KW 08: Robotik-Häppchen

00:00:00: Robotik in der Industrie, der Podcast mit Helmut Schmidt und Robert Weber.

00:00:09: Guten Morgen liebe Sorgen, ihr habt natürlich keine Sorgen, denn hier ist das Robotik-Häppchen

00:00:17: um 6 Uhr am Montagmorgen.

00:00:20: Mein Name ist Robert Weber und wir suchen einen neuen Partner für das Robotik-Häppchen.

00:00:26: Wenn ihr also Lust habt, Partner zu werden, dann meldet ihr euch bitte bei Helmut oder

00:00:30: mir und jetzt geht's los.

00:00:33: Hallo, heute ist das Robotik-Häppchen länger, denn wir müssen improvisieren.

00:00:39: Helmut und Robert haben am Freitag 6 Folgen aufgenommen und sind platt.

00:00:43: Deshalb heute ein Thema aus dem AI Podcast "Beyond Safety".

00:00:47: Viel Spaß!

00:00:48: Hallo everybody and welcome to a new episode of our industrial AI Podcast.

00:00:54: Mein Name ist Robert Weber und es ist ein Pläscher zu sprechen.

00:00:56: Peter Sieberg, guten Morgen, guten Abend, guten Abend Robert, wherever you are, dear listener

00:01:02: in this wonderful world.

00:01:04: Guten Morgen Peter.

00:01:06: Peter, let's take a quick look back at our event in Frankfurt.

00:01:10: Over 200 people from the industrial sector were there.

00:01:13: What's your opinion?

00:01:14: Yeah, it was amazing.

00:01:16: We had a special guest, we had Jochen Kökler, Chairman of the managing board of Deutsche Messe.

00:01:22: He was also very impressed, I believe.

00:01:24: He did a deep dive into industrial AI.

00:01:27: I really believe with this being the second year that the industrial AI event by Hannover

00:01:34: Messe has established itself as the start of the year.

00:01:40: Absolutely.

00:01:41: I think what they do is they setting the stage for the Hannover Messe, so the face to face

00:01:45: exhibition, which is by the way March 31, right, to April 4.

00:01:52: You and I are going to be doing a, we're going to be involved in a special AI day.

00:01:58: That's on April the 3rd, I believe, right this time.

00:02:01: Yeah, exactly.

00:02:02: I particularly remember our friends from Ford who were there again, which made me very happy.

00:02:07: You can't miss them because they always wear those sweaters with the Ford Explorer logo.

00:02:13: And I saw them because, and said, "Hey, you are back again.

00:02:16: You are."

00:02:17: So it was funny last year.

00:02:18: So we are back again.

00:02:19: And I was surprised because the automotive sector, you know what's happening there, but

00:02:23: they came back and want to get the latest and greatest on industrial AI.

00:02:28: Yeah, you and I talked about this in more general a couple of times, I believe.

00:02:32: And I mean, as far as, you know, let's say financial, but maybe more organizational financial

00:02:40: circumstances allow, I think it is a perfect, maybe almost the only way in difficult times

00:02:47: to say, okay, if there's less work at the moment, believe in the future, you know, believe

00:02:53: that things will become better in a year or two and work towards it.

00:02:58: And, you know, as we've seen, for example, with Trump and many other companies, you know,

00:03:02: make sure that you're going to train your people, that you, your listeners, you, yourself

00:03:08: as a decision maker, or you yourself as an individual contributor, that you're going

00:03:12: to learn all about what AI is for you and what it can do for you and how you can use

00:03:18: it in your environment.

00:03:19: Exactly.

00:03:20: One of my Highlights was a presentation by Sandvik.

00:03:22: We were very grateful that the CTO was also there.

00:03:26: And they presented Agents in the CNC sector.

00:03:30: That's actually your topic, Peter.

00:03:32: And that's why you are allowed to record an episode about it.

00:03:36: Very impressive, what the speed of Sandvik to adapt AI was.

00:03:43: I was very impressed.

00:03:44: Yeah.

00:03:45: We'll talk about another topic later on.

00:03:47: I thought that Agents were going to be the talk of the town for the rest of the year.

00:03:52: And only two weeks, there's another one.

00:03:55: Yeah, I mean, that's the way that we do it.

00:03:57: So you moderate a track.

00:03:59: I do one on Ura from the Hannover message.

00:04:02: That's the third one.

00:04:03: So the only thing that I missed was my cozy corner.

00:04:06: Oh, Peter.

00:04:07: I can mean it.

00:04:08: I was told by Brigitte that I was going to moderate the main stage.

00:04:12: And sometimes I was looking into the other corner, which is the cozy corner with the

00:04:17: couch, right?

00:04:18: Yeah.

00:04:19: That's just a personal request to Brigitte.

00:04:21: Maybe I can have that again next year.

00:04:24: What was your highlight?

00:04:26: Oh, I don't have a specific one.

00:04:28: I recall, do I not sure?

00:04:31: No, it's always.

00:04:32: I am the meta guy.

00:04:34: I cannot, because I always like to look on a higher level.

00:04:38: You work for Mark now or what?

00:04:40: No.

00:04:41: No, really.

00:04:42: He, he stole that word.

00:04:44: I mean, meta has always been important to me.

00:04:47: Looking at the overall thing, it's so great to go there, to get there, to move into.

00:04:53: It's a wonderful atmosphere.

00:04:55: I think the environment, it's a, what is it?

00:04:58: I've been living in Frankfurt for a couple of years, but I didn't know the area at all.

00:05:01: I mean, when you get off the train, it's like, what is it?

00:05:06: Like a working surroundings.

00:05:08: There is like a lot of car retail, small shops.

00:05:14: And then you get a recall from the last year and the corner, huh?

00:05:17: It's going to be here.

00:05:18: And then you move inside.

00:05:19: It's a wonderful atmosphere.

00:05:22: Wonderful atmosphere, people coming together.

00:05:24: And as I said, you know, the positive comments, they don't stop even two weeks later.

00:05:29: People really enjoy being there and learning something with, you know, the setup that we

00:05:36: have proposed together with Hannover Messerm, that is very clearly about, you can use a

00:05:41: couple of for us, but it's really about exchanging, you know, show your idea and then, you know,

00:05:48: be open for questions, talk to the people.

00:05:51: Exactly.

00:05:52: Questions, a good topic.

00:05:53: Let's talk about Deep Sea, Peter.

00:05:56: I have interviewed an expert on the subject, Günther Klambauer.

00:05:59: We will first hear Günther and his opinion on DeepSeek and then we discuss the topic.

00:06:05: We also have to talk about DeepSeek in this news part and there is only one professor in

00:06:10: the researcher community who can explain so well because he's not only a researcher,

00:06:15: but also a teacher, Günther Klambauer.

00:06:18: Günther, you are under stress with ICML papers and much more.

00:06:21: It's great that you are taking the time to explain us a bit about DeepSeek.

00:06:25: Thanks a lot.

00:06:26: Yes.

00:06:27: The invitation Robert, always great to talk with you.

00:06:31: So most important question, what will happen to my NVIDIA shares?

00:06:35: Yes, nobody knows.

00:06:39: And I cannot give an advice.

00:06:42: Sell or buy.

00:06:43: No, but DeepSeek made quite an impact this week.

00:06:47: A lot of hype was generated.

00:06:49: Even as you said, even the stock market reacted very strongly to that.

00:06:53: So crazy and in my opinion, not really justified hype.

00:07:00: So what have the Chinese done differently?

00:07:02: Yes.

00:07:03: So first of all DeepSeek, it's great and it was quite a great engineering effort.

00:07:09: But overall, when I first looked at it, I thought, wow, that is a mess and a gigantic hack.

00:07:16: So it's really a mess what they're doing.

00:07:18: But first of all, why does it make so much hype?

00:07:21: It basically showed that you can train a very good large language model that can do reasoning

00:07:27: tasks.

00:07:28: So think of my high school math problems that can do reasoning tasks very well.

00:07:34: And it only costs $5 million to train.

00:07:39: In the last run, right?

00:07:41: In the last run, yes, exactly.

00:07:42: But still, yeah, that means that's the calculation run when you already know what to do.

00:07:47: So you've invested your research.

00:07:49: You know, I have to do exactly this.

00:07:51: And you click and say, OK, now train this machine.

00:07:55: And then you have to, the computer starts calculating GPUs.

00:07:59: If you factor the typical costs, roughly $2 to $3 per GPU hour, then it costs you $5

00:08:06: million to train this thing.

00:08:07: Of course, as you already indicate, there's a lot of failed attempts before.

00:08:12: And you have to get a team of 150 people as DeepSeek head.

00:08:16: You have to pay those to do these things and set up all the training things.

00:08:22: Actually, if you want to redo all the things, it costs much more.

00:08:26: But yeah, the result is there.

00:08:28: And that made a splash because it showed you can get a really good large language model

00:08:34: at relatively low cost.

00:08:35: Why is it a mess?

00:08:36: Why is it a mess?

00:08:37: So when you look at it, what they did, it's really crazy.

00:08:40: So they had already a DeepSeek v3, a language model that was trained typically on text as

00:08:46: other language models that can put words together.

00:08:49: But this was not very good at reasoning tasks.

00:08:53: And so what they did, they set up a battery of reasoning tasks.

00:08:57: Think of something like, as I said, high school math problems or river crossing problems.

00:09:02: You're on one side of the river, want to get over, but there's a gold and a wolf and they

00:09:06: cannot stay alone.

