Best of LinkedIn: Software-Defined & AI-Defined Vehicles CW 33/ 34

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

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.

00:00:21: you can find more info In the description

00:00:22: right.

00:00:23: so automakers Right now are essentially trying to figure out how to program a car in like, single afternoon without accidentally hacking their own brakes.

00:00:34: And today we're really looking at... how they're actually trying to pull that off.

00:00:37: Right, because I mean... That's the ultimate tension in the industry right now.

00:00:41: Totally So.

00:00:42: we are digging into most critical insights that surfaced across LinkedIn over last two weeks specifically cutting through all of the noise around software-defined vehicles and this whole new frontier of AI defined vehicle.

00:00:54: Yeah for any mobility professional listening to this The shift happening at the lab it basically dictates everything on road.

00:01:02: The traditional development cycle is just collapsing, and that completely rewrites the rules for how these machines are engineered in the first place.

00:01:09: Okay so let's unpack this collapsing timeline a bit because you know historically if you were building safety critical automotive software You had very rigid like five-year development cycle.

00:01:21: Oh yeah massively time Right!

00:01:22: You have the luxury of time basically to design test prototype validate all it.

00:01:28: But today that timeline is just immense.

00:01:34: Automakers are, they're racing to launch new architectures.

00:01:37: but you obviously cannot rush safety and security compliance without-

00:01:40: Without catastrophic consequences.

00:01:42: exactly

00:01:43: which

00:01:43: creates this massive bottleneck.

00:01:45: But if we look at the engineering insights from Jai Krishnan Thundafil top teams are completely abandoning those legacy workflows To meet these aggressive deadlines

00:02:01: breaking things when you're dealing with a two ton machine?

00:02:03: Well, decouple the software from physical hardware as early as possible.

00:02:09: He highlighted three major shifts here.

00:02:11: so first teams are testing on virtual ECUs because instead of waiting months for a physical prototype board to be manufactured.

00:02:20: Which

00:02:20: is the traditional hardware-in-the-loop setup, right?

00:02:22: Exactly!

00:02:22: Hardware and the loop are so slow... So they just simulate the silicon entirely in the cloud now Wow Yeah They validate the code on a digital twin of the hardware before the physical chip even exists.

00:02:34: That makes total sense.

00:02:35: You aren't paralyzed waiting on the supply chain or hardware fabrication.

00:02:40: Just see if you're code even compiles correctly.

00:02:43: Right, that speed is absolutely crucial.

00:02:45: and then the second major shift he noted is transitioning the code base to Rust.

00:02:50: Interesting Yeah because older languages like C or C++ Air they leave a lot of room for memory leaks Or buffer overflows

00:02:57: Which hackers just love to exploit.

00:02:59: Oh!

00:03:00: They love it.

00:03:01: But rust fundamentally changes That by using this strict ownership model at compile time It basically mathematically prevents entire classes of memory safety vulnerabilities before the code is even allowed to run.

00:03:14: Wait, really?

00:03:15: Yeah it forces the developer to fix the vulnerability Before the software will even build.

00:03:20: So you aren't spending like Hundreds of hours hunting down a memory leak during some road test because this system literally wouldn't let You create the leak in first place.

00:03:29: Exactly that's real power.

00:03:31: and third PC mentioned baking compliance directly into the daily CICD pipelines.

00:03:37: The continuous integration and continuous deployment pipeline?

00:03:40: Yeah, usually checking for ISO two six to six-two.

00:03:43: you know functional safety standards right.

00:03:46: And ISO two one four three four of the cybersecurity ones.

00:03:48: That's

00:03:49: usually a nightmare

00:03:50: it is.

00:03:50: It's this massive agonizing audit at the very end of a multi year cycle.

00:03:54: but now They're automating those trees checks into the daily build.

00:03:58: Meaning,

00:03:58: this system is just constantly auditing itself every single time an engineer commits a line of code?

00:04:03: Yeah!

00:04:03: I mean that's brilliant way to prevent technical debt.

00:04:06: but here where kind have push back little bit.

00:04:10: Okay lay it on

00:04:10: me Especially when we start talking about injecting AI in CICD workflows To speed things up even more If you apply AI to Code Generation Aren't running risk for scaling up garbage output?

00:04:26: It feels a bit like, I don't know.

00:04:28: giving it very fast highly automated typewriter to bad novelist doesn't just create much higher volume of plausible looking mistakes.

00:04:36: That danger is incredibly real.

00:04:39: Vinicius Taddu Zain actually addressed this head-on in his analysis of open source automotive projects.

