Best of LinkedIn: Smart Manufacturing & Industrial AI CW 41/ 42

00:00:00: Provided by Thomas Allgaier and Freeness, based on the most relevant LinkedIn posts about smart manufacturing in industrial AI.

00:00:06: In calendar weeks forty one and forty two.

00:00:09: Freenes is a B to V market research company that supports Smart Manufacturing providers with building feature-by-feature competitive intelligence That shows exactly how their product stacks up against the competition.

00:00:20: You can find more info in the description.

00:00:22: And welcome to the deep dive, you know We're taking this curated stack of professional insights from the last couple of weeks and we're really getting into The messy reality of smart manufacturing and industrial AI.

00:00:32: Yeah,

00:00:33: and we are moving way past the hype phase.

00:00:35: here I mean looking at these discussions From industry leaders.

00:00:39: it's obvious were in This era of high-stakes execution.

00:00:41: now

00:00:42: absolutely No more just talking about AI potential.

00:00:45: We're unpacking massive market consolidation, the shift to edge computing and well what it actually takes to get autonomous systems working on a factory floor without causing total chaos of course right

00:00:56: but blowing things up.

00:00:58: And To ground this we really need to separate The news headlines from the actual plant floor because you know There's a fundamental difference between consumer AI and industrial AI.

00:01:08: Oh that distinction is exactly where we have to start.

00:01:11: Roland Bush had a really great perspective on this recently.

00:01:14: You know, when the public hears AI they think of a large language model drafting a quirky email right?

00:01:20: Or like generating a funny picture of a cat.

00:01:23: Yeah exactly and if that model hallucinates you get a weirdly phrased sentence.

00:01:27: just minor annoyance Right

00:01:29: but If an industrial AI hallucinats A boiler could literally explode or a robotic arm swings thirty degrees too far And Just takes out An entire assembly line.

00:01:39: Oh wow Yeah, the stakes are completely different.

00:01:41: Totally industrial AI powers the physical world power grids automated guided vehicles chemical mixes.

00:01:49: it's working with data that standard public models have never even seen.

00:01:53: like what kind of data?

00:01:54: um Like proprietary machine vibration telemetry or highly specific thermal signatures.

00:02:01: The tolerance for error in those environments isn't just low It's zero.

00:02:05: I mean if their requirements are that stringent.

00:02:08: Building these models has to cost an absolute fortune.

00:02:10: Oh, it does.

00:02:11: Peter Cortes shared some really stark numbers on this.

00:02:14: so last year the US poured two hundred eighty six billion dollars into AI development.

00:02:20: Two hundred and eighty-six billion?

00:02:22: Wow!

00:02:22: Yeah And you're up by comparison they invested just twenty one billion dollars.

00:02:28: Wait...

00:02:28: Really?!

00:02:29: Just twenty one.

00:02:30: So I mean if look at that ten to one ratio purely for consumer AI The race is over.

00:02:35: the U.S wins.

00:02:36: It looks that way.

00:02:37: Yeah,

00:02:38: but capital

00:02:39: alone can't buy that proprietary telemetry we were just talking about and That is Europe's massive structural advantage because

00:02:46: of the older manufacturing base.

00:02:47: Exactly you've got these century-old companies sitting on mountains of proprietary production data Historical sensor readings deep engineering context.

00:02:56: You can't just scrape that off the internet right.

00:02:58: it takes decades to build that up

00:03:00: decades And it requires immense trust from customers.

00:03:04: That's why Siemens is committing over one billion euros to industrial AI, Over the next three years.

00:03:09: they know their historical data ecosystem Is an insurmountable mode?

00:03:12: Okay But let me push back on this money versus data thing for a second.

00:03:16: because having decades of proprietary sense of readings sitting On a server isn't that basically like sitting on A massive untapped oil reserve but you completely lack The refinery to process it?

00:03:28: ooh It's a good way To put.

00:03:30: I

00:03:31: mean, the data is just crude oil.

00:03:32: It has potential sure but without the AI infrastructure The refinery to process it into actionable insights that Data doesn't drive the engine.

00:03:42: That analogy holds up.

00:03:42: perfectly honestly raw data on its own does not improve yield On a production line and if you look at the major players they clearly know They need to buy or build those refineries right now

00:03:52: like the PTC deal

00:03:53: Exactly.