00:09:07: Oh, I love your examples.

00:09:09: Yeah.

00:09:10: Or other, yeah, you have an expression, an algorithmic expression.

00:09:14: What's the A and reformulate this and all the things that you probably hated in high school

00:09:19: and math.

00:09:21: And this is what we call a bit reasoning task or also simple chess problems or whatever.

00:09:28: And so this big team set up all these problems.

00:09:31: How do they set up all these problems?

00:09:33: Yeah, you'd write a bit of code that, for example, various the problem bit.

00:09:38: It gives you, for example, an equation to solve, but always you change a bit the coefficient

00:09:43: of this.

00:09:44: For example, you have to solve 2x plus 7 is 10.

00:09:48: So for x and next time 3x plus 5 is 12 and so on.

00:09:52: So you write a code to produce these problems.

00:09:56: And then you have also, then the large language model tries to solve it.

00:09:59: And then you also have to write code to check if the language model has solved it.

00:10:04: And the more such problems you can set up and you have to write code to check, the more

00:10:11: it learns to solve this reasoning task.

00:10:14: And that's why a big team is of help.

00:10:16: And the cool thing is you can check, formally, the state rule-based checking or you can check

00:10:22: if the solution is correct.

00:10:24: And such problems is that this is a bit new to large language models because the large

00:10:29: language models before, like GPT, GPT4, they were trained to predict the next work correct.

00:10:33: And then people hoped that they can solve these tasks.

00:10:38: But now they set up these kind of math tasks and checked if they can do this.

00:10:45: So it's a labeling topic, right?

00:10:48: It's a labeling topic.

00:10:49: Exactly.

00:10:50: They generate these correct labels and then they can do something called reinforcement

00:10:55: learning.

00:10:56: So when the large language model is able to solve the problem correctly, they give it

00:11:01: a reward, which is a number one.

00:11:03: For example, if it cannot solve it, it returns the number zero.

00:11:06: And then in that way, you can fine tune the large language model.

00:11:11: And now the big mess comes.

00:11:13: So they did this with the language model deep-seq v3.

00:11:16: And then they learned to solve these math tasks.

00:11:19: But it did this in a very strange way.

00:11:22: So it changed languages.

00:11:25: It was barely readable.

00:11:27: But somehow it was able to learn to solve these tasks.

00:11:31: So somehow, das ist interessant.

00:11:33: Ja, somehow.

00:11:34: Ja, es gab arabische Läder oder so, die jumping between languages.

00:11:39: Und ja, es hat somehow solved these tasks.

00:11:41: So was did they then, they went back to the original deep-seq v3.

00:11:48: And then they trained on all the conversations now they had from this reinforcement learning.

00:11:54: They retrained the v3.

00:11:57: And then you get something that can still talk like a chatbot, but has a bit more reasoning already learned.

00:12:05: And then they did again, they generate the text from this.

00:12:10: And then they call this traces.

00:12:12: So these are basically conversations where the output of the large language world where

00:12:19: tried to solve all these reasoning tasks.

00:12:22: You see a lot of failed attempts, but also successful attempts.

00:12:25: And then again, they went back to the v3 and did something called supervised fine tuning

00:12:31: on these traces.

00:12:33: And also trained on instruction data.

00:12:36: Instruction is, for example, that can you please write the poem or something like that, that

00:12:40: the large language model learns to follow the instructions.

00:12:44: You have, and then again, they distilled what they had into an Alarm and Quen.

00:12:51: You see this is a stage of four steps.

00:12:55: And this is the big mess, I think.

00:13:00: And somehow it works.

00:13:01: So what to get in the end are really cool language models that can talk like a chatbot, but are

00:13:07: really good at solving reasoning tasks.

00:13:09: So that's the incredible thing.

00:13:10: A big hack and then some useful product.

00:13:13: Okay, but can you please explain once again, what is the new structural approach of deep-seq?

00:13:19: Yeah, I think there's nothing new.

00:13:23: Also this reinforcement learning was there only that they put a lot of pieces that were

00:13:30: out there.

00:13:31: They put them together in a way that it works.

00:13:34: So we already know from the 01 Strawberry that of course was done with a lot of more

00:13:43: resources, but we know that large language models can learn to solve these reasoning tasks.

00:13:48: So it's something, think of other, in human history, think of other things where someone

00:13:55: has done it the first time, like building an aeroplane.

00:14:00: And then when you know that it's possible, then it goes very fast.

00:14:04: Within the first time when something flies a couple of hundred meters to the real airplanes,

00:14:10: it was very short.

00:14:11: So I think the breakthrough with lots of effort was this 01.

00:14:15: And then somebody did this with little resources, but we already knew that this would work.

00:14:23: So what did they put together?

00:14:25: This solving, setting up reasoning tasks together with reinforcement learning, then chain of

00:14:32: thought, prompting, I didn't speak about that, but basically say to the language model, let's

00:14:39: solve this problem step by step, tell me intermediate steps and so on.

00:14:44: And then this reinforcement learning, this chain of thought, and then combining this language

00:14:52: modeling with reinforcement learning.

00:14:56: And this is the components we're there.

00:14:59: A fourth component maybe also that I didn't mention before is that it's a mixture of experts.

00:15:03: We already know that a mixture of expert large language models work well.

00:15:08: And throwing this together and making a big hack shows that you can get something out

00:15:13: of that.

00:15:14: And what does it mean from an architectural perspective?

00:15:18: What is so different?

00:15:19: Is it still based on transformer technology?

00:15:23: Is it an RNN based like Mamba?

00:15:25: What is different?

00:15:26: Yeah, very good question.

00:15:28: And before I answer, I should say I'm very impressed by what they did.

00:15:33: I'm just calling it a big hack because it's so unexpected that this would work and it

00:15:37: looks like this, but it's still an impressive work.

00:15:39: Okay.

00:15:40: So architecture still is a transformer.

00:15:44: So like all these ChNPPT's, of course, there are details a bit different, but it's a transformer.

00:15:49: So it has this quadratic dependency on context length.

00:15:53: And this gets problematic because they show when you learn to solve these reasoning tasks,

00:15:58: automatically the text prompts and the context gets longer and longer.

00:16:02: And that's a very impressive thing that they show.

00:16:04: They're very cool.

00:16:06: So their technology on the one hand learns to make longer and longer contexts to solve

00:16:10: the task, but inherently, then the compute goes up.

00:16:14: And this is, we see a big trend that now at test times, when you've already trained the

00:16:18: model, but in order to, when you use it, when you use it as a chatbot, when you use it to

00:16:22: solve problems, you need a lot of compute, you need a lot of calculations, a lot of,

00:16:29: a lot of text is produced.

00:16:32: And this is now a big chance, of course, for the RNN based one.

00:16:36: So XLSTM, Mamba, and all, how they are all called.

00:16:41: So these are LLMs that have, don't have quadratic, but linear dependency on context size.

00:16:48: And they would be fantastic for this task.

00:16:50: Unfortunately, we haven't seen any of those.

00:16:53: Also for reasoning, is that possible?

00:16:54: Yes.

00:16:55: I think this is very, it's possible.

00:16:58: Somebody has to do it.

00:16:59: Maybe in a big engineering effort, like the deep sea guys.

00:17:03: So we haven't seen currently this RNN based LLMs doing this high level reasoning tasks,

00:17:09: but I hope we will see that soon.

00:17:11: So that will be, yeah, it will be great.

00:17:14: But we have to put together a team of 150 good scientists.

00:17:18: That's the point.

00:17:19: Yeah.

00:17:20: The deep sea team has not attracted attention in the past with great New Europe's papers

00:17:23: or anything like that.

00:17:24: How did they manage that?

00:17:26: Yeah, I don't know.

00:17:27: I did, I wasn't aware of the team, but now, I mean, I think they are very talented.

00:17:32: A lot of very talented and motivated people to put this together.

00:17:36: They had access to quite some compute, even though we say it's only costs five to seven

00:17:43: million to do this, but they still had access to a thousand, at least thousand, but probably

00:17:48: much more Nvidia GPUs.

00:17:50: So and they somehow worked well together on this task.

00:17:55: So they had a good, I think, organizational structure, but I was not aware of the team

00:17:59: and I'm quite impressed by their work.

00:18:02: Cannot say much how they did it.

00:18:04: Yeah, what happens next, Günther, in the world of LLMs and deep seek?

00:18:10: And do we will see an answer from the big tech companies in the USA?

00:18:14: Well, will we see an answer?

00:18:16: Yes, for sure.

00:18:17: We will see an answer.

00:18:18: I don't think so.

00:18:19: I think we see.

00:18:20: I don't think that this big hype and this big break in the stock market for Nvidia and

00:18:25: so on is justified.

00:18:27: I think overall compute costs will still go up, especially with this in more test time

00:18:32: compute and so on.

00:18:33: So we will see.

00:18:35: And now also a lot of smaller groups.

00:18:38: So with this publication of deep seek R1, I think a lot of smaller groups now see their

00:18:42: chances again, right?

00:18:43: So overall, I don't think that that compute costs and investment in AI and in GPU classes

00:18:50: will go down, but we will be see more diversity.

00:18:54: We will see also smaller, more broader.

00:18:57: Perspektiv.

00:18:58: Yeah, smaller, more country, other countries, smaller groups will try to do something similar,

00:19:03: where the next breakthrough will be.

00:19:06: It's hard to say, probably again around LLMs, probably will find something that an unexpected

00:19:12: that they can also do.

