Best of LinkedIn: Smart Manufacturing CW 07/ 08

00:00:00: This episode is provided by Thomas Allgaier and Frennis, based on the most relevant LinkedIn posts about smart manufacturing in calendar weeks seven and eight.

00:00:08: Frenis is a B-to-B market research company that supports enterprises across the smart manufacturing industry with the market customer and competitive insights they need to navigate dynamic markets and drive customer centric product development.

00:00:22: It is really good to be back.

00:00:23: And I have to say, looking at the sheer volume of insights we pulled from week seven and eight there's a very distinct narrative forming this time around.

00:00:31: Yeah it feels different

00:00:33: Right, we aren't looking at those visionary slogans anymore.

00:00:36: It seems like the industry has collectively decided to stop dreaming about The Factory of the Future and actually start building the architecture for The Factory right now.

00:00:46: I noticed that exactly A few months ago the feeds were just full of buzzwords so you'd see Metaverse Web Three.

00:00:52: manufacturing things have felt very abstract But this week on our deep dive it's all about repeatable architectures.

00:01:00: It really feels like the pilots are finally ending and The C-suite is asking, you know okay How do we actually deploy this across twenty sites without going completely bankrupt?

00:01:11: That is the exact shift.

00:01:12: We're seeing The Conversation move away from innovation theater, you know those isolated little pilot projects that look amazing in a press release but do nothing else...we are moving to actual production environments where uptime is literally the only metric that matters for anyone!

00:01:27: And if you look at the posts we've curated today There are really three pillars holding this up.

00:01:32: You've got executable digital twins, the rise of industrial AI agents and then the actual physical infrastructure required to support all that.

00:01:41: Let's start with that first pillar, um...digital twins.

00:01:44: because I feel like Digital Twin is a term that has been horribly abused by marketing departments for about a decade now.

00:01:50: Oh absolutely.

00:01:52: But the sources this week, particularly from people like Tim Walsh and Pedro Leone.

00:01:56: They're suggesting a hard pivot to what they are calling executable twins And

00:02:00: executable is critical distinction.

00:02:02: here I mean historically digital twin was often just static three D model right?

00:02:06: It's visualization tool Just pretty

00:02:08: picture on screen

00:02:09: Exactly Really useful for showing client What facility might look like but basically operational dead weight once factory actually opened.

00:02:19: But Tim Walsh brought this up in the context of the pharmaceutical industry, specifically looking at clean rooms.

00:02:25: And for you listening who maybe don't work on pharma validating a clean room is incredibly tedious right?

00:02:30: Tidious

00:02:31: and wildly expensive.

00:02:32: Traditionally, validation is this very physical process.

00:02:36: You build the room you turn on HVAC... ...you literally introduce smoke or tracers to see how air moves and measure contamination.

00:02:43: And if you fail well have physically modify the duct work of layout run test all over again.

00:02:50: Tim's point at that industry moving to executable twins where actually model fluid dynamics in contamination control before bend a single piece sheet metal.

00:02:59: So it

00:03:00: was not just a CAD model it's a full physics simulation of the airflow itself.

00:03:05: It is a stress test, you can model worst case scenarios entirely in software like what happens if a filter clogs?

00:03:12: What happens if the room pressure drops?

00:03:15: Tim noted that this significantly lowers the carbon footprint which makes sense because you aren't running massive physical equipment for weeks just to pass regulatory tests

00:03:25: and drastically accelerates product modifications.

00:03:27: right This concept of, you know tests before you build it appeared in heavy manufacturing too this week.

00:03:33: Pedro Leon had a post about CNC machining that really resonated.

00:03:37: he was talking about the shift from testing on the machine to testing Before The Machine

00:03:42: which is just an economic necessity.

00:03:43: at this point I mean high-end multi axis CNC machines are incredibly capital intensive assets.

00:03:49: if You Are Using That Machine To Verify Your NC Code Running

00:03:54: It Slowly like step-by-step to make sure the tool doesn't

00:03:56: crash.

00:03:57: Exactly, if you're doing that then machine is not making parts and If it's Not Making Parts It Is Not Making Money.

