technology,
but sees as a researcher
also the chances, for example,
for the film, that we
come to a world where you really
have to mark clearly what is real and what is fake.
Okay, that's a very nice
translation actually for the third one.
Because if I've really noticed,
it's about thinking about it differently.
Namely, to go out of real pictures
and to mark real ones and not,
to mark the wrong ones, right?
Exactly, and that's what
the project "Deep Shield" did to the task.
This is a cooperation of the company "SekuBlocks"
with the industry-oriented
business community, short "IABG"
and also the Federal University
in Munich for "IABG".
This is a tech service for
car, air and space travel,
but also for the military.
That means that the employees
have been dealing with the issues for a long time,
how safe information processing can work.
Martin Riedl tells us that
he is the resource manager
of the Innovation Center in the "IABG".
"Detection" does a lot of things.
We said, "Okay, we would like to do something
differently and not connect us
with this tech-race, but instead
from a cryptographic perspective
to secure the information
that we can achieve as a trustee."
And "Deep Shield" actually
is part of a NATO training.
So they are involved,
where it's about, that the troops,
and the technology of the NATO
and their allies are running together.
And then it's about
that you can exchange military pictures
for 100% security,
for example, pictures of drone
or drone sightings.
Okay, honestly, I would never
get the idea. That's a great approach.
But how do you want to make sure
that your pictures that you really mark
are real and that nothing is wrong?
Yes, technically,
I have to admit that it's a bit demanding.
You need three components
in the system, basically.
You need a watermark,
that clearly shows that this picture is real.
Then you need a verificator,
who is testing it, when the picture
is added, the watermark is still there
and the picture has not been changed
in its content.
And then, to transfer this information,
you need something in between,
which, so to speak, combines both.
So a system, which saves the information,
the picture and the watermark
can be added.
Okay, you've already explained it quite well.
I'll try to summarize it again.
We have information, which is written
when the picture is recorded,
which is then placed somewhere.
And it has to be verifiable again and again.
But let's start at the beginning,
how to do it so nicely.
The information has to be written down.
How does that work?
You have a kind of signature signature
in the original picture,
which is hidden directly in the data.
For this, you need a so-called private key,
something that only knows the device,
that the picture is recording
and otherwise really nobody.
And with the help of this key,
a computer program can then write
a unique signature in the data of the picture.
And this private key,
it always comes together with a public key,
which can then make this signature
readable again for machines,
but not write it yourself.
And the private key, which you can theoretically
put somewhere in a operating system,
Christopher Langewig tells that this is the CTO
of Secure Blocks.
But if I am in a high-security environment,
then I need it on an electronic level.
And there, if my drone is recorded
by the enemy,
then I don't want this key
to be extracted there,
because then I can simulate this drone
and operate it in an infarction.
Okay, but is this key
then built into the drone?
Yes, at least that's the goal.
But you can also imagine that
for example for a camera on a cell phone.
But that's not all now,
because you might want to know
afterwards, if a picture has been changed,
then the content has changed from the picture.
And for that, semantic data
will also be raised about the picture,
very simply presented in AI,
almost the picture in a text together
and also other techniques.
But to verify the picture later
or to falsify it later,
you have to somehow lay down this information
about this picture,
which you just talked about.
How does that actually work?
We have a technology
that you already know since the
2010s, especially from cryptocurrencies
like Bitcoin, namely the blockchain.
Yes, I have told you all the time
when you talked about it.
Now you have already thought about
what to do with blockchains.
You have just said that often
in terms of cryptocurrencies.
Well, if you really want
to tell it in full space now,
for example, where it is written,
for example, how many bitcoins I have,
such a kind of data management system.
Yes, you can understand it like that.
I try to explain it quite simply
how it works.
That's your job today. You always have to break everything down.
I don't find it that easy at this point.
But you have information that you want to write.
In the case of Deepsheet,
that's about the water sign you hear in the picture.
Now imagine you take a block and a pen
and write it on it.
Of course, that's not particularly safe
because no one else can take a block
and a pen and write it on it.
This water sign belongs to a completely different picture.
But what the blockchain does,
it saves as a local copy
what you have written on the block,
not only on your computer,
but on a whole network
from many computers that are all connected.
And with that, this information is decentralized.
That means, if you see someone
packing another picture to the water sign,
then the network, that's not true,
because all computers in the network
have information.
Okay, so I have to imagine that.
All computers in this system
have the information for the water sign
and for the picture.
That's why it's almost unchangeable
for all individuals.
Exactly, because they are connected
with all the other information
to other pictures and other water signs.
And that depends on a kind of chain
from information.
If a new picture is seen with a water sign,
then a new block
is inserted into this blockchain
and it's about fake-true computer operations.
And the blockchain
is really a block
longer in this system,
this network, fake-true security.
And at the same time,
this technology is also open source,
says Christopher Langewig, the CTO of CQblocks.
That means, everyone can
operate this blockchain,
prepare it, it is worldwide in use.
So it develops more and more
to a standard in this environment.
We also use this technology.
Simply because, for example,
for cryptocurrencies, many instances
with the technology are already trusted.
And then the verificator
can now read from the blockchain
at the end of the system.
What water sign should the picture have
and what are the semantic
original information about it
and can verify real content like that.
OK, but in retrospect,
you can also understand
on the blockchain
what changes, for example,
to remove the water sign.
Yes, theoretically, removing the water sign
is actually no problem for the system.
