Best of LinkedIn: Software-Defined & AI-Defined Vehicles CW 29/ 30

00:00:00: Provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts about software-defined AI defined vehicles in calendar weeks twenty nine and thirty.

00:00:09: Frenness supports tier one automotive suppliers with early stage market validation for their R&D efforts by combining secondary research direct OEM expert interviews and facilitated customer meetings.

00:00:20: you can find more info.

00:00:23: So welcome to the deep dive, where we are unpacking the COP discussions for mobility professionals across LinkedIn.

00:00:30: Yeah thanks for having me.

00:00:32: We're really looking at a massive evolution right now.

00:00:36: A single line of compromise code in modern vehicle costs automakers an average of ten point seven million dollars per incident, which is just

00:00:45: wild.

00:00:45: Oh it's staggering and I mean that ten point seven million figure?

00:00:48: That is just the financial fallout right?

00:00:50: It doesn't even touch the brand damage or you know The actual physical safety risks on the road.

00:00:55: the fundamental architecture of the automobile Is essentially being ripped apart and rebuilt from the ground up

00:01:00: its like.

00:01:00: we aren't Even building cars anymore.

00:01:02: We're basically building these highly vulnerable server racks On wheels, right?

00:01:06: Pretty much yeah.

00:01:07: And the people designing these systems are out there trying to build the airplane while they're flying it.

00:01:11: Exactly

00:01:13: Which is our mission for you on this deep dive?

00:01:15: We are cutting up the fluff And we want to look at the raw mechanics of this shift from a software-defined to an AI defined vehicle.

00:01:24: We'll be hitting architecture, cascading safety threats and how OEMs are actually attempting to execute

00:01:32: it.".

00:01:32: The terminology itself is shifting so fast that even the supply chain can barely keep up!

00:01:37: If you trace the chatter back to CES twenty-twenty six, The industry has definitively moved past software defined.

00:01:45: Totally shifted.

00:01:45: right now It's the AI defined vehicle and some practitioners on LinkedIn are even pushing it further to call up the experience Defined Vehicle.

00:01:52: see that sounds kind of like a marketing pivot to me does

00:01:55: sound like yeah

00:01:56: But I want to break down what?

00:01:58: AI defined actually means for an engineering team, you know working on a random Tuesday morning because You can't just slap the larger zonal compute unit into a chassis and suddenly call it in AI vehicle.

00:02:09: No.

00:02:10: There has to be a mechanical shift In how these entire systems are built And all really comes down at this core concept called decoupling Because fundamentally automotive engineering Has always been trapped by silicon timelines.

00:02:24: Historically, you couldn't effectively test your software until the physical hardware like the system on chip was actually manufactured and sitting physically on your test bench.

00:02:34: Which is a massive bottleneck

00:02:36: Exactly!

00:02:36: You were entirely bottlenecked by hardware availability.

00:02:39: but decoupling completely shatters that dependency.

00:02:42: And they are doing this using virtualization, right?

00:02:44: Because the sources mentioned these approaches like EBSivion being deployed to decouple cockpit software development from those physical hardware timelines.

00:02:53: Yeah ebsivion is a great example.

00:02:55: So how does that actually work under The Hood?

00:02:57: Well

00:02:57: think of it as creating a flawless digital twin Of the hardware's architecture.

00:03:03: Instead of waiting for a physical chip From Qualcomm or NVIDIA The engineering teams use A Hardware Abstraction Layer in the cloud.

00:03:10: Oh, so it's all cloud-based simulation.

00:03:13: Right!

00:03:13: It simulates the exact instruction sets The memory constraints Everything of future physical chip.

00:03:19: So software developers can write Compile and aggressively test that whole digital experience in a virtual environment.

00:03:27: That makes total sense.

00:03:28: By time the physical boards are finally minted The software is already completely mature

00:03:32: Exactly!

00:03:33: It speeds up the iteration loop immensely.