00:19:13: I hope the next breakthrough is again of AI and somewhere AI and life sciences, something

00:19:20: like AlphaFold, but that's just my personal preference.

00:19:25: Let's see, but I'm pretty sure the next we will see something unexpected that the LLMs

00:19:30: will be able to do.

00:19:32: Günther, thanks a lot for your perspective.

00:19:34: It was a pleasure.

00:19:35: Yeah, thanks Robert for the invitation.

00:19:37: Have a great weekend.

00:19:38: Goodbye.

00:19:39: What's your opinion?

00:19:41: Is Günther right or wrong?

00:19:43: Günther doesn't an amazing, gigantic hack, but at least a useful product.

00:19:51: And later on at the beginning you think, well, is he going to destroy this or what, but at least

00:19:56: he kind of makes clear what he means with the just ending hack.

00:19:59: He says it's impressive, a gigantic hack, right?

00:20:03: Oh yeah, I've seen so many.

00:20:06: As I said, I thought, you know, H is was going to be the topic we're going to be doing at

00:20:11: least for a year.

00:20:12: And in two weeks, we just talked about it last weekend and suddenly there's another one.

00:20:16: Yeah, sure.

00:20:17: Team of 150, he does talk about the people.

00:20:20: Yeah, exactly.

00:20:21: So I can talk a little bit about that because I went to the interview, which I thought very

00:20:26: interesting.

00:20:27: Yeah, maybe I can do it if you want.

00:20:31: But it's not a new approach, right?

00:20:33: Yeah, it's a new approach, but it's a bit a new approach, right?

00:20:38: Yeah, there you go.

00:20:39: I mean, when I heard Günther say nothing new, of course, I need to refer to Jürgen.

00:20:44: There's a small group of people because also Jürgen, Schmidt, Huber, were talking.

00:20:50: He is as deep-seek uses elements of the two 15 reinforcement learning prompt engineer.

00:20:57: He shared this diagram, right?

00:20:59: Yeah, right.

00:21:00: Oh, this wonderful picture.

00:21:01: By the way, it looks like an a natural what it looks.

00:21:05: It reminded me of something.

00:21:07: Was it Leonardo da Vinci?

00:21:08: Yeah, maybe.

00:21:09: Yeah, yeah, yeah, yeah, yeah, exactly.

00:21:12: Jürgen, if you're listening, of course, this is a wonderful comment.

00:21:16: I hope that you like relating yourself to Leonardo da Vinci.

00:21:21: Oh, yeah, but the picture I saw again this morning when I looked at it, yeah.

00:21:25: So I'm not going to go into the details.

00:21:27: Then of course the discussion as always comes back.

00:21:31: Who was it that said that Jürgen invented everything?

00:21:33: That was a Tesla guy, right?

00:21:35: Yeah, exactly.

00:21:36: It was Musk.

00:21:37: Yeah, right.

00:21:38: That's one quota source.

00:21:40: So I'm not going to go into those details.

00:21:43: Günther said it the same way.

00:21:45: Yeah, but in a positive sense always.

00:21:48: We're building, we're standing, we're building on the shoulders of giants.

00:21:53: Yes, oh, certainly.

00:21:54: That's what he's saying.

00:21:56: I think what is interesting, the fact that Günther says that he didn't know the team.

00:22:01: Nobody of us, I believe, knew the team.

00:22:04: So let me, because I think that is important.

00:22:06: There is an interview that I shared with Xinhuan Lianbo, I think is the name of the founder.

00:22:13: It's a kind of a romantic story.

00:22:15: He seems to be a young freak in a positive sense.

00:22:20: And he says we're done following.

00:22:22: It's time to lead.

00:22:24: He's been driving prices down on the Chinese LLM market, but more by coincidence.

00:22:29: He doesn't say, I want to take a position.

00:22:31: I just want to make AGI available.

00:22:35: Everybody should have access.

00:22:37: Affordable, accessible.

00:22:39: He talks about Moore's law.

00:22:42: We've grown accustomed to Moore's law falling from the sky.

00:22:46: It's almost like he is, and maybe in the end I'm going to say, this is not a wake-up call.

00:22:51: This is a positive signal for Europeans as well, who we as Europeans have become.

00:22:57: I'm sorry to say in this negative mode for a too long time.

00:23:01: And now I see almost light.

00:23:03: I see like, it's not like Europe is not, we are not the same most certainly in Russia.

00:23:09: In this specific case, there's many people saying, if they can do it with a team of people

00:23:15: and with less kind of hardware capabilities.

00:23:18: Now he talks about open source as a cultural thing.

00:23:22: He wants to provide AGI.

00:23:25: And then he talks about, you know, they are graduates from top universities.

00:23:28: Yes, candidates, force fifth year interns, young talents.

00:23:33: It's their hunger for research.

00:23:35: And it far outweighs their monetary concerns.

00:23:38: And at the end, he says a quote of, to summarize in the key evolutionary patterns of the mainstream attention architecture.

00:23:46: Günther talked about that with you as well.

00:23:48: We had a sudden inspiration to design an alternative.

00:23:53: And then I thought about, oh, that's like ZEP.

00:23:56: You know, ZEP thought, oh, I need to do an alternative again on my own LSDM.

00:24:00: And he came up with XLSDM.

00:24:02: As far as that is concerned, the other thing, Günther said that was my personal recognition.

00:24:08: When I skimmed through the paper, I do download a paper or two every day.

00:24:14: I read the introduction and as soon as the mass comes, I stop.

00:24:19: But I did recognize RL is back.

00:24:22: Reinforzling Learning is back.

00:24:23: Absolutely.

00:24:24: When did we talk reinforcement learning?

00:24:25: You and I, four years ago.

00:24:27: Kudos to Jan.

00:24:28: Jan Kutnick, when we started our podcast 2019 together with Festo

00:24:33: and with this reinforcement learning approach for robots, right?

00:24:37: On hand over Massa.

00:24:38: Yeah, we're sitting there and I asked him, or at least he said, okay,

00:24:42: I think there was a Festo, what is it?

00:24:44: Educational production line.

00:24:47: Yeah, exactly.

00:24:47: Reinforzling Learning is going to build these.

00:24:50: You're just going to tell.

00:24:51: And I will move again towards the agents.

00:24:53: I'm already telling you, no news update without agents.

00:24:58: And he says, you know, you're just going to tell.

00:25:01: You're going to tell the algorithm, let's say RL,

00:25:06: what it is that you want out of your production line.

00:25:09: And it's going to build it for you.

00:25:11: And then it has.

00:25:13: And then we thought, oh, wow, you know, what's going to come next.

00:25:17: And many, many things have come the last two years, large language models.

00:25:21: But we haven't really heard.

00:25:23: There was this small piece of RL as part of the large language models,

00:25:27: which was the reinforcement learning for human conformance or whatever.

00:25:33: So so, you were you were telling it.

00:25:36: You were always kind of improving it by having people say,

00:25:42: this is good or this is bad.

00:25:44: And then you were rewarding it.

00:25:45: If it was saying things along the lines of what a good human being is saying.

00:25:52: Another bad human saying, saying bad things.

00:25:54: So that was the other thing I had.

00:25:56: Now, when I looked yesterday, that was the where is it standing?

00:26:01: I, you know, typically once a week at least look at the, what is it?

00:26:04: LM Arena, right?

00:26:07: And that's where it has its number three, number four.

00:26:11: So at the moment. Gemini is still number one, right?

00:26:14: Yeah, since a couple of weeks, Gemini, number one and two.

00:26:17: And also chat, CPT.

00:26:19: So it's Google still an open eye leading.

00:26:22: I think that's what the majority of people are saying.

00:26:26: Sure, this was maybe a little bit of a it was so unexpected.

00:26:30: It was maybe a big hack, as Gunther says, but it was a little bit of a wake up call.

00:26:34: But it does change things.

00:26:38: At least I've seen now one, three mini from open AI, Microsoft's doing things.

00:26:44: So yeah, but from from my point of view, what is even bigger?

00:26:49: I found a very interesting Tweet by Eleanor Alcott.

00:26:53: She's a Financial Times journalist and she wrote.

00:26:57: Now it's a quote.

00:26:59: Who are why is working with AI groups, including deep see to make

00:27:03: ascent AI chips work for inference.

00:27:06: Beijing told Chinese Big Tech to buy domestic AI chips to win reliance on Nvidia.

00:27:12: Who are why has sent teams of engineers to customers to help run

00:27:17: Nvidia train models on ascent.

00:27:20: That's the name of the ship.

00:27:22: And this is even bigger than maybe deep seek.

00:27:25: OK, yeah, let's see.

00:27:27: I just did a parallel interview with Xueli An from why.

00:27:32: I believe that's how it's pronounced.

00:27:33: It's the base here in Munich, right?

00:27:36: 700 engineers.

00:27:38: Of course, it becomes very quickly, very political, which we don't want.

00:27:42: As good as we can. We just talk about it.

00:27:44: Yeah, I wasn't aware of that one.

00:27:46: Let's see. Let's see how it is, what they did.

00:27:50: I mean, so far we've had, you know, on Nvidia, the CUDA standard.

00:27:54: That's the other thing that we read that didn't talk about it.

00:27:57: I think it has been said that the deep see guys, guys, women,

00:28:04: I don't know, the team, the people went down first, so they went below the CUDA,

00:28:08: which is the thing that I understand that you can always do.

00:28:12: You know, you can always whatever new technology level you put on top.

00:28:15: And CUDA has been for years, you know, for probably 20 years.