00:04:02: Plus

00:04:02: You Risk A Massive Crash!

00:04:04: If The Code Is Wrong... ...You Might Just Destroy a Fifty Thousand Dollar Spindle In A Fraction Of Second Pager's

00:04:08: Argument Is That The Executable Twin Allows You To Simulate That Complete Machine Environment.

00:04:14: The Kinematics Tool Geometry Fixed Your Interference All of It in Virtual Space.

00:04:19: The Goal Is When You Finally Walk Up And Hit Cycle Start On Real Machine you have a hundred percent confidence.

00:04:25: Zero scrap, zero collision, zero downtime.

00:04:29: But I do have to play devil's advocate here for a second.

00:04:31: We heard about simulation four years, why is this suddenly different now?

00:04:35: Steven Spacek had opposed his fifty five analysis where he actually seemed pretty skeptical of the whole current AI powered twin narrative.

00:04:43: Yeah,

00:04:44: Stephen provides very necessary reality check there.

00:04:47: There was terrible tendency.

00:04:48: right now just sprinkle on everything and assume it magically solves the problem.

00:04:53: Stevens argument that successful twins absolutely must be built physics first not algorithms.

00:04:58: If your model doesn't understand the fundamental kinematics and physical limitations of the tool, AI cannot save you.

00:05:04: So can just like feed a neural network?

00:05:06: A bunch of pass sensor data and hope it accurately predicts future.

00:05:10: I mean You CAN but will be incredibly fragile if machine encounters a physical situation It hasn't seen in training data.

00:05:16: The AI just guesses.

00:05:18: Physics doesn't guess.

00:05:19: Yeah, that makes sense.

00:05:20: But stevan actually went further.

00:05:22: he emphasized That the loop must be closed.

00:05:25: a simulation that just sits on A server and never talks to The real machine is basically Just a video game.

00:05:31: To Be a true twin it has to ingest Real-time data

00:05:35: like vibrations temperatures that sort of vibrations

00:05:37: temperature spindle load exactly And It uses that real world Data to continuously correct.

00:05:43: the virtual model

00:05:44: got it.

00:05:45: so the Model might say This cut should take X amount of power, but the real machine comes back and says actually I'm using x plus ten percent.

00:05:54: So The Twin updates itself to reflect that new reality precisely

00:05:58: And this whole concept stales up massively.

00:06:01: Matthias Heinecke highlighted the digital twin composer That was seen at CES.

00:06:05: We aren't just talking about modeling one CNC machine anymore, we're talking the entire factory floor.

00:06:11: He mentioned companies like PepsiCo are using this to simulate entire factory layouts.

00:06:15: That's interesting because usually you only do that kind of heavy layout simulation when your building a brand new green field plant.

00:06:22: And that is exactly the mistake people make.

00:06:25: Matias points out they are using it to find hidden capacity in existing plants.

00:06:30: By simulating the flow of materials and people dynamically, They can find bottlenecks they didn't even know existed.

00:06:37: It's all about squeezing more value Out-of-the physical assets you already own Rather than pouring millions into concrete for a new building

00:06:44: Which actually brings us perfectly To this second major theme Of our deep dive.

00:06:48: If we have these executable twins and all this data flowing from the machines, why aren't we seeing massive productivity gains?

00:06:56: Yeah

00:06:56: that's a great question.

00:06:57: Kudzai Mandatresa shared at the statistic.

00:07:00: it was frankly bit depressing.

00:07:02: total factory productivity has basically been flat or even declining despite billions of dollars spent on IoT automation over last decade.

00:07:10: It

00:07:10: is classic productivity paradox.

00:07:13: We are absolutely drowning in data but starving for actual insights.

00:07:17: Could I argues that the bottleneck isn't data access anymore?

00:07:21: We have connected everything that can be connected.

00:07:23: The bottleneck is decision intelligence, we've spent ten years building slightly better dashboards but a dashboard just gives human more homework to do.

00:07:32: Yeah

00:07:33: here's beautiful graph showing you're losing money today.