00:03:35: OK, so that handles the software side.

00:03:37: But to actually run that software in the real world... The physical nervous system of a car has to change too right?

00:03:44: Because the sources note General Motors is adopting an Ethernet backbone for their zonal architecture and I kind think about it like this We are moving from a traditional CAN bus, which kind of broadcasts messages to every component like you know a localized neighborhood dial-up network.

00:04:00: That is a brilliant way to phrase it.

00:04:01: yeah right

00:04:02: and we're moving too A native ethernet backbone Which acts Way more Like a citywide fiber optic grid routing massive packets Of data instantly.

00:04:12: You have to, I mean the bandwidth requirements for AI inference make that upgrade completely non-negotiable at this point.

00:04:18: But wait you can't just rip out fifty years of CAN bus architecture overnight.

00:04:23: how do you integrate this massive Ethernet backbone when half the car's components are still running on legacy telematics and simple microcontrollers?

00:04:32: Well the secret is you don't actually rip it up...you build a hierarchy.

00:04:35: So classic auto SARS layered architecture remains the baseline for those basic functions.

00:04:41: Oh, I see.

00:04:41: so like your power windows and seat motors?

00:04:43: Yeah they're still using that legacy communication.

00:04:46: They don't need fiber optic speeds to roll down a window.

00:04:48: That makes sense.

00:04:49: But for the massive data pipes Like the programmable perception networks...that's where The new iE-E-Atoe point three D meter standard comes in.

00:04:56: Okay let's break that sander down A bit.

00:04:58: How does ie eight oh two point three d new meter handle the data differently To eliminate the bottlenecks?

00:05:03: So it creates native Ethernet nodes specifically designed for those perception networks.

00:05:09: In a legacy setup, A camera sends raw data to a central gateway which then has to translate and route It.

00:05:15: And

00:05:15: that translation...creates latency

00:05:18: Exactly!

00:05:18: Creates a delay.

00:05:20: But this new standard allows the cameras The radar or the lidar To plug directly into high-speed ethernet grid as native nodes.

00:05:28: Oh wow so they are speaking the exact same packet switched language as the central AI computer right out of the gate.

00:05:34: Precisely,

00:05:34: no translation No bottleneck.

00:05:36: and automakers are actually being scored on this transition Right now through a new AR zero to ar five framework

00:05:43: A or zero two ar-five.

00:05:44: What does that measure?

00:05:45: It grades their electrical and electronic architecture consolidation maturity.

00:05:49: So it's a huge metric right now.

00:05:50: But you know, consolidating all that compute power into a central brain has to introduce a massive physical constraint, right?

00:05:58: Heat!

00:05:59: The sources highlight that thermal management is being completely reframed.

00:06:03: Oh...it really is It no longer just about keeping in combustion engine from overheating?

00:06:07: Right because Engine heat is distributed across the block but AI compute heat is incredibly concentrated in one spot.

00:06:14: Exactly!

00:06:15: When you are running heavy machine learning models on a centralized system-on chip, You generate these intense localized heat fluxes that traditional automotive liquid cooling loops just aren't designed to handle.

00:06:27: And

00:06:27: if the chip gets too hot and thermally throttles what happens?

00:06:31: Well your autonomous driving systems starts losing frame rates

00:06:33: Which is terrifying at seventy miles an hour

00:06:35: Right?

00:06:36: So thermal management now this huge strategic enabler Superior Cooling literally translates to higher autonomous performance.

00:06:44: Wow.

00:06:45: Okay, so you've built this high-speed fiber optic city on wheels and you've cooled the central brain enough to run these massive AI models.

00:06:53: but that creates an entirely new problem.

00:06:56: by networking all of these previously isolated systems together a single broken lock can compromise the entire City right?

00:07:03: Oh absolutely!

00:07:03: The attack surface isn't just a single ECU anymore it is literally all connected.