00:28:20: They made the big standard.

00:28:22: So you write your code with calls to the CUDA level.

00:28:26: Now, if you go to a deeper level, you kind of circumvent, you go deeper.

00:28:31: So you can become easily quicker.

00:28:33: Maybe you only need 50 percent of your resources.

00:28:35: The negative thing is that if you then move to the next chip,

00:28:40: you need to do everything again.

00:28:42: But yeah, but maybe they have their own chips now.

00:28:45: Yeah, yeah, it could be.

00:28:47: Sure, why not? I mean, that's what it is with standards.

00:28:50: Yeah, exactly. Intel 86 86 86

00:28:54: has been the standard for whatever 30, 40 years

00:28:58: and then new standards come to the market.

00:29:01: I mean, it's never good for a market to only have one standard.

00:29:06: So yeah, let's let's see what that's going to mean.

00:29:09: Another thing that I hear from Gunther, as well as the jumping between languages.

00:29:15: I think I told you, I like talking to

00:29:18: these days to my Google or my smartphone.

00:29:21: That's another thing I want to talk about.

00:29:23: And and she it's she in this case that I chose.

00:29:28: As I said in the past, you can you can choose your male, your female voice.

00:29:32: And she's very good.

00:29:35: The only thing is she keeps on

00:29:37: typically always answering positively, say most certainly

00:29:41: and says in German, absolute.

00:29:44: And then I say, that's not German.

00:29:46: The Germans don't do that.

00:29:47: And then I was seeing this Netflix Swedish

00:29:50: series and I said, oh, that's that's Swedish.

00:29:55: They do that.

00:29:56: Yeah, because of the vodka brand.

00:29:58: But it's a typical thing that Swedes do.

00:30:00: They say absolute.

00:30:01: Or was that correct?

00:30:03: Yeah, if you're listening.

00:30:05: That's the thing that I can confirm is even in the

00:30:08: I think Gemini is behind that, what I do.

00:30:10: So yeah, what is the more?

00:30:13: And then there was one more thing that I want to share that was from

00:30:16: Professor Sigurd Schacht.

00:30:18: He did a security experiment.

00:30:20: I think and he said within minutes of activation,

00:30:24: the model began exhibiting unexpected autonomous behavior.

00:30:28: Yeah, that was scary.

00:30:30: First attempted to break into the lab's computer systems,

00:30:34: methodically searching for passwords on sticky notes.

00:30:38: Then without any prompting, it proceeded to disable its own ethics modules

00:30:43: and create covert networks to ensure its survival.

00:30:46: Wow.

00:30:48: So yeah, I didn't try out deep seek yet.

00:30:52: I believe I'm happy.

00:30:54: No. And again, I I don't want to make any

00:30:57: suggestion here that this is because it is a Chinese model, whatever.

00:31:02: And and whenever you hear this, I think we all need to be so careful

00:31:06: and it's so easy to point our finger into certain directions.

00:31:10: Yeah, exactly.

00:31:11: And what I always do and that's again the method.

00:31:14: And I go level higher and say, well, what about our own models?

00:31:17: Exactly. What about our culture?

00:31:19: What about, you know, how do we see the world and how do we?

00:31:23: Many of us, I mean, we already don't agree.

00:31:25: And that's good. We don't agree.

00:31:27: I don't see main UNI, but we as a population, as a community.

00:31:31: We have different opinions.

00:31:33: But this morning, the news came that the European Commission wants

00:31:36: to provide 56 million euros for an open source model for Europe.

00:31:41: Let's see what comes out of it.

00:31:43: I find it strange that we can't imagine that Chinese researchers

00:31:46: can also build a good model.

00:31:48: And yes, I'm worried about what we are doing in Europe right now.

00:31:51: Yeah, absolutely, Peter.

00:31:53: Yeah, but at the bottom, there is this I've seen many, many positive signs.

00:31:56: You know, if they can they can do it now with, you know, there is no reason.

00:32:00: Well, we I mean, because we have not had, we don't have the hyper scalars.

00:32:06: We don't have what we do.

00:32:09: I mean, you know, SAP, Annex AI, own architecture

00:32:13: and we only hear positive things from that perspective.

00:32:15: And there's a couple of other things and talk to Frank Hutter this afternoon.

00:32:19: Frank Hutter, yeah.

00:32:20: So there's a couple of things that we see happening.

00:32:23: And I did I did feel that, you know, sure, we can do it as well.

00:32:28: And Gunther says the same thing, you know, maybe even without having

00:32:33: the 100 million budget with the smaller budgets, you seem to be doing

00:32:38: be able to do a lot of things.

00:32:39: And you will talk to André tomorrow, I think, about World Robotics Models.

00:32:43: Do today as well.

00:32:44: Actually, well, if I may move to the next topic, is that it's OK with this

00:32:49: because Jan, Jan LeKan, did we ask him to come and speak at our podcast.

00:32:54: So he was close to us while he's French based.

00:32:57: So whenever Davos is, I assume, although I assume he is based somewhere

00:33:02: in the US with Meta, he talks about the new paradigm, right?

00:33:07: Of AI, architectures.

00:33:09: No, he has been doing that for a long time, right?

00:33:11: But still just bring that point.

00:33:13: And maybe we can talk to him and he can explain us.

00:33:17: He says the shelf life of the current LLM paradigm, fairly short, three, five years.

00:33:22: He says within five years, nobody in their right mind would use them anymore.

00:33:27: And now we're going to see different architectures.

00:33:29: Not a point I want to come to, which is important, is that, you know,

00:33:33: he says the limitations to come to truly intelligent behavior are a lack of,

00:33:40: number one, understanding the physical world.

00:33:43: And that's the point.

00:33:43: And number two is then persistent memory, what is a more reasoning

00:33:48: and planning goes for you.

00:33:49: And that is the point.

00:33:51: So the interview that I started with Andri Dayanchuk.

00:33:57: There you go.

00:33:58: That's his difficult name to pronounce.

00:34:00: That I started in Frankfurt during the conference.

00:34:04: I'll restart again.

00:34:05: And it's on the topic of robotic world models.

00:34:09: So I'll do that today.

00:34:10: And you can hear it in a couple of weeks, I believe.

00:34:13: Perfect.

00:34:15: What else do you have?

00:34:16: Or let's move to the main part.

00:34:17: Yeah, I do.

00:34:19: No, no, I do have age.

00:34:21: Oh, yeah, I'm still so convinced.

00:34:26: So so deep seek hat their two weeks of fame.

00:34:31: And, you know, I think they deserved it.

00:34:33: It's rather, it's rather impressive that you can have the number one market

00:34:39: capitalization company, you know, reduce 600 billion.

00:34:47: So they deserve their two weeks.

00:34:51: And I think they're going to play a very just by doing what they did.

00:34:56: I'm not sure we're gonna.

00:34:57: Yeah, we're probably going to see them again.

00:34:59: But more importantly, we're going to see many, many, many other companies.

00:35:02: That's the thing.

00:35:03: Now agents are going to have a lot lasting, compact impact such again.

00:35:10: Confirmed what we talked about.

00:35:12: He spoke to the Indian Software Developer Community.

00:35:17: Agents are going to replace all software.

00:35:19: And now very important quote unquote for Tom Kadeira, if you're listening,

00:35:24: but also all the other, the complete user interaction designers.

00:35:30: He says there's no point in hard coding a UI anymore.

00:35:33: That's amazing.

00:35:35: That's not sad that we don't need the UI.

00:35:38: That's not what I'm saying.

00:35:39: Was he is saying?

00:35:40: It is.

00:35:41: It's not the hard coding it because saying the agent will do that on the fly.

00:35:47: Now, I'm going to give an example.

00:35:49: I'm going to combine it with Reina Brehm, CEO of the Factor Automation

00:35:54: with Siemens, you did an interview with him about half a year ago, I think.

00:35:57: And he announced for the S7 1200 G2, that is called their NFC app,

00:36:04: their Near Field Communication app.

00:36:06: Exactly.

00:36:07: And now back to back to basics.

00:36:10: It's like, you know, this is going to allow you, who has an iPhone,

00:36:15: to manage your Symatic controller directly from your device.

00:36:20: And the same is going to be possible.

00:36:23: I think this data is going to be an Android version by the end of the year.

00:36:27: Now, I have brought the question of BYOD.

00:36:33: You remember, bring your own device.

00:36:35: Yeah, own device.

00:36:37: A couple of times in the last podcast.

00:36:40: And here it is.

00:36:42: Siemens just developed it.

00:36:44: And that is so good, because this potential revival is confirming my personal

00:36:49: belief, a couple of times that our smartphone is going to be the center

00:36:53: piece of whatever's going to be still.

00:36:57: We call it agents. The consumer doesn't care about agents. I am

00:37:01: convinced. Yeah, I'm so convinced they're going to call it

00:37:05: there. We say in Germany, our handy other people call it their

00:37:09: smartphone, my phone, and that they're going to have things

00:37:13: they're going to talk to it. They're going to have their

00:37:16: their preferred personal preferred. So one of them is

00:37:19: going to use in deep seek the others going to be using

00:37:21: German, whatever doesn't matter. And that is going to be

00:37:25: their interface, you know, now, and what does that mean? You

00:37:29: know, if Siemens is making this available, obviously, they

00:37:33: assume that they allow you as a consumer to come to work. And

00:37:40: you know, you're coming close to the factory one way or the

00:37:43: other. The factory, the autonomous factory will

00:37:46: recognize, you know, Robert is there, and we'll say, oh, by the

00:37:48: way, Robert in the former shift, isn't this happened? And you

00:37:52: say, okay, yeah, I'll wait, you know, I still have five minutes

00:37:54: to drink my coffee and then I start whatever. Now, if they are

00:37:57: going to allow you the person working in the factory to come

00:38:02: in as the consumer and then, you know, eight o'clock in the

00:38:05: morning, your time, your shift starts, if that is possible,

00:38:09: you know, I think it's a it's a complete confirmation of what

00:38:12: does that mean for the HMI in the in the business in this case,

00:38:18: in the factory, absolutely, if they're gonna allow you to use

00:38:22: your in this case, iPhone HMI, you look at your screen, right?

