Rise of AI Virtual Chat 2021 · Berlin

AI in Manufacturing: Teaching Robots to Learn by Demonstration

Ronnie Vuine · CEO, micropsi industries

At a Rise of AI Virtual Chat in February 2021, held in partnership with AWS, Fabian Westerheide interviews Ronnie Vuine, CEO of micropsi industries, about applying AI to manufacturing. Vuine explains how his company's software lets industrial robots learn skills, such as finding a logo on a screen in a live demo or locating features on refrigerators during factory testing, from four or five human demonstrations instead of precise pre-programmed coordinates: the robot watches a worker perform the task, uploads the recordings to AWS for training, and can execute the learned skill after about three hours of cloud computation. He traces micropsi's roots to a Minecraft bot project and to Dietrich Doerner's psi theory of intelligence, and describes why the company deliberately moved away from pursuing AGI toward a sellable product. The conversation then turns to why AI adoption in manufacturing remains low, with Westerheide citing a figure that only six percent of manufacturing businesses use AI, and Vuine arguing that bolting sensors onto old machines is not the same as redesigning production as a software-driven system. He closes with advice on building AI capability in-house, a note on data privacy between customers, and predictions about GPT-3's near-term impact.

Key takeaways

  • Micropsi industries trains industrial robots to perform tasks from four or five human demonstrations rather than precise pre-programmed coordinates.
  • Ronnie Vuine says pursuing AGI directly is commercially unpredictable, calling it "alchemy" rather than engineering or science.
  • The company's skill-learning system runs on an Nvidia GPU compute box and uses AWS cloud training, evaluating its camera feed twenty times per second.
  • Only six percent of manufacturing businesses currently use AI, according to a figure cited by Fabian Westerheide.
  • Vuine argues most factories only bolt sensors onto old machines instead of redesigning production as a software-driven system, unlike Tesla's approach.

Detailed notes

From a Minecraft bot to robots that learn spatial movement

Vuine says micropsi industries grew out of an earlier project to build an autonomous Minecraft bot that learned its own motivations rather than being told what to do. The team realized they had taken a shortcut by treating the game world as discrete, letting the bot "teleport" between known locations instead of learning continuous movement from one point to another. Solving that problem of moving from A to B without precise measurement, he says, is what pulled the company into robotics. Vuine argues there is a deep, not fully understood connection between this kind of spatial reasoning and intelligence more broadly, since animals and people evolved to handle continuous three-dimensional movement. He traces the company's name and approach to Dietrich Doerner's psi theory, a theoretical-psychology model of how intelligence, emotion and failure work that Doerner developed independently of AI research; Vuine and co-founder Joscha Bach read Doerner's work in a reading circle at Humboldt University and began implementing early versions of it in Java.

Why micropsi chose product over AGI

Vuine says he avoids discussing AGI with investors because pursuing it is "not engineering and not science, it's more like alchemy": progress cannot be planned or budgeted in a systematic way. Running a venture-backed company means working toward a roughly ten-year exit horizon, which he breaks down as three years to product-market fit and four years to revenue. After about two years of early, more exploratory work at micropsi, it became clear that AGI-adjacent research, such as artificial emotion, was not predictable enough to build a business on, since a moody robot has no commercial value even if it made the system more human-like.

How the skill-learning system works

In a live demo, an off-the-shelf industrial robot fitted with a standard, low-resolution camera (VGA-level, comparable to a twenty-five dollar webcam) learns to find a logo on a screen after a worker demonstrates the movement four or five times over about twenty minutes; the recorded demonstrations are uploaded to AWS, where the skill is computed in roughly three hours. The compute box driving the robot is described as essentially a gaming PC with an Nvidia GPU running Linux, evaluating the camera feed twenty times per second and making a continuous, self-correcting approach rather than one precise measured movement, the way humans plug in a USB cable by feel rather than exact geometry. A live customer example is testing refrigerators on a production line, where the robot must find features that always look similar but are never in exactly the same place. Vuine says customer data is never shared or pooled across clients; other customers benefit only because the larger pool of data lets micropsi backtest algorithm changes for robustness and speed.

Why manufacturing AI adoption is still low

Westerheide cites a statistic that only six percent of manufacturing businesses currently use AI. Vuine attributes the gap to the difficulty of connecting a "fluffy" software world to machines that are sometimes forty years old and built to do one physical thing well. He says the common approach of sticking sensors onto legacy machines to gather vibration or temperature data for predictive maintenance is a partial fix, not real digitization. He contrasts Tesla, which he says built its cars and factories as digital systems from the ground up, with German carmakers that he says have roughly 250 individually "smart" modules but no unifying operating system connecting them. His vision of a fully software-driven "general factory" points to additive manufacturing and 3D printing as the model, and he notes that continuous, fluid production such as beer brewing is already closer to full automation than discrete assembly work such as car manufacturing, which still requires many workers to move and fit parts together.

Advice for manufacturers: build AI capability in-house

Vuine recommends that manufacturers build AI capability internally rather than outsourcing digitalization entirely to consultancies, arguing that companies need enough in-house understanding to select and deploy AI products without losing the ability to update their own production later. He compares this to the shift from mainframe computing to PCs: businesses did not need MIT-trained engineers to run Microsoft Excel, just an IT person who was "good enough." On ownership, he confirms that a customer such as BMW would own the skills its robots learn through micropsi's software, just as it owns other purchased production inputs.

Q&A highlights

Asked about legacy hardware, Vuine says manufacturers are less locked into old plants than they think, since production lines are typically updated every five to seven years regardless of AI; adoption tends to start with better-funded sectors like automotive and then trickle down. On brain-computer interfaces, he notes that while such interfaces could offer very high input bandwidth, language itself has lower throughput than typing, so professionals doing a task well often still prefer a keyboard. Asked whether micropsi combines classical symbolic AI with machine learning, he says the current system is almost purely machine learning, but predicts the next wave of AI will combine pattern detection with reasoning, likely including probabilistic or Bayesian methods, since deep learning alone "finds patterns" but "doesn't reason." On GPT-3, he predicts real market impact within two to three years and warns that once AI-written text becomes indistinguishable from human writing, all online text becomes "fakeable," which he expects to undermine the practice of using text as a signal of consensus and push people back toward judging whether an argument shows real reasoning.