00:04:45: Oh, he did?

00:04:45: Yeah He pointed out that.

00:04:46: look AI did not invent bad patches But it absolutely weaponized their speed and scale

00:04:50: because generating the code Is basically free.

00:04:52: now I mean a junior developer can generate a thousand lines Of code in seconds

00:04:56: exactly.

00:04:56: And that completely breaks The review model Because AI makes producing the code In the bug reports dramatically cheaper.

00:05:03: but reviewing them requires deep technical judgment,

00:05:06: right?

00:05:06: You still need a human to read it.

00:05:08: And that is still incredibly expensive in terms of human hours.

00:05:12: so the contributor gets this massive productivity gain but The Human Maintainer just gets crushed under the verification bill.

00:05:19: Wow

00:05:20: result is this flood of plausible looking But functionally flawed code.

00:05:26: So what does the solution then?

00:05:28: if AI is just clogging the pipeline with hallucinations do critical engineering teams Just I don't know ban it completely?

00:05:35: Well, banning the tool just puts you at a massive competitive disadvantage.

00:05:38: Right!

00:05:39: You'd fall behind instantly.

00:05:40: Exactly The solution is building automated quality gates with much higher filtering capacity.

00:05:46: You establish strict machine-readable requirements for tests and you ought to reject any incomplete change without human ever even looking at it.

00:05:55: Oh I see Yeah...you reserve human maintainer attention only for the AI contributions that actually managed to pass those rigorous digital gates.

00:06:03: That heavily disciplined approach perfectly aligns with the GM engineering research that Alok Wary was highlighting.

00:06:09: Oh

00:06:10: right, yeah

00:06:10: Yeah, they looked at how to practically apply AI across vehicle software workflows.

00:06:16: And the conclusion wasn't just to let the AI write everything unchecked.

00:06:20: No because they recognize the limitations of current models.

00:06:23: Their approach bounds the AI

00:06:25: Exactly.

00:06:26: They emphasize keeping AI tasks highly focused Providing very specific vehicle context and ensuring every single output is easily reviewable by a human.

00:06:36: The goal isn't treat AI as replacement for engineering accountability.

00:06:39: It's about surfacing issues earlier so human engineers can spend more time on complex architectural decisions.

00:06:46: Right, you augment the human expertise and don't replace the liability.

00:06:50: but I mean extreme.

00:06:51: end of this.

00:06:51: acceleration is still just wild to consider.

00:06:55: Emily Cho and Malhar Patel highlighted Applied Intuition's Dana which is an AI agent that claims to span entire software defined vehicle life cycle.

00:07:04: The pitch was it allows an engineer essentially program a vehicle within one afternoon

00:07:09: Moving from a written requirement to software architecture, to code generation and review in an afternoon.

00:07:15: I mean that is a staggering leap from five-year cycle!

00:07:18: It's staggering on paper but generating vehicle software in single afternoons.

00:07:26: The physical vehicle hardware isn't designed to handle it.

00:07:29: Right, the software needs a home and legacy homes are broken

00:07:32: Exactly!

00:07:32: So we're seeing this massive bottleneck shift... ...the friction is no longer just about writing code but the platform architecture itself.

00:07:41: Yeah!

00:07:42: Mate Al's Biver pointed that out really clearly handing your software engineers.

00:07:46: the most advanced expensive silicon won't solve anything if you underline E architecture.

00:07:52: The

00:07:53: electrical and electronic architecture,

00:07:54: right?

00:07:54: If that architecture is still siloed it doesn't matter.

00:07:57: It's

00:07:57: like um...it's like renovating an old house.

00:07:59: you can buy the most advanced expensive smart home system on-the market which

00:08:03: Is your new silicon.

00:08:04: yeah

00:08:05: Right but that won't fix bad outdated plumbing.

00:08:08: Which is the silo legacy architecture?

00:08:11: I mean, if you have a hundred different tiny computers scattered around the car each controlling a single window or a single sensor.

00:08:18: You can't push a central software update to all of them?

00:08:21: No,

00:08:21: you can

00:08:22: actually have to knock down walls to get a modern zonal architecture

00:08:26: that spot on The underlying physical foundation has to fundamentally decouple the hardware from the control logic.

00:08:34: right.

00:08:35: So in a zonal architecture, instead of running a dedicated wire from front bumper sensor all the way to central brain you run it into local zone hub and then that hub sends single high speed data stream back to main compute cluster.

00:08:50: And just to prove this isn't theoretical marketing speak.