00:07:36: good luck figuring out why

00:07:38: Exactly.

00:07:39: Kudzai believes the solution is a hard shift toward industrial AI agents.

00:07:43: And to be clear, these aren't just chat interfaces.

00:07:45: they are an orchestration layer.

00:07:47: They sit right between data and physical action.

00:07:50: But let's get specific for listeners.

00:07:53: What does agent actually do in factory setting?

00:07:56: Michael Finocchiero had great take on this specifically pushing back against the idea that AI manufacturing is just a chat bot.

00:08:03: Yeah, Michael's point is that AI actually wins on operator workflow.

00:08:07: It's the unglamorous stuff.

00:08:08: instead of an operator spending an hour at the end of a shift writing a summary The agent auto generates it based on machine logs.

00:08:15: Oh That's huge

00:08:16: right.

00:08:16: and Instead of relying on tribal knowledge like you know oh kick the Machine here if he makes this weird rattling noise the agent turns that tribal knowledge into a validated, repeatable checklist.

00:08:28: So it's way less about typing hey chat bot please optimize my global supply chain and much more about.

00:08:35: here is the exact step-by-step procedure to fix this specific jam on line three?

00:08:40: Yes

00:08:40: exactly.

00:08:41: but Michael also warned about trust which is key.

00:08:44: you have to rigorously validate these agents.

00:08:46: You absolutely cannot have an AI hallucinating a safety protocol in a factory floor.

00:08:51: Definitely not Which naturally links AI actually entering the physical world.

00:08:56: Robotics!

00:08:57: Oh, this was a massive topic across The Feeds.

00:09:00: we had Roho Garg talking about China's strategy and Oscar to Hongku from AWS Talking About Humanoids.

00:09:06: it really feels like We are on the verge of a robot explosion?

00:09:09: We are but the underlying motivations Are quite different depending On who you look at.

00:09:14: roho garg noted that china has made humanoid robots A top national priority.

00:09:19: They are heavily focused on mass commercialization and what they call embodied intelligence.

00:09:35: That is the main argument.

00:09:37: If you buy a traditional six-axis robot arm today, You have to bolt it to the floor build a safety cage around It and bring the work to the robot?

00:09:44: It is very fixed automation.

00:09:46: Yeah Oscar argues that humanoids break that entire paradigm because they can walk upstairs Step over loose cables And basically navigate The messy unstructured reality of A normal brownfield factory!

00:09:57: You don't adapt your factory To the robot...the robot adapts to Your Factory.

00:10:01: I do have to push back on that narrative a bit though.

00:10:04: A humanoid robot is incredibly complex.

00:10:07: It's heavy, it has lithium batteries its balancing dynamically.

00:10:10: Kelly Nagel who was referencing Siemens Alex Greenberg raised a massive red flag about exactly this

00:10:17: and rightfully so.

00:10:18: You cannot just drop a two hundred pound bipedal robot into a busy factory aisle and just hope for the best.

00:10:25: Kelly emphasized that this requires heavy multidisciplinary planning in simulation.

00:10:29: if a robot ships and falls or it misunderstands environment, drops payload its massive safety hazard which brings us right back to the simulation point we started with.

00:10:39: you need simulate humanoid robots inside digital twin before ever deploy physically.

00:10:45: There is another risk to all this automation that I hadn't really considered until i read Caleb Eastman's post, he was talking about brittleness.

00:10:53: Ah yes!

00:10:55: This probably the most provocative thought of my entire week.

00:10:59: Caleb argues that as we use AI to optimize every single ounce of Slack out-of the manufacturing system, We inadvertently make the entire organization incredibly brittle.

00:11:08: Because slack is also recovery time right?

00:11:10: Right like if you have extra inventory lying around or extra people on shift You can handle a sudden crisis

00:11:15: Exactly!

00:11:16: If the process knowledge lives entirely inside an AI agent Or some black box machine learning model fewer and fewer humans actually understand how The business runs.

00:11:25: at a granular level.

00:11:27: His quote was chilling.