Quotes

The thing about AGI is it is not engineering and it's not science, it's more like alchemy.

Ronnie Vuine

Build the capabilities yourselves, don't allow people to talk you into the idea that you need external specialists that do the digitalization and do the AI for you.

Ronnie Vuine

Questions answered in this talk

How does micropsi industries train industrial robots without programming exact positions?

According to Ronnie Vuine at a Rise of AI Virtual Chat in February 2021, a worker demonstrates the desired movement four or five times in front of a camera, the recordings are uploaded to AWS, and the robot's skill is computed in about three hours of cloud training.

Why is AI adoption still low in manufacturing?

Vuine explains that most factories only attach sensors to decades-old machines to collect data, such as vibration readings for predictive maintenance, rather than redesigning production as a software-driven system the way Tesla built its factories from the ground up.

Should manufacturers build AI in-house or buy it from vendors?

Vuine recommends buying packaged AI products, comparing this to buying a PC, but says staff still need to understand the underlying technology well enough to evaluate vendors and keep ownership of how their production is updated in the future.

Mentioned

skill learning by demonstration, spatial reasoning, psi theory, AGI, additive manufacturing, predictive maintenance, GPT-3, Digital Services Act

Keep exploring

More from Ronnie Vuine

More from Rise of AI Virtual Chat 2021

More on Enterprise AI & Adoption

Transcript

Auto-generated transcript, not edited. · 11,164 words

Show the full transcript

good morning to a special episode of our virtual chat my name is fabian bessaid i'm an entrepreneur and investor with a passion for artificial intelligence and this is a series with conversation a series of conversations with leading minds in the eye we had last year do i would say corona gave us the push we visualized and we had 18 episodes with a bright variety you can check them all out on our website and on youtube and today we have a double feature one in the morning one in the afternoon um please say hello in the chat so if you're here live today use and swap cut the chat say hello welcome to the world tell us where you are where you're from um i know we have different time zones second is in between you can ask questions please use the comment and the questions tool i will see them here later at the end we will go through the questions and it will ask them and directly to ronnie who is already in the

backstage area um roughly the agenda for the day so we have now 45 minutes uh to 50 minutes with ronnie it's a interview we do we also have a prototype of this robot here which we will likely see if it works or not as usual with prototypes um we will have the round tables right after so stay online here at 11 we will finish on time you can switch to the round tables then we have one hour of break for the team and for yourself and then we have the second virtual chat with michael brim at one another sets of round tables one hour after what did you miss since we talked last time most of us have seen us mostly at the risify summit in november we finally get more money for germany so this article is from 2019 when the german government announced 10 billion for startups to invest in the future well the money is finally coming through bundestag accepted everything they are not setting up the programs two-thirds of the money

will be deployed this year through the institutional like kafi and coparion and heartache f and the rest will be set in new formats that is good because it will hopefully attract another 60 billion of private money for deep tech in germany um also i would like to mention that 1 billion is dedicated only to ai from government money plus private money out there and this is good it would bring us and i would say in the top five of funding environments which we have globally still behind the usa still behind china but in europe this is a pole position and i think it's great on the second thing is we have to expect more regulation european commission and the european parliament announced or already have committed new regulation last year it's called the digital service act now the german government tries to implement this for our law making this will we will update you over the year we have enough startups out there we have the researchers we have the money but still the industry

is not using enough ai and especially the manufacturing industry we will have one roundtable which is focused and said only six percent of manufacturing business out there businesses are actually using ai and i will speak later with ronnie about this why we have this case what are the bottlenecks and how can we overcome this so topic today is can we use ai to upgrade our industry and um this is critical for us because i say this is our last chance the b2c the end consumer thing you know it's already in the hand of google of apple of facebook etc but we speak about all the manufacturing the automotive the logistics the production the cnc everything out there it's still very offline very analog so there's a huge opportunity to bring these companies and this processes online virtual make them connected but it hasn't happened yet and this will our be our deep dive for today um my guest is ronnie wiene and we always have a discussion how to pronounce his name right or wrong

i think there is no right of wrong so it's ronnie and i call him a friend he was at my wedding even he said like oh it's raining i don't want to come he came and that was wonderful i thank him for this um also with asgard my venture capital firm i invested in this company pretty early and was one of my first true ai investments really true ai investments he was one of my first guests in my podcast series i was doing he spoke several times for rise of ai i think my favorite talk was why science fiction is a self-fulfilling prophecy and it will be the truth and royalty one day well because people who read science fiction are later becoming programmers developers and ceos um i would call him a philosopher and a tech nerd at the same time so you can drink a glass of whiskey with him smoke a cigar and talk for hours and you will never miss anything you said like oh that wasn't interesting he's also the

founder and ceo of microsoft industries um a company which has by now raised plenty of money one of the best funded ai companies which we have in germany one of the top ones in the world for his field really really strong team good investors but i was thinking about this i know him when he was still doing just software um his company is teaching robots new skills that sounds easy it's kind of complicated i would say we play a short video and at this time we will give give ronnie time to come in the studio corona compliant thank you so welcome honey um it's really nice to see you face to face finally yes we haven't done this in a while uh circumstances have been have not been favorable for this type of thing when you came in you were like a little bit like like my puppy yes like overwhelmed with so many people and influences and noises i mean did you use cobra is it an excuse or is it just the circumstances