00:08:53: Thomas Alex shared an incredible proof point on this exact transition.

00:08:57: The BMW stat?

00:08:58: Yes!

00:08:59: BMW literally removed six hundred meters cable by shifting to a zonal architecture.

00:09:05: Wow!

00:09:06: Six hundred meters.

00:09:07: think about the weight savings, the manufacturing simplification and just the raw copper cost.

00:09:12: save there.

00:09:13: it proves this architectural shift is real and measurable.

00:09:16: on the assembly line right now

00:09:17: It's a complete rewiring of The Vehicles Nervous System.

00:09:21: By the way, if you are finding these architectural deep dives valuable just a quick reminder to hit subscribe so that we don't miss our future additions.

00:09:28: because

00:09:34: Right.

00:09:42: This is a perfect example of the new paradigm because BMW isn't just buying individual chips anymore, Qualcomm has providing the entire Snapdragon digital chassis

00:09:52: right?

00:09:52: It's a vertically integrated platform play.

00:09:55: it

00:09:55: Is a massive shift in procurement.

00:09:57: The competitive advantage is no longer about you know who can source the cheapest microcontrollers?

00:10:02: No!

00:10:03: its About securing A scalable centralized computing platform that can actually support AI software updates over the next decade.

00:10:12: They're basically buying a digital foundation for their ADS and cockpit programs.

00:10:17: But then you look at other legacy automakers, and they're taking a very different approach to this consolidation.

00:10:22: I saw a post from Raul Arredondo discussing Honda and Nissan... Oh right!

00:10:28: They are jointly developing a shared SDV operating system which is heavily based on Nissan's underlying technology

00:10:35: Which really highlights the ultimate dilemma for OEMs right now Do you build by or partner?

00:10:40: Exactly It raises a critical question about risk.

00:10:43: Is pooling resources to build a shared baseline OS a strategic necessity because the capital expenditure is just too high to do it alone?

00:10:52: Or, Is It A Massive Risk For Two Legacy OEMs To Build A Solution Together That Perhaps Neither Is Fully Fundamentally Knowledgeable About At A Software First Level.

00:11:14: But, you know while the industry wrestles with platform architecture The marketing departments are already rushing to crown next evolution.

00:11:20: Oh here we go!

00:11:21: The AI-defined vehicle.

00:11:23: The AIDV

00:11:26: Putting AI in front of a vehicle name is trend at this moment.

00:11:29: Of course

00:11:29: it's.

00:11:30: It raises massive questions about safety security and actual business value.

00:11:36: Prasant Gauravahala offered very grounded perspective.

00:11:40: What did he say?

00:11:41: He argued that ADV doesn't replace SDV.

00:11:43: It's not a competing standard.

00:11:46: The AI capabilities emerge directly from those software-defined foundations.

00:11:51: You cannot have an AI defined car if you haven't first defined the hardware with software.

00:11:56: Right, so it is evolution and not replacement.

00:11:59: Exactly.

00:11:59: And Pedro Pacheco issued a pretty stark warning to OEMs chasing this label.

00:12:04: Yeah he did.

00:12:05: He argued that automakers need stop building tech for tech sake.

00:12:10: quote unquote best aid V but maximizing AI's actual business value for the customer.

00:12:17: And to achieve that indispensable value, The AI has to actually understand the physical world

00:12:22: right?

00:12:23: Andre C your deck noted that you know.

00:12:24: setting the cabin temperature via voice prompt is a nice novelty But a true.

00:12:29: in vehicle.

00:12:30: AI must reason about its physical environment.

00:12:34: It has to understand context the weather outside the traffic conditions even the driver stress level.

00:12:40: Oh, wow.

00:12:41: And Robert Moran echoed that pointing out that unlocking this full vehicle context requires centralized compute.

00:12:48: it's not just about throwing more TOPS Terra operations per second at a silo system.

00:12:53: It's about giving the AI access to whole vehicles.

00:12:56: data stream safely Right But see?

00:12:58: This raises huge red flag for me.

00:13:00: Okay If we have an A.I.. That has access to the Full Vehicle Context The Steering The Breaking The Cameras The Cabin Audio How can we possibly secure that?

00:13:10: I mean, the compliance battle seems impossible.

00:13:12: If the AI is connected to the cloud and also connected to breaks what's actually a distinction between safety and security in this new paradigm?

00:13:19: Blurring the line of those two concepts was really costly mistake and Florian Leitner-Fischer framed it perfectly for this era.

00:13:27: Oh!

00:13:27: How does he define that?