00:03:54: Neil Barua announced that PTC agreed to an all-cash acquisition by Schneider Electric, and the value on that is an astonishing twenty two point six billion dollars.

00:04:05: Wow!

00:04:06: Yeah.

00:04:06: Twenty two billion?

00:04:08: You don't see deals of that magnitude unless a company is aggressively assembling all pieces into one unified end-to-end industrial software in AI portfolio.

00:04:17: So if you're plant manager listening this right... A twenty two billion dollar deal sounds impressive But what does that mean for daily operations?

00:04:25: Are manufacturers actually ready to let the AI make decisions on the floor.

00:04:31: Well, The short answer is that the technology is ready but the organizations themselves not so much.

00:04:37: no classic adoption bottleneck

00:04:39: right.

00:04:40: Chris Stevens offered a really grounded reality check on this.

00:04:43: he pointed out That fully autonomous factories are technically achievable Right now

00:04:48: like today Really

00:04:49: Today.

00:04:50: As an example, Procter & Gamble is already using AI inspection systems to cut their scrap rates by up to twenty percent.

00:04:56: The tech isn't the bottleneck — it's implementation costs and complexity—and mostly human willingness to trust in algorithm over a veteran

00:05:05: operator.".

00:05:06: Which brings us back on the core issue because trust requires context!

00:05:10: All these billions of investments mean absolutely nothing if the AI is operating at a vacuum.

00:05:15: Ujwal Udramdurg hit this exactly...

00:05:19: What did he say?

00:05:20: He argued that factories don't actually have a data problem.

00:05:23: They have a meaning problem.

00:05:24: Oh, I like thought

00:05:25: right

00:05:26: Meaning problem

00:05:26: because if you drop an advanced AI model onto a shop floor without mapping the specific domain context You're just empowering a very fast Very confident system to make terrible decisions based on a fractured understanding of reality.

00:05:38: Yeah A fast bad decision is still about

00:05:41: exactly and Anchil saw who shared a great breakdown Of what happens when you actually solve that meeting problem.

00:05:48: A team was using SAP Juul to analyze non-conformance records.

00:05:52: Okay, so like quality control alerts?

00:05:54: Yeah

00:05:55: exactly normally an operator sees a flat alert Like temperature violation on line four and then they have spend hours hunting down why it happened.

00:06:05: Because the temperature spike doesn't tell you if the cooling valve failed or If the sensor itself is just broken

00:06:11: right but because this AI was given deep context didn't flash or red light.

00:06:16: It correlated the violation with a specific mixing operation, tied it to a specific routing step and even cross-reference maintenance logs to flag equipment

00:06:25: degradation.".

00:06:26: Oh wow!

00:06:27: So its'a multi dimensional diagnostic tool?

00:06:30: Exactly though Sahu is very careful point out that these are just potential root causes.

00:06:35: you still absolutely need human expert validate ADI's logic against reality

00:06:40: And that human validation.

00:06:42: But getting that level of analysis means your underlying systems have to speak the exact same language.

00:06:48: Boris Shalimov talked about how critical it is to build a semantic thread across the V model.

00:06:53: Okay, for anyone listening who doesn't live in Systems Engineering The v-model Is basically the life cycle of our product Right?

00:07:00: Yep Starts broad at the top with concepts Goes down into detailed design Hits manufacturing At the bottom and goes back up For testing and service.

00:07:09: And traditionally, every one of those steps uses completely different software naming conventions.

00:07:14: It's a mess and Shalemov argues that without a knowledge graph linking all of that A true semantic thread your data integration is really just bureaucracy

00:07:25: via API.

00:07:27: I mean That has to hit close to home for a lot of engineering teams.

00:07:30: it's painful.

00:07:31: someone has to constantly export rename and reconcile spreadsheets.

00:07:36: If you throw an unsupervised AI into the mass It might just guess at relationships.

00:07:41: Which is terrifying!

00:07:42: Right, a plausible guess.

00:07:43: it's not verified engineering traceability

00:07:46: Really quick before we talk about how dangerous that can be.

00:07:48: If you are finding this deep dive valuable make sure to subscribe.