that you have significant less interaction with everyone but i i like to think i was just leading by example by staying home and not doing the macho thing and you know proving that you know i'm not afraid of the virus and i'll go to the office but i'm um i'm an introvert enough to so i don't have to pretend i didn't enjoy it so is this like a positive thing for us introverts at least it's not as hard for us that is for i mean that's you know something that i sort of see how hard is for people who like to be out there and who do not need to recharge their batteries by being alone as often as i do so i i don't know if it's a good thing um but at least it's significantly easier for us i think okay so for for the audience out there um watch the movie x machina and then think about him because he's living not as fancy but he's he purchased himself i call it

a castle a while ago and we we joked about it i purchased something too i paid with crypto he paid with crypto that's the time um and uh he is he is developing their amazing stuff i would say somewhere in the random way no one no one can come to him yes it has thick walls uh and it's it's not a literal castle but it's um very close to one and it has castle-like properties and uh i can some some people told me when you're in your seat at your fireplace in your swiveling chair all you need is a white cat to um the james bond gesture would be would be complete it's not what i'm trying to do i just like it but okay i can live with that image if we think about mainstream let's say in five to ten years i mean i just quoted ex machina are we the good guys or the bad guys i like to think some of us are the good are the are the good guys

um it's like you know you can't avoid technology you can't just not interact with it and then hope that you will be fine you need to interact with it and while interacting it being one of the good be one of the good guys you know there's in ai and in robotics of course all the military stuff is going on all the super shady surveillance stuff is going on and um you can just and this is you know no matter what we do it's going to go on right this is this is it's always been like that uh it's quite a quite it's just a law of um how things operate uh so some good guys need to be in it too okay so you were running a company before called texter which was i would say kind of a german version of a kindle yes even before the kindle came out so you did did a little bit of it was tech but it was different tell us what is different running an ai company

and a non-ai company you know the first wave before i will speak later speak with michael about the ecosystem overall but where do you see like the difference and where did you underestimate the difficulty of running a robotic slash ai company i don't think i did the the it is very difficult by the way if anybody's wondering but you know running any company is very difficult um all the ones that do real innovation um and it wasn't that difficult that different than texture actually it's your your main challenges i mean of course you work with people who are good at their field and i did then and i do now and the main challenge for you know from a management perspective or from a vision perspective is actually quite similar you need to keep the organization from believing the stories it tells itself and you you need to force it constantly force it to listen to what the market does um and you know not not not believe the pretty lies um you make up

um inevitably right this is just naturally in any organization this happens so um you know keep keeping an organization tuned to the market while it is doing wildly innovative stuff um and you know many of the ideas we have it would be so nice if they were true and it's just so tempting after you've told this story so many times if you would have been right and if you're so deeply invested with you know so much lifetime and so much energy into things you did and then at some point the market tells you either i don't care um or this doesn't work and then being able to go back and change it and you know abandon ideas that you liked and that you're publicly invested in by having spoken about them is always the hardest and i tend to think we're doing a slightly better job at microbes industries at this than we did at texture this is what experience gives you right you've seen it you know how difficult this is so if i

compare it to companies that's the main difference for me but it's really not in the tech the tech is always hard and it's it's not easy building a kindle and we failed so um i don't think ai is particularly hard in that regard doing companies that do anything meaningful um and you know do it right this is the hard part okay so the good news for you out there ai is not that hard and that's true um i would speak about different phases so when we got to know each other through your former co-founder uli and you were developing a minecraft bot which was autonomous so so you just said okay we built this work model with all those models and the self learning itself but at the end this ai is able to survive it's able to learn the game a little bit like alpha go out there from the core idea learn the game itself survive and you used in game to train it and then you but you ended up in robotics

so tell us a little bit about this skip i mean i've been with your journey but for everyone out there is they see now how you train robots not virtual bots it's it's less random than you might think so and we're ai people so near my background as you said is in philosophy and computer science and one of the fields where these intersect naturally is artificial intelligence so artificial intelligence or more succinctly you know what is intelligence how do you take stuff that's dumb like rocks or you know what we're made from and make it behave intelligently how how does that magic happen that's what really interested me and has always interested me and um the pursuit of this um if you want to explore this in software um just you know start writing code at some point and then okay how do i make this smarter um and doing a minecraft bot was like a step on that way so how do i build a system that doesn't know what it's supposed to do

but you know it gets artificially motivational system so it knows what it needs to learn and how it needs to learn things in its virtual world to find its way around that and that's what we've always been interested how do you not tell a system what it needs to learn but how do it you know how do you make it so it you know finds its way without you um essentially paving that path and we did that in minecraft and then at some point we realized uh we had cheated um yeah we knew we were cheating when we were doing it took a shortcut we took we took we took shortcuts and the shortcuts were we pretended minecraft was discrete so there was like individual places in the world that the software bot could essentially teleport too and it you know it could tell these places apart and it knew what there could be done where there was food in some places and you could could chop down trees and do all the minecraft things

in other places but we really didn't know how to get from a to b so making a continuous movement um we didn't know how to do it and we said okay how do we do that and now we're getting into robotics territory how do you get from a to be in a continuous movement and how do you learn this um and this is really where the connection is and um i also believe that there's a deep connection that at least i don't really understand between spatial reasoning which is you know what this boils down to in the end and intelligence because that's where it has to have come from right that's what natural system what animals and people had to deal with there's a world it's 3d or 60 three translations and three rotations and in that world you need to make continuous movements and you need to understand how the geometry of this world works um and there's an argument to be made that the facilities that nature evolved in us to handle this