Because when it is done, then it is clear
that the picture was somehow processed.
It is no longer the original.
And add an original water sign later.
That doesn't work, because you then
miss this secure key.
What is actually a challenge
for the water sign and is
when the picture is uploaded to social media,
is turned, is cut,
exactly numbers like robust
the water signs are currently.
They didn't know that in the interview.
But the robustness score
has improved in the last five months
by 13 percent
and also the speed
to write the whole thing in the blockchain
has increased by 75 percent again.
And currently, they improve their product
further in the framework of a NATO competition,
where their product is used in NATO exercises,
I had already said.
Yes, and also there are talks, for example with Europol,
we have to exchange pictures internationally.
But Martin Riedel,
he can be well-prepared
for secret services,
when the information, for example, is exchanged.
These are projects that of course
could bring a lot of money
to develop the technology further.
But I don't want
to let society fall more under the table,
because here you profit,
in any case, massively because
you have the risks of misinformation
or misinformation
in a world that is becoming increasingly difficult
and then can significantly reduce.
Would this technology also require a certain width?
And that is a prerequisite.
Yes, and with that a certain width
you also plan to develop a kind of chip
with which this technology
can then be built directly into devices.
Plus, there are still a few pilot projects
and of course you have to constantly develop the product.
Max, now you have presented
all of us very remarkable projects.
So thank you for that.
And I remember that it was a seven at the end.
So seven teams that have met each other there.
And I think you really
chose three incredibly exciting things out of it.
But please don't pick me up
any longer. Who won it now?
Yes, definitely one of the three that I presented.
So either Cretin AI,
who reviews technology images and videos
as fact checkers,
or It's Real.media,
editors on the knowledge of professionals
in the field of film foundation,
or Deep Shield,
the one with secure technology in the devices themselves
and the blockchain in real life.
I think I want to mark
the whole Fake's Fetching Secure.
What do you think? Who won?
So I would say
of the three projects,
maybe it's because of my job,
I'm sorry, Cretin AI,
because I was able to understand best
how it actually works.
So maybe they won?
Let's take a look.
We're going to the late afternoon
of the Challenge.
So it's just before 7 o'clock.
The winners should be announced at 7 o'clock.
Yes, all teams have pitched
the Jury final
and are now in a big room,
talking to each other.
And you really notice that there is
really tension in the air.
Then there will be prepared sex glasses.
And finally the Jury comes from
and only costs.
He takes the microphone.
First of all, thanks to all teams for their great work
and then he knows who won.
With loads of creativity
I have a very interesting technological foundation
and that is
It's Real Media.
That's so funny, I was thinking
if I were the Jury now,
I think I would have drawn
It's Real Media,
if I were from my personal
I would have drawn Gretchen AI.
So it's funny, somehow I still had
such a feeling.
In any case, he deserves Gretchen AI.
They actually became second place.
But still I would be interested
I have already speculated in my head
Why did you win?
Yes, Janus Kostath von Sprint,
who is also part of the Jury,
he told me, especially because
the team has made a huge development
in this 13-month Challenge.
The director Christoph Biel
is super innovative
and very special.
And also because her team
has actually cut off the best
during the evaluation.
All the directors were tested
by Sprint
and they cut them off the best.
Of course I also spoke to
Augustin Russo and Christoph Biel
right after the award ceremony
and she asked how it is now
to win the Challenge.
I am very happy because
I put a lot of work on that.
Yes, it feels really good.
We really didn't expect that.
It was a nice surprise.
I was just a bit overwhelmed.
Did you have a nice party tonight?
We still have to go to Mannheim
but we definitely celebrate at the weekend.
You were celebrated at the weekend?
Yes, of course.
Then have a wonderful time.
I really like how modest they are.
They sound so really surprising
and I don't think we understood
what actually happened.
Yes, I also had the feeling that
it would really be a surprise.
Thank you very much Max.
I learned so much and there are
exciting approaches in it.
I think they are really important
for our future,
as we will feel in the internet
and also in the real world.
Because deepfakes are always a big problem.
And I don't think that will be the last episode
where you and I talk about deepfakes.
Then we will look at it again.
I think so too.
Thank you very much Maximilian Brose.
My colleague, thank you for the research
and that you brought the day there at the competition.
I enjoyed it a lot.
Now I would like to talk about our listeners.
Because they always write us
many great emails.
For example, Uwe wrote
that we should talk a bit more about
European ideas and more about European chatbots.
I'm happy about such emails
because I think it's good if you challenge us
and give us some input on what we can do better.
Of course we want that.
And of course I'm more happy about this episode
because you presented German projects.
European projects,
the whole competition for European projects,
but yes, the winner teams come from Germany.
So then we did everything right.
And if you have the need
and really want to encourage you
to write to us,
to let us know your wishes,
maybe ideas, then you can write to us
under
[email protected]
And I understand that with the chatbots.
We also got these emails more often now.
The problem is always a bit
where research is being done,
which chatbots, so to speak,
are being discussed a lot,
where a lot of innovation is being put in,
where a lot of money is and unfortunately it is often
these big tech giants from the USA.
But of course we always like to
write something like that.
And next week there will be a episode with Fedi and Reif.
They were on the KONKON on the way
and there they recorded a episode
where there is a lot of content
that is trusted.
I think that fits again
fantastically to our episode
that we recorded today.
It's almost an endless transition.
So listen to it, it's worth it
and I say thank you, see you next time.
PodcastTranscript AI