00:07:51: so catch all our future additions.

00:07:53: We've got alot more coming.

00:07:54: Yes definitely hit subscribe.

00:07:56: So speaking of the lack of traceability introducing massive risk.

00:08:00: Ravi Kunji brought up great scenario involving bill materials or BOM

00:08:05: The literal DNA product Every screw, every casing.

00:08:09: Exactly!

00:08:10: Kunju pointed out that while a BOM might look like a simple spreadsheet letting unsupervised AI agents edit it is incredibly dangerous.

00:08:18: He said It's like letting an eager intern post to your general ledger just because it looks like Excel.

00:08:23: Oh man That's a disaster waiting to happen.

00:08:27: An AI confidently making the wrong edit could cause factory order millions in the wrong parts

00:08:32: Or halt-align Because dependency was deleted.

00:08:36: Letting a wildly enthusiastic sous chef rewrite a Michelin star recipe, but they've never actually tasted the ingredients.

00:08:43: Yes!

00:08:44: Perfect analogy.

00:08:46: We're rushing to hand over

00:08:48: what?

00:08:48: To AI before we rigidly define why?

00:08:51: Yeah that enthusiasm is just totally outpacing governance.

00:08:56: It really is.

00:08:57: But once you do establish that governance, You can start pushing AI out of the server room and into the physical world which is driving this massive pivot toward edge computing and physical AI.

00:09:08: Right because The Cloud isn't always the answer.

00:09:10: Helene O'Jilliburt shared a scenario That perfectly illustrates us.

00:09:14: Imagine a chemical plant running twenty four seven And they get recurring valve failure.

00:09:20: Okay.

00:09:20: The engineering team takes their highly proprietary process data, exact catalyst ratios thermal profiles and they feed it into a public cloud AI to troubleshoot.

00:09:30: I mean the Cloud model probably finds the fix in seconds.

00:09:33: It does its incredibly fast.

00:09:36: but the trade-off that plant just uploaded Its most closely guarded recipes tool multi tenant server They literally gave away a core competitive advantage via an API call.

00:09:48: Oh, that's a huge security risk.

00:09:50: Huge!

00:09:51: Gilliburne notes.

00:09:52: with industrial ransomware up sixty-four percent.

00:09:55: edge deployment isn't just about speed it is data sovereignty and protecting IP.

00:10:00: Absolutely when you process the telemetry in public cloud You're handing your crown jewels to someone else For real time control loops.

00:10:08: It has to happen locally at the Edge.

00:10:10: An edge deployment getting astonishingly small now.

00:10:14: Dennis Dombrowski highlighted tiny ML embedding AI directly onto microchips inside individual sensors.

00:10:20: Like for high-speed optical sorting on a recycling line?

00:10:23: Exactly!

00:10:24: Hundreds of items flying past a lens every second, the physics of the network just don't allow for cloud compute there.

00:10:30: by the time server responds that item is gone...

00:10:34: Right and it fundamentally changes battery life too.

00:10:37: If remote sensor in pipeline constantly streaming raw data over cellular its dead in days.

00:10:44: But with tiny ML It runs the model locally, it just watches data and only transmits a tiny one-bite alert if it detects problem.

00:10:53: That shifts battery life from days to years...and ensures autonomy even of network drops!

00:10:59: But scaling that autonomy back up into enterprise software level?

00:11:04: That's where gets hard…

00:11:05: It is but Holger Wilde shared some really promising stuff about Siemens Envoy platform – an agentic AI foundation built on AWS.

00:11:15: So, they aren't custom coding every single agent?

00:11:18: No.

00:11:19: They used a shared foundation and because of that... ...they scaled to nine distinct functional agents across different teams in just one week.

00:11:26: His main point was demos.

00:11:28: don't scale Foundations do.

00:11:30: That makes total sense.

00:11:32: Demos are easy in controlled environment But what about putting these models into actual physical robots?

00:11:38: Ah!

00:11:39: Physical AI Right.

00:11:40: David Randall, Andresi and Shri Ella Prolu discuss the new open physical AI tool chain from AWS in NVIDIA.

00:11:46: To close the loop between the virtual simulation... ...and the physical hardware?