3d or 60 complexity in the world is the same technology that we're using for other things when we're being intelligent when we're doing things it's not all of it but a part of you know the the miracle of intelligence in humans and animals um is happening around spatial reasoning and that's what special reasoning spatial reasoning space agreement three rotating 3d objects in your head right this is hard um imagining that something has a back side yes moving around things understanding when you will run into it um understanding that two that you know two views that you have if you take a picture with your phone from here and from there and all you get is you know two planes of pixels how do you understand that you're looking at the same object just from two different sides so that the ai which has a lot of input and sensors is able to gather this and it's built i call it a work model is understanding that there is a three-dimensional space to move this leads

to the first audience questions which i like because uh christian becca he asks is makeup c related to dietrich sterner's original psy theory and i say yes so tell us a little bit there is research behind it like brain cognitive brain stuff behind it or like things or even or even wilder yeah so first of all the answer is yes um this micropsy was like a modest attempt at recreating what this returner had done in the 90s um in in his work uh dildo turner is a german psychologist who had uh from a theoretical psychology sort of reinvented half of ai without noticing so he's not an ai guy he's not an aia guy he's a he came from research on um how does it how does it happen that intelligence goes wrong right when you have a bunch of highly trained people operating a complex system you know like a nuclear reactor um and then you know they they're brilliant guys they're engineers they've been trained um they've been doing this all their life

and then a sequence of events happens and they take a sequence of action and it's completely catastrophic how can that happen um how do people fail and then from from this research there's a good book about this the logic dismissed links logic of failure which i highly highly recommend it especially for managers so this is what he did and then at some point he asked okay i really need a more general model of how this works what is failing there how does intelligence work how do people represent knowledge why are they doing things why are they switching strategies when does it happen um how does emotion play into it right in the western tradition we tend to think that emotion is sort of like a disturbance of rational thought that progresses and that's if that would be true nature wouldn't have evolved it right that's obviously wrong um so it plays a role in intelligence um and that's something that uh disregard was interested in and in our university days um we read that joshua

also co-founder of michael yes who is uh by the way at mit and he is used to be he used to be at his emmy now um with a couple of very interesting that's where our talents go either they're here with us in berlin or they end up in the valley yeah so and together with yosha we read um there was like a just a reading circle around joshua who was a lecturer at humboldt university and i was studying with him at that time and we read donor and at some point we said you know this looks this this this sounds like we can just implement this in java because it was the early 2000s so we were all doing java um so um you know we started to write code um and said you know this is like a micro version of what did richter is doing properly because we're just students and this is the relation very little of dietrich derna is in our robot control today but the way we approach ai

and the way we think about intelligence and um representation and uh you know what the interesting problems are um is very much informed um by this background that we have so when we met we joked a little bit about that you're one of the few teams who are working on agi like artificial intelligence but you're too realistic to pursue it and you said you yes it's a dream one day we want to achieve but we will now go for i call a product market a real product raising money and you know do you convince investors more like hey i can solve all what problems one day because they're amazing or i have a use case here like i can make money yeah i i i tend to sort of not talk about agi when i talk to investors you were the exception because you clearly enjoyed it um no the the um the thing about agi is it is it is not it is not engineering and it's not science um it's more like alchemy

you essentially muck around with ideas that you have and you try them but you can't really do that in a systematic way so that doesn't mean it's not real and it doesn't mean we won't get there it's just bad business because you can't predict you know how much money you need to spend on what um to get what results exactly and then you know some of the things you need to do to get to agi which is working representation or work on artificial emotion doesn't really help anybody it makes a very intelligent system if you've built the full system but you know if we added emotions to our robots right now nobody wants a moody robot one of the advantages of robots is they're never moody uh they just do what you tell them um and that's sort of the point so so not all of what you need to do for agi and it's by the way it's debatable whether you need emote emotion but um i tend to think it at least helps

to do something that's credibly and that can be approached as an as a human-like intelligence by by humans um is is really hard to justify in a commercial context and we did some mucking around in the early days um and that gave us perspective of what's you know what is what the probable distance is and where realistically we could get to something that would make money and if you're in a venture-backed company and you do ai you need to have a horizon of 10 years to exit which means three years to product market fit and four years until you start making money and was just after two years very clear that some of the agi stuff that we were doing wasn't predictable enough to pursue it while also you know being a business okay while um you mentioned no one wants a moody robot i can recommend to read the book qualityland uh which has hands full of moody self-driving caps and it is very entertaining if you are in a cab which thinks no

i don't want to take the shortest way today i'm a little bit bored and drives randomly around and drops you off at a random place a little bit like if you take a cab in hong kong or mexico happens you're the same that they drop you off where you not want to be but where they think it could fit um before we jump a little bit about you know let's speak a second about the where you currently are and then let's showcase the the robot you brought up so you started very theoretical um for years you had nothing to show except of some interesting videos and theories and then you came into it like where are you standing today like company wise you have paying customers you're shipping you're selling you're making revenue right yes we do um very serious people buy ai to from us to put it in their factories and solve problems they so far couldn't solve and when i say solve i mean you know automate things that so far couldn't

couldn't be automated for instance in testing right you need to find something that always looks similar but never the same and it's never in the same place so you need an intelligent and intelligent system that you can train on site uh to your product um to to find these features so for instance uh one thing that's ongoing right now is we're testing refrigerators um with robots that are driven by a system that's that has learned how to move and how to find the interesting things about the refrigerators in the very factory so the way this works is a customer buys a robot just some industrial robot they hook up our system they hook up some cameras then their engineers show the robots how humans would make the movements that they want so they just you know bring the worker that's been doing it the robot watches with its two little eyes is cameras and infers what people are doing there and then after a couple days in a bit of tuning you can essentially swap

the robot in and they will be doing uh what so far people have so we will bring in now the robot and showcase it mark please and um he will just start playing around we will comment it so for me my understanding is the first problem is a robot can't learn normally and it's always doing the same task but your robot is able to act on new environments so he's learning he's self-improving yes so traditionally so this is an industrial role but we don't make it right now you can buy it off the shelf you buy this off the shelf essentially just the software that drives it right now um is by us so normally yes you teach these robots to go from a to b right to precise positions on the table and it will be able to repeat that very precisely but knowing where to go if things differ like in this case the position of that microbes industries logo m on the screen is hard so this is an extremely simple case