00:07:15: Yeah,

00:07:15: these advanced networks mean vulnerabilities cascade instantly.

00:07:19: a hacker gaining access through say a simple bluetooth module can pivot across the ethernet backbone and go straight into the breaking controller.

00:07:27: man

00:07:28: that is scary

00:07:29: it is and its forcing complete collision between two historically separate engineering disciplines functional safety and cybersecurity.

00:07:36: Because those used to be entirely separate silos, you had the safety team over here making sure that breaks didn't physically fail And security teams were there so nobody hacked the infotainment screen

00:07:47: Exactly!

00:07:48: Now practitioners on LinkedIn are showing major friction as these two teams are forced to merge.

00:07:55: Where is this friction coming from?

00:07:57: It's largely driven by standards overlap.

00:07:59: People are highlighting this really confusing terminology clash between ISO, two six to sixty-two which governs functional safety and ISO Two one four three for which govern cybersecurity.

00:08:09: Oh I bet that's a nightmare for the paperwork.

00:08:11: It really is.

00:08:12: both frameworks demand rigorous risk assessments But they use similar words to mean very different things.

00:08:18: it creates massive integration headaches in the engineering base.

00:08:21: Well, let's dig into the functional safety side for a second because there was a post by Aladdin hamdi that provided A really clear breakdown of how iso two six to sixty-two Utilizes acyl classifications to grade risk.

00:08:35: Yes That Was a great Breakdown.

00:08:36: He Explained How it Scales from Acel a Up To acl D Based on Three Factors Severity Of The Failure Exposure to the Driving Situation and Control Ability By the Driver.

00:08:47: Right an acal a Is Pretty Straight Forward.

00:08:50: Think of your Automatic dramatic windshield wipers.

00:08:52: Right, if they fail you just turn them on manually?

00:08:54: Exactly!

00:08:55: If the code crashes it's annoying perhaps slightly hazardous and a heavy downpour.

00:08:59: but...the driver has high controllability that can easily reach out to fix it.

00:09:03: But

00:09:03: then you jump up to ASL-D The most demanding tier And this covers critical systems like electric power steering break by wire and these high speed ADS functions.

00:09:12: Life or death system?

00:09:13: Yeah exactly

00:09:14: So To achieve ACLD compliance for central compute architecture What are the actual mechanics involved?

00:09:21: I mean, you can't just rely on writing bug-free code.

00:09:24: Can you

00:09:24: not even close?

00:09:25: Hmm acld requires hardware level fault detection operating in milliseconds and The primary mechanism for this is lockstep processing

00:09:35: lock step process.

00:09:36: Yeah You physically have two CPU cores running the exact same mathematical operation simultaneously And there is a checker circuit that compares their outputs every few nanoseconds.

00:09:45: Every few nanos seconds?

00:09:46: That is incredibly fast!

00:09:48: It has to be, if a cosmic ray flips a bit or software glitch causes microsecond discrepancy the system instantly detects fault and triggers a safe state protocol.

00:09:57: Like safely pulling car over

00:09:59: Exactly.

00:10:00: But it demands massive expensive hardware redundancy

00:10:03: Wow.

00:10:03: And hey, quick interjection.

00:10:05: if you're finding this deep dive into ASIL and network vulnerabilities valuable make sure to subscribe so we catch our future additions.

00:10:11: but returning to the safety aspect there is another standard to consider here too right?

00:10:15: Yes it is also vital to separate that kind of internal hardware failure from SOTIF which is another Standard heavily discussed in The Sources Right Now.

00:10:25: How does SOTIV differ?

00:10:26: form ASIL?

00:10:27: Well ASIL covers internal failures like a broken sensor wire or crashed CPU thread.

00:10:33: But SODIF covers failures... stemming from external real-world performance limitations where the hardware actually works perfectly fine.

00:10:41: Oh, I see!