00:13:29: Safety deals with random failures like hardware fault or sensor degradation.

00:13:34: maybe software bug.

00:13:35: The system fails because something broke.

00:13:37: It's an engineering problem based on probabilities and redundancy.

00:13:40: So safety is the system breaking on its own?

00:13:43: Correct, but security deals with adaptive intelligent attackers.

00:13:49: The system doesn't just fail.

00:13:50: someone is actively exploiting it.

00:13:52: a safety certificate Is essentially timeless once the physics are proven.

00:13:57: they're proven right.

00:13:58: But a security assessment is just a snapshot because the threat landscape evolves The moment the vehicle leaves,

00:14:05: that makes total sense.

00:14:06: Yeah and if you confuse the two You end up with a car That is formally safe on paper but can be remotely manipulated by a hacker through the infotainment system.

00:14:16: Okay,

00:14:16: that distinction is terrifying but so necessary!

00:14:19: Oh yeah?

00:14:20: And this exact compliance battle managing the full context while separating safety from security it's already creating clear winners and losers.

00:14:27: It really is.

00:14:28: Cindeep Karey shared an incredible scoring matrix evaluating OEMs on SDV & Digital Cockpit Safety & Compliance.

00:14:35: Yeah...this was fascinating.

00:14:37: How are legacy players stacking up against disruptors in space?

00:14:41: because I was pretty surprised.

00:14:42: Right, it's a fascinating reversal of the usual narrative.

00:14:45: Mercedes-Benz is actually leading the pack at fifty seven out of sixty possible points

00:14:50: Which is wild!

00:14:51: Why are they scoring so high when the assumption usually that tech first startups lead this software race?

00:14:57: Well...it

00:14:57: comes down to architecture and legal liability.

00:15:00: Mercedes has certified level three system DrivePilot Meaning.

00:15:05: they actually assume legal liability when the system is engaged.

00:15:09: Wow, okay to do that safely their proprietary MBOS uses hypervisor isolation.

00:15:14: What exactly?

00:15:15: Is there?

00:15:15: so a hypervisor is a software layer That creates virtual walls between different operating systems running on this same exact chip.

00:15:22: Oh

00:15:22: I see.

00:15:23: it guarantees that safety critical elements like breaking are totally separated from the infotainment screen.

00:15:28: if A hacker breaches the Android automotive side They hit a virtual brick wall and just cannot access the drivetrain.

00:15:35: That

00:15:35: is huge!

00:15:36: And where did Tesla land on that matrix?

00:15:37: They lagged significantly at twenty-nine out of sixty.

00:15:40: A massive gap, why do they score so low compared to Mercedes?

00:15:43: Critical

00:15:44: compliance gaps.

00:15:45: They rely heavily in a center screen architecture... ...that essentially presents single point of failure.

00:15:50: Ah, go wrong.

00:15:51: Furthermore, they lack UNR-on-one.

00:15:53: fifty seven certification.

00:15:55: Which is what?

00:15:56: It's the strict international regulatory framework specifically governing automated lane keeping systems.

00:16:02: Right, so on paper The legacy OEM is currently beating the tech first OEM when it comes to the rigorous intersection of architecture and compliance.

00:16:11: That is so interesting.

00:16:13: But you know if we connect this to the bigger picture having the best hypervisor isolation?

00:16:18: The fastest cscd pipelines And the highest compliance scores on paper ultimately serves one single goal

00:16:25: getting autonomous leads actually operating on the street.

00:16:28: Exactly, but yet... The final hurdles.

00:16:30: stopping these vehicles aren't in code or chips and even U.N regulations!

00:16:35: The bottlenecks are real world.

00:16:36: we have to pivot from vehicle itself into messy environment it operates in.

00:16:41: Yeah

00:16:41: you have too

00:16:42: Because a perfect algorithm doesn't mean anything.

00:16:44: if car isn't legally allowed leave parking lot.

00:16:47: Mike Murphy shared eye opening stats that prove the bottle next for autonomous vehicles is no longer technical.

00:16:55: For a decade the hard question was can the car drive itself safely?

00:17:00: Now that's the easy part.

00:17:01: So what is actually stopping?

00:17:03: The rollout?

00:17:03: if the tech is there What's the holdup?

00:17:06: old statutes real estate and the power grid.

00:17:08: wait

00:17:09: real estate.

00:17:10: Yeah, look at zoops for example.

00:17:12: They have a manufacturing plant capable of building ten thousand vehicles year.