um the thing about this is you can train this with like four or five demonstrations and like training i mean we have done it before so how long do you need to train a robot now to do a simple task like this for this this is like 20 minutes and then you're waiting like three hours or so or three half i think um for the cloud to compute the skill and normally you have to hire someone right like a full-time employee to do the same job for many use cases yeah i mean if you if you did that by hand without learning agi um you know i mean this this is solvable with a standard industrial camera today unlike the fridge case that i that i mentioned earlier but yeah then you have somebody who needs to understand neural networks um and um they need to gather the data and they need to know what they're uh doing when they're preparing the data um and with this essentially you show the robot a position above

the m uh like four or five times this is enough for this demo um and it will just infer what it needs to do so everyone out there who's interested in this they could contact you um interesting so a little bit more so this is off-the-shelf robot yeah you have customized the camera i think all right so the camera is an off-the-shelf camera we're not making it either so this is a boring camera the demo would work with a just a standard webcam for 25 bucks of amazon there's nothing special about the camera we like that one because we control the shutter and we are running a robot um there are cases this is not one of them where we need precise movements and then a precise camera helps but the resolution isn't even that high i think that's like 640 or some you know vga or something and then this is so the software which is yours we don't even see i mean sure so it's here in the button right you you your

final product is like what how large is it like there's there's a there's a little box looks like a desktop pc um or a like a set top box on your on your tv uh it drives the robot's motion in real time so it evaluates the camera 20 seconds of time 20 times a second um sees the image and decides okay how do i need to move to make the right movement the way to think about this is traditional robots when you for instance want to plug in a cable into a phone for charging would do a measurement they would try to find out where is the phone you know in 6d spatial reasoning so what's the position and what's the rotations of that phone and i make a single um measured movement towards yes like perfect shot so we humans don't do that no like you usb cables right you try it we make a continuous movement we wobble about this is the connection to minecraft earlier right how do you get from a

to b without knowing uh the other side without having done the measurement without having the ability to be precise so this is what the software does it's evaluates the camera and makes this approaching cautious movement correcting itself while it approaches and that's happening on the compute box that's uh inside this this demo because it's a partnership today you you run even here on aws the skills when we create skills so the compute box in there is connected to the aws cloud and we're using their gpus and their storage for training the skills so when you record data you show something to the robot um the data gets uploaded to aws it gets processed by um secret um and very advanced algorithms by us but on amazon's cloud info so you use aws for training yes essentially and here you have what we thought like nvidia chips or you feel like customized board did you or is it official the the the compute box that's in there the edge device that's in there is essentially just

the gaming pc so there's an nvidia gpu in there it runs linux and we use very little capacity of that gpu for the forward pass i thank you for the demo yes mark it's wonderful and i would like to ask you before we take a seat um it's like like synchronizing when a client is when a client is skilling the eye do you benefit from this is like do you learn with everyone else or do you have to pre-train everything and then it's built on top like where is are there synergies if you have more than one client there are synergies from more than one client but not because we mix data and certainly not because we you know show data from one client to the other it's a you know people are commonly worried about um you know are my competitors going to see this data or are they going to benefit from me uploading this data the answer is no of course not that you know would not be that would not be

a good proposition the way other people benefit from more data being available available to us is mostly because we can backtest changes to the algorithm so you know whenever we change the algorithm with unit we make it more precise we make it faster we make it make it more robust against different light conditions all these things we have a wealth of data available from all our customers today to test all the changes against this data so we know if we make that change to the algorithm we would still be able to solve all these cases at the same quality or better but it's not that we're training a single ai that knows how to you know pick our logo um and test fridges and do all the other things not saying it's not possible it's just not necessary for the application that we're doing and there will be no benefit from it it's also that in the industrial space that we're working with people are super concerned about their data and rightly so so we

make very strong guarantees about who is going to see the data and what will it be used to and that's by the way why people let us do this in a cloud right uh if if we if we would just freely give away the data uh then people would very understandably and rightfully so uh have a problem with giving us that data for training so i would like to close a little bit about microsoft industries and jump in the middle level with you yeah because the second half of the conversation and um i mean the typicals can ai help us to upgrade the industries and we spoke before and you helped me a lot to figure out the questions you told me it's like all the plans out there and now cloud providers are coming and trying to provide them with electricity or cloud data management but they don't find the plug it's like you startups like you are already you have solutions uh big like from ibm to aws to dodge telecom t systems

they're all out there and selling the bases the infrastructure for it and it's not there yet like what are the reasons the time questions is it or is it like institutional problem it's made it's mainly you know proper technology reasons um i would say we all got used to being able to install software that's what i can play yeah install the app on your iphone yeah and not just that it's easy but that you're also dealing with hardware that uh wants to run software um so yeah we all got used to pcs and then phones and then tablets and upgrading the software on there is very easy but in an industrial setting you're working with machines that are sometimes 40 years old and it's not just that they don't contain chips so you can't just install stuff on it um it's also they you know they need to be from metal um um that the core thing that they're doing um has nothing to do with software right they're deforming just a piece of plastic

for instance um so um it's not really clear what when you're a very fluffy cloud and you have this very solid piece of metal that wants to do one thing and does it well for many many years you know how do they how do they connect and there's a couple of obvious ideas right please you can just you can just you can what people have been doing for a couple years now is you just stick a sensor to the piece of metal and then at least you have some data about the machine right i've seen a lot of startups who say hey we're able to measure the temperature or the movement like like vibrations yeah and from the vibration we can derive with the machine might be broken one day or not yes exactly so you you you're desperately trying to do something with data with a 40 year old machine that just doesn't want to do data it wants to do metal and then okay you're at least doing something and you can create