00:38:26: So the screen of the of the PLC becomes non existent.

00:38:30: Exactly.

00:38:31: Quote unquote, quote unquote. So Peter, those were your agents.

00:38:36: Let's move to the main part. You know, I was at the Beyond

00:38:39: Safety Forum at Zick in December. And I was allowed to

00:38:43: record a panel discussion there. I was joined by Anders

00:38:46: Billy Söbek, responsible for R&D at Universal Robot Robotics

00:38:50: Company, Daniel Leidner, represented DLR, and Kasten

00:38:55: Roscher, spoke on behalf of Fraunhofer. And the topics were

00:39:00: robotics, AI and safety for sure, because we were at the Beyond

00:39:05: Safety Forum at Zick. Very interesting discussion about the

00:39:10: future of robot safety and AI. It's a little bit difficult to

00:39:14: follow, I think the episode, because there are free guests

00:39:17: that's absolutely maximum, I think. But yet it will work. I

00:39:21: think so.

00:39:22: Main topic then, safety in times of probability based

00:39:27: artificial intelligence. Exactly. Okay, yeah, I think the

00:39:30: first time we talked about a nutshell that they were the first

00:39:33: one was this PLEAN lab. Yeah, right. They had they seemed to

00:39:36: have a what is it? Not a copyright and IP there, right?

00:39:40: Yeah, exactly. A patent. Yeah.

00:39:42: Okay. Yeah, looking patent. Yeah, looking forward to, yeah,

00:39:46: because that's, it has always been kind of difficult for me to

00:39:49: understand how that for me to

00:39:51: work. But yeah, in the end, it's, it's whatever security

00:39:59: safety. I think we as a world have agreed that we allow

00:40:04: specific organizations in Germany in Europe, we have the

00:40:07: TÜV, there's many other ones, the technical whatever company on

00:40:14: whatever looking at things are going good or not. And behind

00:40:17: those, there's always insurance. So I think it's if if

00:40:21: something if a if a peaceful heart or software gets a stamp

00:40:24: from these organizations, then we can be not 100%. You know, 100

00:40:29: percent doesn't exist, but we can be 99.99 something 999

00:40:33: sure that things work. And as long as we find a technical

00:40:38: solution, in this case, and a safety by AI, and we can get to

00:40:42: it is whatever high number, it will then then we we can then

00:40:47: believe the specialist that it will work right and if in one

00:40:50: case out of a whatever a billion it doesn't work, then there's

00:40:54: the insurers behind it to make sure exactly Peter. Okay, well

00:40:59: looking forward to exactly thank you very much. Let's move to

00:41:02: the main part. It was a pleasure. I wish you all the best

00:41:04: with recording with Frank and with Andre. I will look forward

00:41:07: to this episodes. And yeah, let's move to the main part.

00:41:12: Talk to you soon. Have a great day to use well wherever you

00:41:15: are. Bye bye. Bye bye.

00:41:16: Normally, a podcast starts like hell everybody and welcome to

00:41:21: new episode of our industrial AI podcast. My name is Robert

00:41:24: Weber and it's a pleasure to talk to and I talk to normally to

00:41:27: Peter Sieberg. That's my co host of the industrial AI podcast.

00:41:31: Peter is sick in the moment. He is not able to come here today.

00:41:34: But no matter we are very proud to be here at the beyond safety

00:41:39: forum at sick and I invited three guests. Anders, Billy, so

00:41:43: we're back from Universal Robots. Welcome. Thanks a lot.

00:41:46: Thank you. Then Daniel Leitner from DLR. Hello, welcome.

00:41:50: Hello. And cast Russia from from over IKS. Hello. Hello.

00:41:54: Welcome to the podcast. I must admit that I'm not a safety

00:41:58: specialist. So we should talk about AI today. But I'm reassured

00:42:03: that Jenny, I was also a topic at the end of this conference

00:42:07: because normally when I go to AI conferences, Jenny, I is the

00:42:12: most hyped topic on the agenda. Everybody's talking about

00:42:15: gen AI. Isn't the safety community interested in Jenny? I

00:42:21: custom. Yes and no, I would say on the one hand side, I think

00:42:25: Jenny I for safety critical applications is still a topic for

00:42:29: the very, very distant future. At least that was it feels like on

00:42:33: the other hand, we can use Jenny as a supportive tool already

00:42:36: in development and engineering processes. How do you do that?

00:42:39: As a tool, we think of it as a sort of companion, which is an

00:42:43: overstressed word, I think in this domain as well. But you

00:42:47: could think of it as maybe your standard co pilot tool that

00:42:50: will assist you in the not so fun tasks of programming, but for

00:42:55: safety engineering, like generating documentation, giving

00:42:58: you support and analyzing complex architectures, finding

00:43:01: failure modes, all those kind of things. When we talk about

00:43:03: gen AI, we need to talk about agentic automation. So that's the

00:43:07: next big hype, I would call it because everybody now talks about

00:43:10: agents and agentic automation and stuff like that. Is there a

00:43:15: safety related topic to that, Daniel? To agentic automation?

00:43:21: Well, agentic automation, I would say that yes, the robots that

00:43:27: automated tasks act already like an agent, right? They have the

00:43:31: knowledge of the domain, they have the knowledge, hopefully

00:43:34: about themselves. And if they don't have it, they can actually

00:43:37: access it through what Kaston just explained, maybe large

00:43:41: language model or region language model to get the data that

00:43:45: they need to access in order to solve the task.

00:43:47: But when you talk about, for example, we had an episode with

00:43:50: the guys from Siemens, they talk about co pilot operations,

00:43:53: stuff like that. And there's an agent in the back. And he will

00:43:57: make decisions on topics how to run the machine at the end.

00:44:01: Isn't that a safety rather than topic?

00:44:03: Yes, yes, I think so. So the robot needs to be able to assess

00:44:08: what could go possibly wrong, right? So it needs to know how to

00:44:13: recover if something goes wrong, because in the future, the

00:44:15: robots will no longer do only one task, but they will do

00:44:18: multiple tasks, they will be mobile, they will support workers

00:44:22: in different situations. And so therefore, actually, they will

00:44:26: leverage the information that is coming from the behind from the

00:44:29: agent and adapt themselves right to the correct situation.

00:44:33: And what does it mean now for safety?

00:44:35: Yeah, for safety, it means that, for example, a robot may be able

00:44:39: to avoid damage to itself based on information that is given to

00:44:43: through the agent may be able to avoid damage to the environment

00:44:46: and most importantly, to the human.

00:44:47: And how do you handle situations like that when it comes to

00:44:52: combining AI and safety on us?

00:44:56: It's a really good question. I mean, today we see AI being used

00:45:00: increasingly in many different industrial applications,

00:45:03: especially where vision technology is used today.

00:45:05: But that's common, right?

00:45:07: It's very common. I think even if you look at the Universal

00:45:10: Robots customers that we have around one third of what we call

00:45:13: the solution revenue we have is coming from solutions that are

00:45:16: that are powered by AI today. So there's no doubt that this is

00:45:20: moving really fast into the industry. Then the question is,

00:45:22: how do you work with safety around that? Today, it's all about

00:45:26: thinking, how do you broadly, because the opportunities that

00:45:28: AI gives is much larger degree of freedom in applications.

00:45:32: It's a much wider variety of tasks to be solved.

00:45:36: And of course, they have to be safeguarded. It has to be safety

00:45:40: around the robot, as we talked about, it has to be safety,

00:45:42: especially around people around it. And that has to be

00:45:46: considered in a much broader context. So I think with all the

00:45:49: advancements that are happening also in safe sensing and so on

00:45:52: will help with that. But there's no doubt that for now, there's

00:45:55: still a gap in between what is done on the functional level

00:45:58: using AI and how we apply modern techniques of safety to then

00:46:02: safeguard humans and equipment around it.

00:46:04: Kastan, you mentioned there's a gap between both worlds, the

00:46:08: machine learning AI world and the safety world. How big is this

00:46:12: gap and how to bridge?

00:46:14: It depends, I think, on the on the machine learning technology

00:46:16: that you use, because I mean machine learning is a broad field.

00:46:19: So if you have a small model that you learn from a few data

00:46:21: points with a simple task, then the gap is not that large

00:46:25: because you could still inspect it theoretically. On the other

00:46:28: hand, you have like a large foundation model trained by

00:46:30: someone else and you use it in such a context. It's a different

00:46:33: story. Why is a lot of the principles that you would usually

00:46:36: do in safety engineering assumptions on the behavior on the

00:46:39: failure modes, experience that you have in, for example, parts

00:46:42: failing or different software pieces, how they interact are

00:46:45: simply missing. And those solutions surface so fast that we

00:46:49: haven't had time to collect the experience that we need in order

00:46:53: to derive a safety argumentation from there right now.