00:17:15: okay

00:17:16: But they are currently capped by federal exemption at deploying only twenty five hundred Because of a law written when the fundamental assumption was that every car legally required a steering wheel.

00:17:27: Oh my gosh!

00:17:28: Right, Tesla has ride-heel permits in Phoenix.

00:17:30: they aren't even using yet.

00:17:32: Waymo has meticulously mapped every single street in Washington DC driven three hundred thousand miles there but legally cannot sell a driverless ride to the district.

00:17:42: The technology is literally parked waiting on bureaucracy.

00:17:45: It really is.

00:17:46: The real hurdles are acquiring industrial real estate near city centers, utility interconnection...

00:17:53: Meaning getting the power there?

00:17:55: Yeah!

00:17:55: Getting enough megawatts of power into those parcels to charge a massive fleet simultaneously and navigating local government affairs.

00:18:04: The industry spent billions staffing up with perception engineers to solve the driving problem, but now they desperately need fleet technicians real estate developers and permit expeditors.

00:18:15: And that operational reality was completely backed-up by Ming Ma's post about launching driverless fleet operations in Las Vegas?

00:18:23: Oh yeah totally!

00:18:24: His whole point is success didn't come from a magical AI stack.

00:18:27: it came form the unglamorous

00:18:29: stuff Yeah.

00:18:30: Facility readiness, strict depot safety protocols vehicle logistics and quite literally just keeping the cars clean right.

00:18:38: it required massive cross-functional operational discipline managed by local leadership on the ground.

00:18:43: And if you are a mobility professional listening to this The takeaway is so clear...the next decade of autonomous scaling won't be one by software engineers writing better neural nets, it's going to be won by government affairs teams and supply chain logistics managing the physical movement of atoms.

00:19:02: So let me ask you this assuming we do clear these operational bottlenecks?

00:19:06: The permits are signed ,the depots are built .The power grids are upgraded?

00:19:10: What is the ultimate end game for our physical spaces ?

00:19:14: When the fleets are truly unleashed at scale what happens to the world outside the car?

00:19:20: Sam Schenck painted a brilliant vision.

00:19:23: He argues that the biggest impact of self-driving cars won't actually have anything to do with the act of driving.

00:19:29: Really?

00:19:30: Then

00:19:30: what?!

00:19:30: It will be about fundamentally redesigning the physical world around

00:19:33: us.

00:19:34: How does this change if the car drives themselves?

00:19:37: Think about infrastructure.

00:19:38: we've built entirely to store stationary vehicles.

00:19:41: Our cities are defined by them.

00:19:43: True,

00:19:43: parking lots everywhere

00:19:44: Exactly!

00:19:46: We'll see a drastic reduction in street parking and traditional curbs.

00:19:51: We won't need massive parking garages or sprawling surface lots in valuable city centers because the vehicle just drops you off and leaves to serve someone else.

00:19:58: Oh wow!

00:19:59: we can reclaim all that frontage, we could see streets become significantly narrower Because autonomous cars drive with mathematical precision leading to pedestrian friendly shared green spaces.

00:20:11: That's a beautiful vision.

00:20:13: Yeah The automobile reshaped cities for the worst a century ago and autonomy will drive an equally profound redesign to reclaim our urban footprint.

00:20:22: Reclaiming

00:20:22: this city from the parked car, it is an incredible thought!

00:20:25: Well as we wrap up today's deep dive I want leave you with one final provocative thought inspired by a post from Michael Wells.

00:20:31: Okay We've talked about architecture, AI...and autonomy but he points out a creeping issue.

00:20:37: inside the cabin Digital cockpits are converging on nearly identical layouts.

00:20:42: Oh they ARE

00:20:43: Right.

00:20:43: A large center screen, a standardized row of app icons and voice-activated AI chatbot.

00:20:50: Yeah every car is starting to look the same inside.

00:20:52: Exactly So.

00:20:53: if an AI handles driving And a standardised AI assistant Handles all passenger interaction Every single screen looks exactly the same.

00:21:01: Where will brand identity actually live?

00:21:04: That's great question.

00:21:05: If you took the badge off the steering wheel Of tomorrows software defined vehicle Would even know what car were sitting in.

00:21:12: Probably not

00:21:13: something to mull over as the software continues.

00:21:16: Well, if you enjoyed this episode new episodes drop every two weeks.

00:21:20: also check out our other editions on charging and battery tech commercial fleets and autonomous in mobility markets.

00:21:26: thanks for joining us.

00:21:28: keep questioning The Consensus!

00:21:30: And don't forget to subscribe.

00:21:31: see you next time.