a dashboard and maybe if you're really good you can actually do predictive maintenance which is what you mentioned right maybe you will be able to predict when something's off um and needs to be needs to be maintenanced um but you're not really digitizing your production you're pretending to so next idea this is this is something i mean it's great um but um this isn't going to change the game no no it's like taping something you have a car you just upgrade it in that you buy in you yeah it's it's it's it's not changing anything so you need to you need to accept that you can't move as fast as you can in a probable proper digital environment where you know you have product cycles of two years phones get exchanged roughly every two years and then you have more compute power and also everything wants to do software and in production for very good reasons um we're just in a very different world people invest in these machines they are like they are they're

very good at what they're doing and you can't just digitize them by wanting it to happen so um but how can we digit because we need to it's like i'm overthinking like if we don't do it and we just continue like today in 10 years some countries some nations some other companies have solved it and they're like sweeping us away and we have seen this before because we were not digital either as humans before and then iphone came in all this like boom yeah and it's it's quite obvious that the the the way to you need to take a more systemic approach right you need to move away from we're essentially digitizing individual machines um where it has happened i mean in cars it just happened right um they're rolling computers tesla did and when we say there's a ruling tesla right the german cars are not yet um because they're sort of built from the ground up from the functions of a car and then they sort of digitized in individual places um and

they're still emulating the old world just for you out there um i think in cars there over like 250 uh ai modules they call it like 250 different components which are all smart itself and steering like just electronics or just the heating or just the display but there is no operating system to connect all of them and the advantage of teslas i mean tesla's good example they built from the bottom up at digital production digital factory i mean we can learn from them what do they do right what we can learn from them so i i don't know tesla intimately but it seems to me that uh what's happening there is that people with a california mindset software people are designing something that so far hasn't been a computer so they think about this like software they modularize um they build interfaces and they make everything updateable um over the over the air um so you know you can have a software update for your car and actually makes you better and you know the

car gets features that it didn't have before you're not trying to do something in digital that is as good as the old stuff and just has some digital advantages but you think of the whole thing as a digital system um with physical components and i think that's you know increasingly the world that we're going into when we're looking at production as well um you know instead of starting from individual machines and then connecting them and measuring something about them taking a systemic approach means okay how do i treat this whole system as something that can produce stuff that's a factory right you put in raw materials on one side and something comes out on the other side and it has always been understood how to build a factory where you you know put one type of raw material in and it produces exactly the same thing on the other side that's the easiest case um and now in germany we're already good at variants same materials but very different variants on the other side but

now you know take the radical approach how would you build a general factory how would that's where you put anything in and anything anything like a 3d printer with unlimited access exactly additive manufacturing 3d printing is exactly the right way because the software because it's completely software driven and then you see why it's so interesting to be in robots because um 3d printers are essentially robots right this is a portal robot yes three axes and some of the 3d printing uh plays out there actually use you know like the one i showed just now just but larger six axis robots for to getting to different places um in space and put some material there and if you have that if you have very um abstract machines that can get to any place and do many things in these places you're you are in a software world now now you can imagine a factory that you know doesn't make cars or beer but a factory that can make beer and cars depending on what you what

you do what you do with the software conversation that's like it just won't happen next year right this is really hard but i think this is like the chinese dream for their plan with chef you know for the plant economy because they then they can calculate oh i need 1 000 cars this week and i mean that's why they they do a lot of interaction they do a lot of things so you you say the goal is i call it the factory where everything is automated just two humans you know two livings they're saying that's a dog and a human and a red button and the task of the dog is to make sure that the human is not pressing the red button and by the way many factories there are industries that are like that right productions right production for example production i don't know but for instance beer right um there is no um outside of berlin craft beer circles um this is a matter of pressing a button and then everything happens

um and you just need to pet the dog while the factory makes beer so in some cases that's already possible it's more the case in productions where you have processes where you have essentially fluid materials that just get pressed around through tubes it's harder in discrete production where you have assembly like cars right there's a reason where you need to move parts you need to move parts and you need to put things together and it's fiddly so today you need a lot of humans and the the way forward there is obviously robots right you need you need to do that with additive manufacturing um so so what can you recommend because we have questions here for example a little bit of agi and we'll come back to the end but also like 3d printing and pharmacy or you know what skills we have but what can you recommend anyone we have a lot of audience here who is in the industries at different positions and they are not the coders themselves they are management or

decision making what can you recommend them like to move forward at least in the right direction very hard to do in a very in a general manner i think generally what i can say is you know build the capabilities yourselves let not talk any let don't allow people to talk you into the idea that you need external specialists that do the digitalization and do the ai for you you don't need to be able to you know program the neural networks yourself there are products out there ours is one of them um but you in the future will need the ability to understand these things and to handle these uh these these things it's not a matter of you know i just produce stuff and all the computer and all the computer stuff is done by some external uh consultancy or system integrator the the ones that are good at this in the future are the ones that you know producing stuff in the future um are on top of their digital means of production uh

are on top of the ais that they need um they're on top of what's possible and what's not possible yet and i think i think the most general advice that's probably reasonable is to build it in-house to spend the money spend the time invest it don't do any pilots and then somebody builds something for you that you don't really understand and then after two years you need to produce something differently and your core ability your ability to update your production um has been outsourced um to somebody who claims they're good at ai i agree but you mentioned good core ability because ai is also in recruiting you can use ai like soft as a service out there for you mention your finances your cash flow stock trading um so i personally see a lot of ai not in the back office yes from accounting bookkeeping you know all these where i say like hey you can purchase a solution out there you can use software to make your life better but you say when it's

your core competence is for example production you have it in-house just as an example i recently spoke with kerber kerber is huge i didn't know them before and what they did good is they have a digital arm now that so they have a venture arm to invest into startups but they also have digital venture which is more like an in-house consultancy which is doing two things help digital transformation in the whole company and build prototypes internally with all the business units i know that siemens is doing and ollie Waltinger he is one of the topic leaders later in half an hour he's running the whole thing at zmens where they have one core research center and then they have business units all with market knowledge but where do you see whether you draw the line like let's buy the solutions because it has cost millions to develop for the startup out there versus no we want to have it internally so my favorite analogy is pcs um in the 1970s it was easy to believe