00:11:28: He said,

00:11:29: highly optimized

00:11:30: systems do not fail gradually.

00:11:32: they fail all at once.

00:11:33: That is terrifying because if the AI goes down or the model drifts and starts giving bad instructions And nobody on the floor knows how to manually schedule The production line anymore

00:11:43: then you are completely dead in the water.

00:11:46: You lose what Caleb calls recoverability.

00:11:50: It's a trade-off we really don't discuss enough in this industry.

00:11:53: We are trading long-term resilience for short term efficiency and Caleb is basically asking if we have gone way too far.

00:11:59: That actually leads perfectly into the infrastructure discussion, because you can't run industrial agents or executable twins or humanoid robots If your network infrastructure cannot handle it.

00:12:10: And this week The whole cloud versus edge debate really heated up.

00:12:13: It did Moosarot Hussein dropped a hammer on that one with the phrase physics doesn't care about your Clyde strategy.

00:12:19: I love that line so much but Explain the technical reality behind it for us.

00:12:24: Why can't we just run everything in the cloud like standard IT does?

00:12:27: It is purely a function of latency.

00:12:29: Moosarot broke down really well, around trip to the Cloud.

00:12:33: So sending data from physical sensor through local gateway out to server firm Virginia processing and send command all way back.

00:12:43: that takes time maybe two hundred milliseconds on good day Which

00:12:47: sounds incredibly fast!

00:12:48: To a human

00:12:49: Clicking web page.

00:12:50: yes to a high-speed stamping press with the twenty millisecond cycle time.

00:12:55: It is an absolute eternity, right?

00:12:57: By the time the cloud says stop The part is misaligned...the machine has already stamped ten more bad parts Or worse, it's crashed the die and caused a hundred thousand dollars in damage.

00:13:07: Muserat argument is that for real time control you absolutely must be edge native.

00:13:11: The decision has to happen on machine itself not data center three states away.

00:13:16: But running AI models at edges hard You don't have infinite compute power of cloud sitting on factory floor.

00:13:22: And David Rogers pointed out another huge issue with edge deployments ML Ops.

00:13:26: Yeah, David touched on something that quietly kills a lot of smart manufacturing projects which is the shifting sand of factory data.

00:13:34: Look In a data science lab, data is clean.

00:13:37: In a factory sensors get dirty.

00:13:39: Calibration drifts over time due to temperature or vibration.

00:13:43: if you deploy a pristine AI model to the edge and the underlying sensor data changes slightly Over the next three months The models accuracy just degrades.

00:13:52: So You need A way To manage that Model Remotely?

00:13:54: You Need Mature MLPs Machine Learning Operations.

00:13:57: It's the discipline of continuously monitoring those edge models, retraining them when data inevitably drips and seamlessly redeploying them without stopping production line.

00:14:07: David argues pretty forcefully that with out Mishir MLOPs your Edge AI is just a temporary experiment that is mathematically guaranteed to eventually fail.

00:14:15: Speaking of things failing let us talk about Brownfield for second The older existing factories.

00:14:21: Emory Bro had some very specific advice Because it's easy to say, just use edge AI when you're building a brand new Tesla Gigafactory.

00:14:29: But what if you were running machines from nineteen ninety?

00:14:32: Anne-Marie's insight is so important here.

00:14:35: You don't need to rip and replace.

00:14:37: There's this huge misconception that To do smart manufacturing you have to throw out the old PLCs And buy all new equipment.

00:14:45: She argues for deploying an edge data layer instead.

00:14:48: How does that work?

00:14:49: You basically put a gateway device on top of the old machine, it listens to the legacy analog signals normalizes that data into a modern format and sends where needs go.

00:14:59: she claims this approach leads two four times faster time-to-value than trying actually upgrade the legacy machine controls themselves.

00:15:07: That seems like literally the only viable path for most mid-market manufacturers.

00:15:12: They simply aren't going to scrap a million dollar stamping line.

00:15:14: just get slightly better data.

00:15:15: Exactly,

00:15:16: but once you have that day of flowing where does it go?

00:15:19: This brings us to the MES, The Manufacturing Execution System.