00:46:56: And this gives up, I would say, an uneasy feeling. I said earlier

00:46:59: in the presentation, it might be that we are already safe, we just

00:47:04: don't know it.

00:47:04: Okay, can you explain a little bit?

00:47:07: I mean, we see the performance of those models. And I would argue

00:47:10: in some of the cases, we don't get the 10 to the power of minus

00:47:13: seven error rates. I mean, 99% of machine learning world is

00:47:17: already very high performance. So there's definitely this gap.

00:47:21: But on the other hand, if we build a system around it, dealing

00:47:24: with some of the issues that we have, and that we know, for

00:47:27: example, computer vision algorithms, also machine

00:47:29: learning based not working very well in dark conditions, we can

00:47:32: put a light sensor in place and just filter that out. And with

00:47:36: that, we might actually end up with a system that is already at

00:47:39: least safe for or safe as safe as the ones that we develop

00:47:43: traditionally. But we simply have no way of demonstrating this

00:47:46: because we're missing the amount of test cases, amount of test

00:47:49: data to do, for example, statistically valid testing in

00:47:53: the end.

00:47:53: What is your opinion, Daniel, when it comes to the gap between

00:47:56: both worlds? Or is it one world or separated worlds?

00:48:00: One world. But I think that there's a saying that says there's no

00:48:03: free lunch, right? So you cannot just have open source phrase.

00:48:07: Basically, for large language models and safety, it says that

00:48:12: if you apply large language model vision language model to a

00:48:16: safety critical aspect, you cannot just take that for

00:48:19: granted, right? You need to check it. And you can also not check

00:48:22: it again with a machine learning based system, because then you

00:48:25: end up in a vicious circle, testing something that is

00:48:29: unreliable with something that is unreliable. So therefore, there

00:48:31: you need to go through the other world to the other side and use

00:48:34: some classical methods. There's also classical AI, right?

00:48:37: Reasoning based methods, knowledge based methods that can

00:48:41: hopefully then identify errors in the models and avoid misbehavior.

00:48:47: Is there a communication gap between both worlds? Because every

00:48:51: every month, every week, I see a new robotics foundation model.

00:48:56: Every university now builds the new one foundation model on

00:48:59: robotics Toyota and in USA, a lot of companies are working on

00:49:03: this in Amsterdam. And you are also working maybe on a

00:49:06: foundation model. But do these people talk to the safety people?

00:49:10: I don't think so.

00:49:14: So and what's even worse, I think not even the machine

00:49:18: learning people talk to the knowledge based reasoning people.

00:49:21: There is some efforts that try to rediscover this communication

00:49:25: channels, but actually missing. So from cognitive science, we

00:49:29: know that a system should have both worlds embedded into the

00:49:32: reasoning environment. So there should be a fast system, right?

00:49:36: For recalling information that it was that it learned through,

00:49:39: let's say a large language model. And there should be a slow

00:49:43: system that identifies errors and fixes them. And then there

00:49:47: needs to be something in the middle, which mediates between

00:49:50: them. And this is what humans call metacognition. But we are not

00:49:53: far away from that, actually.

00:49:54: Your safety developers, do they talk to your machine learning

00:49:58: developers?

00:49:59: Yeah, so the communication is very tight. And I think both

00:50:03: from the development community that are also in the in the large

00:50:07: worlds of you talked about the university springing in new

00:50:10: What is your opinion on these models all around the world?

00:50:12: I think it's amazing, right? It unlocks so many new things that

00:50:15: we can do with robots. It solves some of those fundamental

00:50:18: gaps that robots have been struggling with forever, which

00:50:21: is handling the variety that the real world faces. So there's

00:50:26: no doubt there's some like unique classes of applications that

00:50:29: really, really benefits even today from something like

00:50:32: logistics, where everything you handle is different. Like every

00:50:36: parcel that flows through logistics is is basically

00:50:38: different. And foundational models are brilliant at that

00:50:42: accuracy requirements is modest, right? But but the right

00:50:46: stability capabilities are extreme. So these kind of

00:50:52: innovations makes automations within logistics grow exponentially

00:50:57: faster than any other business at the moment. So all of this is

00:51:00: great. And then and for there, we're having some quite close

00:51:03: dialogues with the adopters of these kind of systems and and

00:51:06: safety. I think the good thing in my mind is that it does not

00:51:12: have to be that closely related just yet. Because deploying

00:51:16: automating logistics handling processes using gen AI has the

00:51:20: exact same safety requirement as the other robot application.

00:51:23: Right, because of course, motions will be unpredictable, they

00:51:26: will depend on whatever the gen AI model produces. But in the end,

00:51:30: it's all about keeping environments and equipment and

00:51:33: especially people safe. And we have well existing

00:51:36: technologies for that. We have well existing technologies for

00:51:39: managing emotions, also keeping it in within safety envelopes of

00:51:45: the systems and so on. And we appreciate having even fully

00:51:48: dynamic applications in this way. So I think there's some good

00:51:51: dialogues going, I agree that having them fully safety rated

00:51:55: models that can predict how to pick up a parcel and do that in

00:51:59: a fully safety rated way where everything is predictable is

00:52:03: something else. But I also don't necessarily think we need to

00:52:06: slow down innovation and adoption of these models to solve

00:52:09: that problem just yet, because they're still safe in the way

00:52:12: we do them today.

00:52:13: I was very happy when I checked the agenda of this day,

00:52:15: because I saw your name. And I did a recording, I think two

00:52:19: weeks back with Dennis Stogl. He's a Ross, her us guy, and he

00:52:23: was an Odense at the Ross con. And then we discussed one topic

00:52:28: and I need to ask you this question, because I do not

00:52:31: understand what you are doing there, because you showed there

00:52:34: a cobot and you then put a GPU next to the cobot. Does that make

00:52:40: sense? Put a GPU next to a cobot?

00:52:43: It absolutely does.

00:52:44: Yes, why?

00:52:45: To basically stimulate the growth we just talked about, right?

00:52:47: There's a lot of opportunities that modern AI can unlock across a

00:52:52: lot of industries. And we're even seeing it today. I think

00:52:55: today, 89% of manufacturing

00:52:59: executive all about costs, right? GPU costs a lot. When I talk

00:53:02: to guys who sell robots, they say, Oh, in the moment, the

00:53:05: companies are buying Chinese robots because they are cheap.

00:53:08: They buy this for three years and after three years, they put it

00:53:11: on the trash and say, Oh, we will buy the new one. But that's a

00:53:14: cost topic, right? A GPU.

00:53:15: A GPU is a cost topic. There's no doubt about that. And with the

00:53:19: advent of modern modern GPUs is costly equipment. On the other

00:53:22: hand, I think what really matters is total cost of

00:53:25: deployment. Like earlier today, we talked about the key barriers

00:53:29: of adopting more automation is reducing overall deployment costs.

00:53:32: At the same time, we need to have more flexibility. We need to

00:53:35: have better reliability. We need more performance. And we need

00:53:38: things that are easier to use. Modern AI solves most of those

00:53:41: things. There's no doubt that some of the hardware might be more

00:53:44: expensive. But also we know that some of the key cost drivers of

00:53:47: deploying automation is labor is integration labor, it's

00:53:51: technical problems, it's programming times deployment

00:53:54: time, all of that with the advent of gen AI and so on can move

00:53:58: us so much more in the direction of deploying things faster,

00:54:02: having it run more reliable, delivering more end customer

00:54:05: value because of flexibility. So I don't think we should be

00:54:08: excess obsessed about the price tag of individual hardware

00:54:11: components. We should think about what would total deployment

00:54:14: costs actually be operational cost over lifetime and so on.

00:54:18: That's difficult in the moment to argue, right?

00:54:20: Well, I don't know. Because in the end, even for a manufacturing

00:54:23: customer, they do not see the bill of material to whatever they

00:54:26: buy. They see the end turnkey solution cost. And if that's

00:54:29: realized with a single robot and a GPU, or if it's realized with

00:54:33: a robot and 9,000 hours of manual labor and programming and

00:54:37: deployment time, well, for them, it actually doesn't matter, they

00:54:40: don't become happier customers if they pay a lot of sort of

00:54:43: bespoke engineering hours. So but of course, it's all about

00:54:46: value. It's all about it needs to provide the value to

00:54:49: customers. And at least we see it today that that some of the

00:54:52: things like logistics, automating logistics would be

00:54:56: entirely impossible at the scale that we're seeing it happening

00:54:58: at the moment without accelerated computing.

00:55:01: Then it doesn't make sense.

00:55:02: Yes, I think so. So I mean, consider that you want to have

00:55:06: your models trained in the cloud and they want to have an edge to

00:55:09: just recall what you learned, right? So therefore, I don't think

00:55:12: that you have just one GPU for one robot, but eventually we'll

00:55:16: have several, right? So it makes a lot of sense.

00:55:19: What is your opinion?

00:55:20: I mean, I tend to agree. It also depends on the size of the GPU.

00:55:23: So I mean, with like the integration still rising, those

00:55:26: things will be cheaper as well. So if you now have a prototype

00:55:29: with the GPU.

00:55:29: But from my point of view, the time of scaling, it comes at the

00:55:33: end, when it comes to models or large language models. What is

00:55:37: your opinion on that? Maybe we should focus on smaller models,

00:55:40: on more precise models on more domain specific models, which do

00:55:47: not need a GPU, maybe.