if i want to take part in that first digital revolution of introducing just digital technology for instance for accounting you you could have this idea so i okay i need to build um a computing center and people in white coats are going to run it and i need to buy it from ibm and you know i can't afford it as just a small medium business and that's not what happened um you got microsoft excel on an ibm pc um and your i.t dude was not from mit and he was good enough right um and that's that's what we're going to see in industrial production as well we're going to get products ours happens to be one of them uh that packages a piece of technology that is actually extremely complicated and you couldn't build it yourself and makes it usable for a set of use cases um like pcs right it's very generic um but i mean you can't use a pc to drive in a nail but whether you do accounting or word processing

it's sort of general and that's that's what we're going to that's what we're going to get in industry too so packaged products that sort of wrap the complexity of the deep tech that's in there but are easy to use in the sense that you can hire a guy who's good enough at this they don't need to be you know mit labs yeah or or or yeah or you know the dit guy in every company and you have an ai guy in every company that understands enough about this to select the products deploy them well and you'll be good enough to be part of that next revolution that also means he can purchase products so startups out there still can build the complicated tech make it purchasable what let's say maintain its contract service contract hardware purchasing contract but then the ownership so the skills for example when i'm let's say bmw i i use yours the skills stay with me so i train them and the skills are then property of bmw yeah that's how

things should be in a in a in a you know bmw doesn't make everything they use to build their cars themselves of course they you know buy their robot from somebody else and they buy the machines from somebody else um and now they buy the software from someone else exactly and they will they will buy their production um um it or ai from somebody else um i i think that's just a natural way of things and you know what one of the things that have happened in ai because it's so new that everybody said okay no i need to understand everything about it and i heard about neural networks during my studies so i'm just going to grab this and try to be completely on top of it and you don't need to do that right you need to be on top of it enough to select the products to understand what's real and who's bullshitting you um and then you know put the right things into production uh in an efficient way like with

you know any other major innovation did it or not um that we have that we have had so we have plenty of questions now and i will try to go let's say highlighted them a little bit the first one quick is how can you skill yourself so people are in the chat asking like oh this will all come and every market is different we can't go in the market what can you give them like how do you train skill educate themselves um so i think it's it's if you're if you're trying to skill yourself for ai and digital production or ai driven production it's a good it's a it's it doesn't hurt to understand how the base technologies work you don't need to be able to be able to you know just program them but just to understand um how the base technology works and you can't do that in youtube or khan academy or wherever um is is certainly a good idea and then um just i mean this is this is easy to

tell people but you know try to develop a sense of what seems realistic um just from my experience i mean i invested in the ai and i'm the business degree by background i um i sat down with ronnie and he was teaching me for hours what he's actually doing and i said oh that makes sense i read a lot of papers he recommended i read this and said oh that makes sense too and then i got a feeling over the years of feeling like okay what is out there when startups come to me oh we can do this i know you can't i don't believe you and like oh that's cool and then we dive into it so i would say have a core of understanding of the technology to understand the bull i haven't built a detector for for buzzwords at the end yeah yeah exactly there are a lot of passwords out there and you really you really can you can only tell what's real from what's not uh when you have like

a base understanding um of where the research is um but that's all on youtube uh you can you can you that's you you can you can you can find out next uh question from boca schwab and i like it too it's uh he says like since we're mostly dealing with legacy hardware manufacturing um does that need that we first need okay he asked do we need to invest in new manufacturing equipment before we can truly harness the power of ai and if so when we speak about plants who were built 30 40 50 years ago and they will not be built again soon because germany is old old infrastructure everything is already built like if you only focus on new plants very late market so what do you we had the discussion before yeah i don't think we're as locked as when in as we believe even if the plants are old um production lines get updated very regularly okay and you know if you're extremely lucky if you have a plan that you can

run for 50 years and make money off it you know if you can okay continue doing it don't touch it it's fine it's fine um but usually you have like this update cycle where every five to seven years something in your product changes regardless of what you're making the world moves on things change and then you update how you do this so you have this traditional industrial investment cycle and every every company has sort of its own ways of updating uh their uh their their production uh facilities the thinking um is harder to change so we have this investment cycle anyway but you know not trying to um many people when they think about the next update or upgrade to their to their productions i think about okay how can i buy the next best individual parks um the the overall okay can i do things completely differently um with you know new technology is happening in very few places very few people can afford it um but i'm actually you know i'm not alarmist

about it and just it will take some while right it starts in places where people can afford this automotives are good at this right they they have the money to think about okay how is the next wave of digital production going to look like and they will experiment with it and then it trickles down to smaller companies it's not always automotive there are other innovative fields uh too but and then it trickles down um just you know people disperse knowledge disperses and people get good at rethinking uh the rules of the game that they are playing because they really change right you're not assembling something from good parts um if you do software type thinking um but you you build something that's very flexible in the first place and then the requirements for the individual parts change i like this black box idea put something in something from an outskirt like ingredients in pizza out and by the way later in the witch in the round tables perfect just continue this discussion ask your questions

there and ask there the people who are topic leader who are more in the market itself a little bit more focus for fast-moving consumer goods or for industrial ai et cetera another question from varina so how will work be defined soon if there is ai and i work together do i need like a brain machine interface or like she says it's hard to vision the consequences and we will go with michael closer to this but here let's stay with industrial solutions there and the mark so yeah i mean it's it's famously hard to imagine the consequences and if you just look back into history people get it wrong all the time so we don't know what will happen right and that's the pattern um the change comes slower than you think and when it comes it changes more than you think because you can't even imagine where all the change is going to be i don't know if it's going to be brain computer interfaces the brain computer interface has you know an ideal the