00:15:24: Metton Kaplan and Jeffrey Tarratt both discussed evolution this week.

00:15:28: they are strongly warning companies against building a graveyard.

00:15:32: Oh, we've all seen this.

00:15:33: The factory running on some completely custom piece of software written twenty years ago by a guy named Bob who retired a decade ago.

00:15:41: and now everyone is terrified to touch the

00:15:42: code.".

00:15:43: That's the ultimate trap!

00:15:45: Metton & Geoffrey argue that modern MES like the Plex platform they were discussing must be low-code and highly flexible.

00:15:53: You need to be able configure your system into specific workflows without writing custom codes that lock you in technical debt forever.

00:16:00: It needs to be a standard platform, not a bespoke IT project.

00:16:05: So we've covered the executable twins to plan the factory The AI agents to run it and the edge infrastructure To handle the brutal physics of it.

00:16:13: But there was one massive constraint that just kept popping up in the feed this week.

00:16:21: This is truly the elephant in the room right now.

00:16:23: Sustainability is no longer just a nice-to-have bullet point on a corporate slide deck, it's becoming hard physical constraint for industrial growth.

00:16:32: Chris Sturdeo posted a stark warning about diminishing rate of return on AI specifically due to energy costs.

00:16:38: He mentioned his specific stat.

00:16:41: He did.

00:16:42: An AI query like asking an LLM a complex question uses ten to thirty times more energy than the standard internet search, so when you start deploying hundreds of agents that are constantly querying massive models just Instantly

00:16:59: and we are already seeing the real-world physical impact of that digital demand Robert Little shared an update from ABB.

00:17:06: That connects these dots perfectly.

00:17:08: It really does.

00:17:09: a bb is investing one hundred and ten million dollars in us manufacturing capacity right now.

00:17:13: Why?

00:17:14: because?

00:17:14: The demand for heavy electrical infrastructure Transformers switch gear high voltage circuit breakers, it's just off the charts.

00:17:21: And the primary driver of that demand Is new AI data centers

00:17:25: so they I were trying to use.

00:17:26: optimized manufacturing is driving a massive manufacturing boom for the equipment needed just to power Because the grid is simply too complex for human operators to manage manually.

00:17:55: Exactly, you have solar and wind generation coming on-and off intermittently all day.

00:17:59: You have massive unpredictable spikes in demand from EV charging networks And these huge AI data centers.

00:18:06: Valtteri argues that we absolutely need AI running on the grid.

00:18:10: Just spot anomalies and balance electrical load In real time Which

00:18:14: actually brings us right back To the brittleness argument From Caleb Eastman earlier.

00:18:18: We are building a system where we need AI grid just so that we can have enough power to run the AI agents inside.

00:18:29: We are moving from those visionary slogans to repeatable hardened architectures.

00:18:46: We're finally realizing that digital twins need physics, not just pretty graphics.

00:18:51: we're realizing that AI agents need to handle boring unglamorous workflows and not just chat.

00:18:57: And while the cloud is great for storage, physics demands the edge of control.

00:19:03: An underpinning all it's this growing tension between efficiency.

00:19:08: We are building much faster, smarter systems.

00:19:11: But we have to be incredibly careful not build systems that too brittle to survive a shock.

00:19:15: That is the core challenge for twenty-twenty six Building systems which aren't just smart but truly robust

00:19:21: Absolutely.

00:19:22: Though it does leave you with one final thought to mull over.

00:19:25: If we have AI agents running the factories and need AI agents to run the power grid, what happens when the factory AI starts negotiating directly with the grid AI for power allocation during a shortage completely cutting humans out of resource loop?

00:19:39: That is terrifying level optimization.

00:19:42: It really is!

00:19:44: Well that's a lot to chew on this week so we're going wrap our deep dive up right there.

00:19:49: Indeed

00:19:50: if you enjoyed it new episodes drop every two weeks.

00:19:53: Also check out our other editions on digital construction and digital power tools.

00:19:57: Thanks so much for joining us!

00:19:59: See you next time.