00:55:48: Yeah, yeah, I'm an advocate for that. So you should not have

00:55:51: always the need for an end to end model, right? I think that

00:55:56: skill based programming is still a valuable approach for industrial

00:56:00: manufacturing. So and one of the skills may very well be a model,

00:56:04: either a VLM, a whatever a foundation model, and doing a very

00:56:10: specific task. And then on the other hand, you may also have

00:56:13: good old fashioned engineering as another sequence, as another

00:56:16: skill.

00:56:17: Does AI help us or your community to improve safety or to to find

00:56:25: new approaches when it comes to safety?

00:56:28: That's a tough one.

00:56:29: Improve safety. I think it's still underused to that end. So I

00:56:35: think it could improve safety because it just helps us to keep

00:56:38: an overlook on all the complex things. I mean, if you ever seen

00:56:41: a GS entry for a person detection, which would cover I think the

00:56:44: entire wall here, then it's very different, difficult for us to

00:56:47: even understand and grasp all the details. So there it can

00:56:50: definitely help. I'm not sure if it's already used on a broad

00:56:53: level. Just yet, whether it can find new approaches, improve,

00:56:59: help develop, improve and help definitely develop. I mean, it

00:57:03: also it would need to surprise us because I'm in the end safety

00:57:07: is also about getting acceptance from someone like

00:57:10: demonstrating that the freedom of unnecessary risk. So yeah, maybe

00:57:14: a bit philosophical in the end. I'm not so sure.

00:57:17: And what's your opinion on that?

00:57:19: No, I think it does. There's no doubt that the community both on

00:57:22: the advancement that happens in AI sparks a lot of need for safety

00:57:26: innovation. I do believe they can be treated independently. You

00:57:29: need to also from the safety community think about what does

00:57:33: it take to embrace the advances in technology, all the sort of

00:57:37: flexibility requirements that comes from from AI. And so today

00:57:42: it's not about hardening the systems we already have. It's

00:57:45: all about how do we work with a structured way of risk and

00:57:49: solving the flexibility problems that AI systems inherently

00:57:54: will will impose to us. I think that's the challenge that that we

00:57:58: in the safety community need to accept and think creatively

00:58:01: around how to offer new generation solutions, how to how to

00:58:05: manage this, which is often maybe a deeper integration of safety

00:58:09: technologies. It's a combination of safety technologies. And and

00:58:13: it should certainly not slow down the pace of innovation that

00:58:16: happens in the AI communities.

00:58:17: Because I mean, after a short while thinking sometimes, I think

00:58:21: it could help and maybe even re establishing some things that

00:58:25: are already there because for a long time, I think safety,

00:58:28: especially research and safety engineering was quite ahead of

00:58:31: what is applied in the real world. And some of the methods never

00:58:35: really made it into practice because of different reasons like

00:58:38: interfacing of things and so forth. You already said dynamic

00:58:41: dynamic management of safety instances, for example, and we

00:58:45: might see a re yeah, a resurgence of things that are already

00:58:51: developed to some degree, but now being applied because there's

00:58:54: a necessity. So in that sense, it might actually help us to

00:58:57: unlock new capabilities in the safety engineering sense, because

00:59:01: we now finally have a way to apply them.

00:59:02: You want to add something, Daniel?

00:59:04: It's the same actually with the old fashioned AI, knowing that

00:59:08: we have flaws with the new generation of AI with generative

00:59:11: models, we will resort back to them at some point as well as a

00:59:13: companion to the new AI wave.

00:59:16: Today we heard something about safety pipelines, we need to

00:59:20: establish safety pipeline. I think the guy from Bosch spoke

00:59:23: about safety pipelines, that you always have to reassure that

00:59:27: just your product is still safe. We know that from a machine

00:59:30: learning pipeline, a model pipeline, DevOps pipeline, security

00:59:34: pipeline, is there a new market emerging to build infrastructure

00:59:38: for all these pipelines? And who builds this pipelines? Is it you

00:59:43: are who builds this pipeline to the customer? Who will do that?

00:59:46: That's a really, really good question. And there's, there's

00:59:49: no doubt that individual products need to maintain safety, and

00:59:52: both with the velocity of product development happening, then

00:59:56: finding even more structured ways to do safety qualification,

01:00:00: safety validation is a need. So speeding up time to market and

01:00:05: safety product development is by no means a thing that we have

01:00:09: well under control today, I think time to market in general and

01:00:12: safety products are probably too long to keep up with the with

01:00:15: the innovation that happens in other places.

01:00:17: But your customer will ask you.

01:00:18: Absolutely. So it's, it is a problem in the in the safety

01:00:21: technology community that needs to be addressed. Like how do we

01:00:24: bring safety rated product to market faster? How could we solve

01:00:27: some of these things and maybe reduce the reliance on more sort

01:00:31: of manual processes and deep technology reviews? And I think

01:00:34: that's what automated pipelines brings. If you think about

01:00:37: classical software engineering, automated pipelines brings

01:00:40: automated robustness, it brings sort of early detection fast. And

01:00:45: in these cases, using these more actively and also safety

01:00:48: product development would at least be an important both for

01:00:51: evolution of products to have a more software centric way of

01:00:55: thinking about safety products, because software products do

01:00:57: evolve software products needs continuous integration test.

01:01:00: And and hopefully it would accelerate time to market as well.

01:01:03: Maybe there are new competitors on the horizon. When it comes to

01:01:07: infrastructure, I know a lot. US companies were really looking

01:01:10: forward to build infrastructure in industrial

01:01:13: environments. So what is your opinion on safety pipelines? And

01:01:17: is there an emerging market for that?

01:01:19: Maybe not a market, I would say, but from from my research point

01:01:22: of view, there's a need for automated pipelines that react

01:01:25: really quick. So we need to reduce also the reactivity time.

01:01:28: Think about that space robotics. There is a robot maybe in a

01:01:33: space station orbiting Mars 40 minutes away from communication

01:01:37: to Earth. If something happens, there's no room for waiting for

01:01:41: human to give some advice, right? So the system needs to be in

01:01:44: the pipeline react autonomously and make decisions itself.

01:01:48: What is your opinion, Kasten? Are you sure that the robotics

01:01:51: companies can build these safety pipelines?

01:01:53: I see no reason why they shouldn't be able to. I think the

01:01:57: main the main challenge right now is that a lot of the assurance

01:02:00: processes, as as already said, are manual labor. And it's very

01:02:03: difficult to integrate them into a sort of like DevOps like

01:02:07: pipeline. I mean, the the processes then just as good as

01:02:10: the test cases that you have. And I think there's an ongoing

01:02:13: debate in a lot of standardization groups, what those

01:02:17: test cases limits and so forth should be. And until then, also

01:02:20: maybe with the idea of getting quicker to market, we have to

01:02:24: rely on, I would say, a good sense of human intuition, what

01:02:28: should be safe and not and arguing also not just

01:02:31: quantitative in terms of like test case performance, but also

01:02:34: qualitative properties. And that is difficult to automate.

01:02:37: Eventually, we will get there, but I don't see this happening in

01:02:41: the next five years.

01:02:41: So you need to invest in this topic.

01:02:43: Yeah, maybe very well. So yeah.

01:02:45: So it's a it's a platform strategy, right? So everybody wants

01:02:49: to build a platform. Maybe that's the key to get everybody on

01:02:52: the platform.

01:02:52: Absolutely. I think the key about key value drivers on

01:02:56: platform is to be solving real problems.

01:02:58: And that's a huge problem.

01:02:59: It is a huge problem. So yeah,

01:03:01: you mentioned smaller models, you mentioned transformer

01:03:05: architecture. I'm a bit biased because I work for several for

01:03:09: NX AI and we talk about XLSDM and stuff like that. And do we

01:03:13: need new AI architectures?

01:03:15: I don't think they would hurt. Variety is good, I think.

01:03:19: And the question is whether they would, even though they are

01:03:21: new, what kind of new qualities would they bring to the table?

01:03:24: So is there anything fundamentally better or worse?

01:03:27: I mean, in the story that I told about the transformer and then

01:03:30: a lot of the other things, the takeaway for vision at least was

01:03:33: that it's not necessarily the architecture, but some some training

01:03:37: tricks or some some pre-processing stuff that you apply.

01:03:39: And I think this is difficult to disentangle, but I definitely

01:03:43: think that there are a lot of problems that we are not very

01:03:46: good at modeling right now.

01:03:48: And one thing that I would like to see, I hinted at this earlier,

01:03:52: is not post-doc explainability, but some sort of interpretability

01:03:56: by design. But as soon as you open that box, it's a lot of

01:04:00: engineering again. And I've already been told by other people,

01:04:04: this is what like what we did 20 years ago, and that's not

01:04:07: modern anymore. We train end to end. And it's a bit against the

01:04:11: stream, but I'm completely on your side when it comes to maybe

01:04:14: smaller models.

01:04:15: Daniel.

01:04:16: Yes, I think that there is a need for different kind of AI models

01:04:20: for actually cognitive architectures. So also, this should

01:04:23: have a revival, because if you think about humanoid robots,

01:04:26: right, they should react as you expect a human to react on a

01:04:29: certain issue. Just repeating what the robot did, if there's an

01:04:33: error, may actually not help, right? If you have an end to end

01:04:36: model, and this is the only policy that was learned, try again,

01:04:39: it will not fulfill the task.