ideal brain computer interface has the advantage of having extremely high bandwidth keyboards or touch screens have slightly lower bandwidths yeah exactly right it's all about um how do i get information from here into some digital form uh you know and the other way and the input stream is good we have eyes right they're very broadband um we can get a lot of stuff information into your brain uh by eyes getting it out is more difficult you have like your hands and then you you use these 102 keys so you have your chatbot you know your algorithm is translating your thoughts into words to tell exactly yes but language is not necessarily it's very natural so it's very easy for people to use um but it's lower throughput than typing for instance um so if you're if you're a pro and you're good at what you're doing you prefer a keyboard over just speaking to the machine because you're just faster with your keyboard or you're an interface that you're that you're attuned to imagine driving

a car by just giving it directions in language it's i do this with a cab driver yes but uh you're cheating yeah i'm cheating by the way i'm my vision is i mean you have a large tablet there so i believe there will be a lot of tablets one day you will run through your factory and then you have just 20 machines in your tablet and you know where to control them and do like real time and then you have a huge back office where you do all the yeah that's already happening right i mean this um the dashboards and tablets and knowing where the issues in your uh in your on your factory floor are so sort of mobilized manufacturing execution systems are already becoming a reality and i see a lot of good stuff happening in that space so christian becker is asking do you use good good old-fashioned ai combined with machine learning yeah that's a question you would ask a micropsy person or a c person we don't right now this

is almost purely machine learning right now there is there is something to say for uh gophi and i think the next generation of ai after you know all the rewards from deep learning have been have been brought in and when people move on and say okay what's the next step what's the next major change we can get through ai and go back to the technology some combination of reasoning which is what you know deep learning is family famously bad at um it just finds patterns um it doesn't reason some combination of probably probably bayesian reasoning and pattern detection is going to be to be the the next um wave of ai uh that we're uh that we're good like like last two minutes what will be the next wave of the eye because you're closer to it than anyone else here in the room like where where do you see the treachery going what direction yeah i don't know how fast this will will go um unfortunately and i i really don't know how f

how far the people who just throw um cores um at passion detection will get right we we we we got spectacular results out of gpt three this is this language ai uh that's from open ai that uh you know does astonishing things by just finding patterns in what we do on the internet and it turns out you can learn a lot about the world and how to do certain things if you just watch all of humanity do stuff on the internet um which by the way the nsa is doing yes i thought that they're as clever about it as open ai but maybe but you know they essentially read all of reddit um and turns out reading all of um openai with gpd3 right reading all of reddit uh tells you a lot about certain communities of people and how they operate and what they can do so these these techniques get you very far um i would still believe um that there are significant limits to that and we'll start to see them as

we you know explore these in the moment you build products with it yeah yes and by the way the you know the impact of products uh that are gpt3 based once they come will be will be very interesting massive and extremely i would say two years it will take i mean startups are working with it now but until it's really hitting something interesting the market two to three years yes hard to predict i don't know how much of it is accessible for for for startups but once it happens um you know all language that you read on the internet becomes fake-able yeah so like everything everything that you read may have been written by an ai you can't tell anymore just by looking at syntax and by just looking at what you're seeing so the question will be you know do i see any signs of actual thinking and what i'm reading oh yes and that's something we don't do today right we just say okay this is an opinion i sort of second so

i'm going to signal boost it which is a broken behavior pattern anyway and if you just have ai to signal boost essentially anything any opinion can be stated as much as you want and as broadly as you want just the existence of a piece of text that makes a point will not have any value anymore so i think we we might actually move away from this horrible world that we're now in where signal boosting stuff on the internet is to the way to win a public argument um and people will have to they're just going to be forced to go back to the is there anything new in that has somebody is there is there an actually intelligent well-reasoned new argument in this and otherwise i'm just going to ignore it as you know some other stuff that some ai has said and they say a lot of things that they will in a couple of years okay the next wave to really answer the question is not going to be that i think there's

more and uh reasoning spatial reasoning and actual logical reasoning needs to be in there so that's why i believe in the combination with the basis okay as i said we can talk for hours about this if you like what ronnie's doing check out his company microsoft industries add him on linkedin even he's introvert as we most are he is online and you can contact with him he says where's my ronnie thank you very much for being here everyone else please stay in the in the stream for a minute too um because we are now heading to the air topic tables and i would briefly introduce them to you it's a little bit easier for you to choose first they are kept at 50 per person so it doesn't mean you can enter each of them with the problem in the beginning that we have keynotes and all this but networking is coming to shred especially under corona like you don't meet new people you hear them you listen to them but you don't get their

opinions and you can't contact with them with the round tables we're trying to overcome this that's why they are no recording we don't save any data there for one hour talk freely do whatever you want to and use this opportunity um table number one is a combination of olivine daniela nico and henning and they will talk about scaling industrial ai or in other trips they will talk about the ecosystems so we have the bmw siemens infineon all there the ai people talking all this so this table number one and also thank you oli for setting this up he's a very proactive person trying to build this ecosystem with us table number two we have fabian schlatt that's who's head of ai at capgenini and wagner who is specialist at for ai at aws and they had brought up this idea that only six percent of manufacturing is really driven by ai and there's a huge gap so they will talk about scaling ai and production processes and then table number three is with krishnan ai

lead at aws and he prepared a good session about how can you make more revenue with data how what is implementation how can you do the challenges and how that you need a culture of experimentations um i also would like to thank aws for being our partner today especially ralph dieter who is in the background our contact person and is really supporting us this is important honestly we built a whole studio here we have a whole crew camera team a lot of setup check out us out by the way on instagram to see behind the scenes and this is your setup and this is supported by aws so we can do this for free for you you can join you can be here you can share this or later watch this on youtube at the same time and um before we go to the round tables enjoy them now we see again at 1pm and i think i'm short on time i say thank you enjoy your round tables and come back online at 1pm

see you later you

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.