00:10:41: So it's kind of a difference between your eye muscles physically tearing which would be an ASIL failure and simply being blinded by walking out of dark movie theater into the blazing sun?

00:10:51: That is an excellent way to conceptualize that.

00:10:53: Yeah... That'd

00:10:53: so Tiff….

00:10:54: Your eyes work perfectly fine but this system entirely compromised with its environment itself.

00:10:59: Exactly if the camera was blinded in glare or radar gets confused by chain link fence.

00:11:05: The software didn't crash, but the vehicle still fails to break for a pedestrian.

00:11:09: Right...the outcome is still a disaster!

00:11:11: Exactly and because these systems are now centralized on that Ethernet backbone we talked about.. ...The stakes of both ASIL and SOTA failures are just exponentially higher.

00:11:22: Security researchers have already demonstrated the remote manipulation of steering in braking over compromised CAN bus networks

00:11:29: Which exactly why regulators pay so much attention right now.

00:11:33: The sources mentioned the US Senate Commerce Committee unanimously advancing the Connected Vehicle Security Act to address these software vulnerabilities at a federal level.

00:11:42: It's much needed step, but industry is also trying.

00:11:48: There is currently a feasibility study on Autosar QCL, which is exploring the integration of quantum-safe cryptography.

00:11:55: Wait really?

00:11:55: Quantum safe cryptography in cars!

00:11:57: Yeah they're actively engineering encryption layers today to protect against future quantum computers that might be capable breaking current vehicle key exchanges.

00:12:05: That's wild to think about.

00:12:07: So okay we have designed a quantum resistant ASALD certified supercomputer on wheels.

00:12:12: The theoretical engineering is sound But who actually owns code running it?

00:12:17: Ah,

00:12:18: that is the trillion-dollar question or rather half a trillion dollar question.

00:12:22: Right because whoever owns technology layers captures massive pieces of this automotive software and electronics market…and sources project it to hit five hundred nineteen billion dollars by twenty thirty five.

00:12:34: Half a trillion dollars?

00:12:36: That number is completely rewiring legacy business models as we speak.

00:12:41: but capturing it requires seamless integration which

00:12:44: brings up Anorita Dorrell's post rate.

00:12:46: He mapped out what he calls the five-stage platform journey.

00:12:50: Yes, AnuDural's breakdown is great.

00:12:52: it highlights a really critical reality.

00:12:54: Automakers simply cannot build everything from scratch anymore.

00:12:57: The time to market is just too punishing.

00:12:59: So What are they doing instead?

00:13:01: They're relying heavily on pre-certified off the shelf software features for a dyes and over-the-air update mechanisms just to accelerate from the engineering phase to launch.

00:13:11: Because automotive competitiveness is no longer about who can stamp steal fastest, it's basically a turf war over layer ownership.

00:13:18: but attempting build those operating layers entirely in house has been pretty catastrophic for some legacy players.

00:13:25: I mean look at collapse of automated driving alliance between Karyat & Bosch.

00:13:29: That collapses as textbook example.

00:13:31: I mean, you had billions in losses and severe validation shortfalls.

00:13:36: Oh what wrong there?

00:13:37: Well when you try to merge legacy tier one supplier code which was built for isolated microcontrollers using rigid waterfall engineering with modern agile software development loops meant for central compute the hardware abstraction layers just break down.

00:13:54: so integration complexity basically swallowed project whole.

00:13:57: exactly.

00:13:58: So if two giants like that with virtually unlimited resources are stumbling over integration, who is actually managing this platform complexity?

00:14:06: well?

00:14:07: It's playing out very differently across the major OEMs.

00:14:10: If you look at Volkswagen a significant portion of their current cost disadvantage Is being attributed directly to platform and software architecture complexity.

00:14:19: Really so it not factory inefficiency?

00:14:22: No!

00:14:22: The struggle to unify the software layer is physically bogging down there hardware output.

00:14:28: Conversely, you look at BMW and they are taking a totally different approach.