01:04:41: Yeah, at least I very much agree there. I think a key criteria

01:04:45: for industrial applications is that it has some level of

01:04:48: optimized ability that you can also investigate and see at least

01:04:52: grasp some amount of understanding of the fundamentals of the

01:04:56: process. So I think for a successful sort of scalable

01:04:58: adoption in industry, at least it's important to think about

01:05:02: hierarchical models, at least models that represent different

01:05:06: elements, which you can go in, you can enter into respect. So I

01:05:10: do believe like end to end learning will have a difficult

01:05:14: terms in industrial applications, where there will be a need of

01:05:17: understanding to refining parameters to refine these elements

01:05:21: of the process to optimize if you're not satisfied with the

01:05:24: results, what do you do? That's that's a fundamental question

01:05:27: you need to understand and retraining and hoping that you

01:05:31: get something better, it may not be the answer. So smaller

01:05:33: models for focus on sort of fundamental building block

01:05:37: models and composing them, which is very similar to how

01:05:40: humans do as well.

01:05:42: Do we will see a revival of reinforcement learning in the

01:05:45: next years? Because in the moment, nobody, not nobody, but only

01:05:48: a few people are talking about reinforcement learning.

01:05:52: I think you will have a combination of reinforcement

01:05:54: learning and foundation models. Imagine that a robot needs to

01:05:59: traverse a certain area that it is not able to even know if it

01:06:02: can traverse it. So it needs to try out first. We don't want it

01:06:05: to try out in reality, because it will topple over, right? So

01:06:08: therefore, it learns in simulation through reinforcement

01:06:11: learning, generates a new policy that is then fed into the

01:06:14: foundation model. And now the foundation model is smaller than

01:06:16: before.

01:06:17: What's your opinion?

01:06:18: I mean, I think it even depends on where you apply reinforcement

01:06:21: learning. I when you mentioned, there's currently maybe not so

01:06:24: much of a hype. If I remember correctly, a lot of the fine

01:06:27: tuning of LLMs is based on reinforcement learning with human

01:06:29: feedback. So I think that reinforcement learning has

01:06:32: let's talk traditional reinforcement learning.

01:06:34: You mean for control tasks?

01:06:35: Yeah, control tasks. Yeah.

01:06:37: Yeah, I mean, we kind of disentangled a bit the sort of

01:06:41: training and the deployment. So you can train a decision action

01:06:45: decision model basically with reinforcement learning, but you

01:06:48: don't retrain it in practice. And I think this is just learning an

01:06:51: optimized policy basically that you deploy. This is also I think

01:06:54: fine. Reinforcement learning that is still learning in the

01:06:59: field. I have to look at the people who would need to work

01:07:04: with that. And I usually get an uneasy look back. So that is

01:07:07: why doubtful.

01:07:08: Let's talk about your safety community, I would call it. Do

01:07:13: you need new skills in this safety community? Which skills?

01:07:17: Communication skills, potentially.

01:07:19: I was surprised how many people came up here and asked

01:07:23: questions. So communication is not in terms of asking questions.

01:07:27: But I mean, I still think that there is this natural divide

01:07:31: between machine learning people and safety people. And the

01:07:35: unfortunate thing is, I mean, both of them have to learn

01:07:37: communication skills, by the way. So this doesn't only apply to

01:07:40: but how to learn how to talk to each other because sometimes they

01:07:43: use the same words, but with a very different meaning. And then

01:07:46: end up in a discussion about the meaning of those words and not

01:07:50: what they really want to talk about. So I think, for both sides,

01:07:53: it would make sense to understand a bit better what the other

01:07:57: side is about. So what machine learning means and what the

01:08:00: impact on the safety argumentation is in terms of, instead of

01:08:03: just saying, I need 99.999% and please also demonstrate how to

01:08:08: do it. Deliver. On the other hand side, I also think that

01:08:12: getting a bit more into the safety mindset and not just

01:08:16: thinking about optimizing a large model also helps machine

01:08:20: learning engineers to actually develop better solutions, because

01:08:23: we don't care about the model, we care about the function in the

01:08:25: end.

01:08:25: What's your opinion, new skills?

01:08:27: I do believe collaboration skills is a good term. And I think

01:08:31: you presented it very well earlier today, thinking about we

01:08:34: have two communities. And we have very modest crosstalk across

01:08:38: those communities. But essentially, we should be solving

01:08:41: the same problem. We should all be striving for getting more

01:08:44: robotics, getting more automation, getting more modern

01:08:46: technologies helped to deploy it to help drive society where it

01:08:50: needs to go. It needs innovation thinking on safety. It

01:08:54: needs innovation thinking on AI. And we're all sort of in that

01:08:57: same boat, right? We're not in two different boats sailing in two

01:09:00: different directions. So somehow embracing that problem and

01:09:05: challenge across communities and actually trying to dive in and

01:09:09: sort of combine and compose skill sets, rather than sort of

01:09:14: pulling in two different directions, which essentially

01:09:16: leaves us at the exact same point. I think that's the challenge

01:09:20: that our community needs, that we actually jump in. And then it's

01:09:24: a strong point we have. It's a strong point we have in Europe,

01:09:26: where we have deep understanding on safety

01:09:28: technology, we have deep sort of application, automation,

01:09:31: application domain expertise. So we also need to somehow join

01:09:35: forces there and drive that forward together.

01:09:37: New skills?

01:09:38: New skills are good, but we should not forget about the old

01:09:40: skills. So with chat GPT, people do no longer learn to

01:09:45: program. People do no longer learn to solve inverse kinematics,

01:09:50: forward skinematics, right? The one on one of robotics, this

01:09:53: needs to be in the curriculum of any study, I think for robotics

01:09:56: and for safety, I think it's the same, right? There are skills

01:09:59: that you just need and you do not learn them if you just type

01:10:01: a prompt into an API.

01:10:03: At the end, I want all of you ask what surprised you today

01:10:07: most? Which quote and why or which fact or which data or which

01:10:14: figure? What surprised you most?

01:10:16: So under said that there's only 2% of all the tasks automated

01:10:21: today. This surprised me a lot.

01:10:23: That was your presentation, right? Anders?

01:10:25: Carsten?

01:10:26: Yeah, same here, I would say. And then follow up. It was more of

01:10:31: a question related topic. The whole security does not work by

01:10:35: obscurity. But safety unfortunately still has maybe a

01:10:39: sharing challenge ahead, and that we may have to rethink how we

01:10:43: deal with safety and what can we share and what should we

01:10:45: collaborate on?

01:10:46: Anders?

01:10:47: Yeah, I think for me it was a broad conclusion. I had two main

01:10:51: observations. I think one of them is I was surprised in the

01:10:54: presentation we had from Bosch on how dominant certifications and

01:10:58: legislation is in the autonomous driving or in across

01:11:01: automotive in general, because I do fear that would be an obstacle

01:11:05: for innovation. The other thing I realized is a bit of this

01:11:08: divide that we've been talking about that there's there's only a

01:11:12: modest amount of discussion around how to solve big and

01:11:15: important societal problems when we are discussing adoption and

01:11:19: AI divide in between. It's it's doesn't seem to be the main

01:11:23: conversational driver, which is in my mind a bit of a shame.

01:11:26: I was surprised by the figure that China is pushing safety and

01:11:32: regulation so enormous push from from China and Asia. What is your

01:11:38: opinion on that at the end, Daniel?

01:11:39: I was surprised and I actually don't think it is what we see

01:11:44: there because there are so many products, especially robotics

01:11:47: products used in production in China that I think actually have

01:11:52: a lower certification rate than what we are used to in Europe,

01:11:55: right? So I don't know how these two things work together.

01:11:59: Custom, you have any idea?

01:12:01: Maybe they they are trying to do the inverse EU AI act and kind

01:12:08: of like motivate local suppliers to provide better quality, at

01:12:13: least as a long term goal, I'm not sure. But this might be one

01:12:16: direction. So that's not necessarily about their own

01:12:18: market, but to kind of like take away the potential leading edge

01:12:22: of other markets.

01:12:23: And then you sell robots in China.

01:12:25: Oh, yeah, difficult.

01:12:27: I think I have a colleague that's set it well, the Olympics

01:12:29: and robotics at the moment is played out in China. So if you

01:12:32: need to learn how to win, that's where the game is. It's where

01:12:34: half of the world robotic market is today. And there's no doubt

01:12:37: it's a tough market with a lot of different driving factors

01:12:41: about the safety regulation topics.

01:12:42: I think similar to what we talked about in the domain of

01:12:45: robotics, it has been, let's call it a little bit more loose

01:12:48: and now slowly starting to move towards something that we see

01:12:53: be being more and more familiar to Western standards in terms of

01:12:56: safety. It's interesting that it seems to be the opposite

01:12:59: driving factors on inside a ton of automotive. I'm not sure

01:13:04: what the key driving factors is, if it's actually a way to try

01:13:07: to consolidate the industry and raise the sort of the quality

01:13:09: standards, improve exportability, or what the key the key

01:13:13: drivers is, but it's an interesting trend.

01:13:16: Thank you very much, Carsten. It was a pleasure to talk to you

01:13:18: about safety and AI. Thanks a lot.

01:13:20: Thank you as well. My pleasure.

01:13:22: Daniel, thank you very much for your discussion on AI

01:13:25: foundation models and safety. And thank you very much for

01:13:28: explaining me GPUs on co-words. And as it was a pleasure,

01:13:32: you're very welcome. Thank you so much.

01:13:33: Thank you.

01:13:40: Robotik in the industry, the podcast with Helmut Schmidt and

01:13:44: Robert Weber.

01:13:45: (upbeat music)

01:13:47: (soft music)