00:11:51: Exactly because a robotics team will get a synthetic model working perfectly on a simulator but then lose six months trying to make the real robot actually replicated on the floor!

00:12:01: The research is wild.

00:12:02: currently one-in-seven startups are building physical AI drones smart logistic spots

00:12:08: But seventy-four percent of those builders feel the economy is totally unready to absorb what they're making.

00:12:13: Okay, wait let me push back on that.

00:12:16: if one in seven startups are building this stuff and seventy four percent say enterprise isn't ready Are we just funding solutions?

00:12:23: That are wandering around looking for a problem.

00:12:25: I mean it's a fair point but The issue Isn't a lack Of problems?

00:12:28: plant managers have A massive list of tasks.

00:12:31: They want automated right.

00:12:32: the bottleneck Is integration.

00:12:34: A manager doesn't want a smart robot if it means ripping out existing conveyor belts.

00:12:38: It's all about how the robot shakes hands with the factory's nervous system.

00:12:42: And that nervous system is the manufacturing execution system, right?

00:12:46: The MES and the digital twins sitting above them?

00:12:49: Yes But we have to be careful using terms like Digital Twin.

00:12:53: Right now Ricardo Moroni pointed out the industry calls every simple dashboard a digital twin.

00:12:58: Oh yeah Slapping a two-D chart on the screen doesn't make it a twin.

00:13:01: Exactly, Moroni says most organizations don't even need full physics based virtual replica day one.

00:13:08: They just need reliable asset state visualization and simple what if scenarios.

00:13:12: But when an organization does build a true digital thread The benefits are crazy.

00:13:17: Tobias Hess brought at BSH the home appliance manufacturer.

00:13:20: Oh they used the Siemens accelerator portfolio right?

00:13:22: Yeah

00:13:23: connected their entire life cycle from product requirements straight to the floor By breaking down those silos, they got time to market by up to fifty percent.

00:13:31: Cutting physical product launch times in half is just staggering.

00:13:36: Machine builders are using this to validate equipment before ordering the steel.

00:13:39: Bozena Eminin and Mustafa Savasi highlighted Mandeli and Grubb using the Siemens CNC digital twin.

00:13:45: Right!

00:13:46: C-N-C The code that tells the physical drill exactly how to move or cut metal.

00:13:51: And traditionally if you make a mistake testing that code A titanium spindle crashes and costs tens of thousands dollars.

00:13:57: But with a digital twin, they do virtual commissioning.

00:14:01: They prove out the programs in software before The Physical Machine is even built.

00:14:05: Yes measure twice cut once entirely digitized and it's scaling up!

00:14:10: Clary Cho and Brady Bjornsson talked about HD Hyundai & Siemens opening A Global Marine Center.

00:14:15: A

00:14:15: Digital Shipyard?

00:14:17: Yes

00:14:17: simulating production sequencing of massive steel blocks for commercial vessels...the efficiency gains totally reshape profit margins.

00:14:25: It's incredible.

00:14:26: But tracking all that complexity means the MES has to evolve.

00:14:30: Silvio so often Roshio Thompson compared the old MES with a static paper map, it just recorded history.

00:14:36: Yeah but modern MES is like dynamic rerouting GPS app on your phone.

00:14:40: Exactly

00:14:40: It detects bottle neck and actively re-routes automated vehicles into different stations.

00:14:45: It tells you what do next.

00:14:47: based real time physics

00:14:48: The intelligence finally there

00:14:51: We keep circling back exact same hurdle

00:14:53: The human element.

00:14:55: Right.

00:14:56: If the MES reroutes production, will the veteran shift supervisor actually follow a new path?

00:15:01: Or overwrite it because they don't understand algorithm.

00:15:04: And that friction is reality of the shop floor today which leaves us with critical thought to mullover.

00:15:10: if ultimate bottleneck isn't technology anymore but human willingness and trust how you redesign cultural handoffs between your veteran engineers and autonomous ages.

00:15:21: Because if you don't engineer that cultural trust, your multi-billion dollar AI system is just going to be another dashboard everyone ignores.

00:15:28: Exactly well If you enjoyed this episode new episodes drop every two weeks.

00:15:33: also check out our other editions on digital construction and connected tools and equipment.

00:15:38: Thanks for listening And don't forget to subscribe.