00:14:32: What's BMW doing?

00:14:33: They're building vehicles like the IX-III in the X-V on completely divergent physical hardware platforms.

00:14:39: so one electric, one combustion but running them on an identical software architecture.

00:14:44: Oh wow!

00:14:45: So they write code once push over the air update And both the Electric and Gas SUV get new interface

00:14:51: Right There successfully decoupling digital platform from the physical platform constraints.

00:14:58: That is smart, but for the heavy lift layers that automakers just can't efficiently build themselves like mapping natural language processing deep autonomous driving data sets we are seeing massive concessions to the tech sector.

00:15:10: Oh absolutely!

00:15:12: The TEP giants are effectively colonizing the vehicles experience layer right now.

00:15:15: Yeah

00:15:16: I mean Ford is integrating Apple Maps natively into its next generation EV platforms starting in twenty-twenty seven.

00:15:22: Mercedes-Benz partnered with Google Cloud to power MVOS.

00:15:25: Stellantis joined Mobileye's RM network.

00:15:28: And you have to contrast those legacy tech partnerships With the strategy of Chinese EV challengers like XPENG.

00:15:35: Right, they just unveiled their LO III SUV in Munich.

00:15:38: Yeah

00:15:38: and They aren't trying to capture The European premium segment by Just undercutting on price or battery costs.

00:15:46: They are targeting that market Purely On the strength Of Their Software & Design.

00:15:50: The software itself is the luxury differentiator

00:15:53: which perfectly pivots.

00:15:54: The conversation from boardroom strategy to real world execution because the dividing line between surviving this transition and falling behind Is entirely dependent on over-the-air?

00:16:04: Execution right ta's

00:16:05: where the rubber meets, the road yeah.

00:16:07: And it has historically been a massive advantage for pure play evs like Tesla Over legacy oems like Ford.

00:16:14: but Ford is adapting.

00:16:15: The sources highlighted a recent Ford Pass update that totally revamped remote vehicle controls.

00:16:20: Yeah, those are big deal!

00:16:21: They moved to slide-based interactions and integrated weather aware climate settings.

00:16:26: It sounds super simple but refining the user experience layer is incredibly difficult for a legacy hardware company to pull off.

00:16:33: it Is because that user experience dictates your premium pricing power.

00:16:37: today Consumer surveys consistently show that assisted an autonomous driving capabilities Are replacing traditional luxury metrics.

00:16:45: People don't care about the engine size anymore.

00:16:47: Right,

00:16:47: consumers care way less about engine displacement and more about how much of the tedious driving task that AI can handle for them.

00:16:54: Though you know the economics of total autonomy like the robo-taxi market specifically they are currently burning massive amounts of capital.

00:17:03: The business model isn't even settled yet.

00:17:05: Are charging per fare?

00:17:07: Uber?

00:17:07: Are you selling a subscription, licensing the stack

00:17:10: or do you take on the astronomical capital expense of owning and maintaining the fleet yourself.

00:17:15: The capital required to survive the EV in autonomy transition is just brutal right now.

00:17:21: if you look at the challenger space Lucid's per vehicle losses currently far exceed NIOs.

00:17:27: it's a tough market But let's set the financial burn rate aside for a second, because there is a much deeper issue hiding in the reality of autonomous deployment.

00:17:36: We are putting neural networks in charge of two-ton machines on public roads

00:17:40: Yeah and the streets are messy

00:17:41: Very messy.

00:17:42: What about the regulatory and ethical guardrails?

00:17:45: For all this?

00:17:46: Well The SAE's AI committee Is actively developing standards for safe deployment but There is A ton Of real world friction.

00:17:56: In HTSA, recently started scrutinizing Robotaxi safety protocols after a Waymo vehicle got stuck and actually delayed emergency responders.

00:18:04: Oh

00:18:04: man!

00:18:05: And that edge case failure perfectly highlights one of the most provocative discussions from The Sources I think?

00:18:11: An industry voice brought up the limitations of process compliance frameworks like A-Space.

00:18:16: Yes, this is such a fascinating point!

00:18:18: Let's really dig into it.

00:18:20: Yeah because A-Space basically dictates software process quality right?

00:18:24: It ensures you documented your code You tested and followed rigorous engineering steps.

00:18:28: But this post argued that Process Compliance cannot guarantee ethical engineering intent on its own

00:18:32: Exactly And they cited Dieselgate as ultimate cautionary precedent for this.

00:18:37: Oh That makes perfect sense...

00:18:38: It Is a Profound Point Because during Dieselgate, the engineering process was likely impeccable.

00:18:45: The code was clean...the testing was thorough…the documentation was perfect…process compliance was achieved….yet this system was intentionally designed to deceive emissions-testing!

00:18:58: not the morality of what you built.

00:18:59: And

00:18:59: if auditing explicitly written rule-based code failed to prevent an ethical disaster like that, how do we possibly audit the ethics of a AI defined vehicle?

00:19:08: That is the exact friction point.

00:19:10: the industry's running head first and two machine learning models don't use simple.

00:19:14: If Then rules they use inference.

00:19:16: there are just massive webs of weights and biases.

00:19:19: it's statistical.

00:19:19: yeah.

00:19:20: When an autonomous system makes a split second decision to swerve away from the pedestrian but hits apart car, it isn't executing pre-written ethical script.

00:19:28: It is calculating statistical probability based on its training data.

00:19:32: It's a total black box.

00:19:33: You can have perfectly compliant asbestos development process for your AI model, but if the training data contains inherent biases or reward function prioritizes wrong outcome The car behaves unethically in real world anyway

00:19:47: Exactly!

00:19:47: A process checklist cannot validate moral alignment of neural network

00:19:52: Which is exactly why regulatory bodies are struggling so much.

00:19:56: You can't just open up the code to see if the car will make the right ethical choice.

00:19:59: The transition to AI-defined vehicles means shifting from auditing code, to auditing behavioral inferences

00:20:06: And that is a science we haven't fully mastered yet.

00:20:09: It's a staggering hurdle for industry to clear.

00:20:11: Process compliance just does not equal moral compliance.

00:20:14: It doesn't, and as we look at the macro evolution of all these challenges from the decoupling architectures to ASLD safety to the ethics of inference models We are really witnessing a much larger convergence happening.

00:20:27: A convergence with what?

00:20:29: Well leave your listeners with this thought The automotive industry in the robotics industry Are no longer separate fields.

00:20:35: They're actively converging around the development Of physical AI and humanoid technology.

00:20:40: Oh, because whether you are building a humanoid robot or level four autonomous vehicle.

00:20:45: You're fundamentally trying to solve the exact same foundational problem right?

00:20:49: Precisely

00:20:50: How does machine process sensory data and navigate chaotic physical world autonomously

00:20:56: Exactly?

00:20:58: And as result of this convergence memory and story suppliers Are quietly signing these massive strategic agreements in background.

00:21:06: They're acting as the new kingmakers right now.

00:21:08: Really?

00:21:09: Memory suppliers!

00:21:10: Yeah, because they are only entities capable of meeting this sheer immense compute needs in these intelligent physical AI vehicles.

00:21:18: The Ultimate Battle isn't just about who writes best software anymore.

00:21:21: It's about securing the silicon and memory that makes it possible on the first place.

00:21:26: Wow...it is a fundamental rewrite for what it means to be an automaker.

00:21:31: If you enjoyed this episode, new episodes drop every two weeks.

00:21:34: Also check out our other editions on charging and battery tech.

00:21:37: commercial fleets an autonomous in mobility markets.

00:21:40: Thank You for joining us today.

00:21:41: keep pushing the boundaries up there And don't forget to subscribe.