Latest / eMotors: Electric Revolution / Mind of the Machine
Transcript
- 0:00Welcome everyone. Let's jump right into today's
- 0:02explainer because we are looking at a massive absolute paradigm
- 0:07shift happening right now in the world of self driving cars.
- 0:10You know, for the better part of a decade we've kind of been
- 0:12stuck in this almost there phase, right?
- 0:15The industry has been chipping away at the problem using one
- 0:17very specific playbook. But recently the script got
- 0:20completely flipped. We are officially moving away
- 0:23from rigid hand coded rules and moving toward machines that are
- 0:26literally learning to understand the physical world all on their
- 0:29own. It is a wildly fascinating
- 0:31evolution and I am absolutely thrilled to unpack it with you
- 0:34today. OK, let's dive into this
- 0:37300,000. Just take a second to really
- 0:40process that number 30. Oh, oh, oh oh.
- 0:43That is the staggering amount of human written lines of C code
- 0:47that Tesla recently just deleted from its selfdriving system.
- 0:50Completely scrapped. For years, engineers spent
- 0:53countless hours trying to brute force the act of driving by
- 0:56meticulously writing rules for literally every possible
- 0:59scenario. They essentially believed
- 1:00driving could be solved like a giant math equation.
- 1:03And now all of that core logic that previously dictated exactly
- 1:06how the car should behave, decades of incredibly complex
- 1:09engineering is just gone. It's been replaced by a
- 1:11completely different approach. This quote from Elon Musk
- 1:14actually sums up the sheer magnitude of this shift
- 1:17perfectly. He notes that in their new
- 1:19system, there is no line of code that explicitly tells the car,
- 1:22hey, there's a roundabout here. And honestly that makes total
- 1:25sense because hard coding if then rules for every single
- 1:28weird, chaotic real world scenario.
- 1:30It's fundamentally impossible. Think about it, a plastic bag
- 1:34blowing across a dark highway at dusk, or a cyclist making
- 1:36totally unpredictable, erratic hand signals while trying to
- 1:39balance a cup of coffee. The real world is just
- 1:42infinitely too messy for traditional programming to
- 1:44anticipate the whole shebang. Which brings us to Section 1,
- 1:48The problem with rules. David versus Goliath.
- 1:52We are witnessing a massive philosophical clash right now
- 1:55between traditional legacy automakers and a completely new,
- 1:59nimble AI first approach that's challenging everything we
- 2:02thought we knew. So there is a stark contrast
- 2:06splitting the industry right now.
- 2:08On one side you have the traditional approach championed
- 2:10by companies like Waymo. They rely heavily on explicit
- 2:12code, extremely expensive Lidar sensors and these centimeter
- 2:16accurate high definition maps. But then on the flip side,
- 2:19you've got the neural network approach used by companies like
- 2:21Tesla. This relies on a vision only,
- 2:23end to end AI system. Instead of being programmed, the
- 2:26car's behavior is learned organically from millions of
- 2:29human examples. It's a fundamental divergent in
- 2:31how even tackle the problem and the limitations of that older
- 2:35traditional rule based approach. Well, they're becoming glaringly
- 2:38obvious. Sure, these systems offer really
- 2:41transparent decision making, but they are incredibly brittle.
- 2:44They fail the absolute moment they encounter a novel
- 2:47situation. Imagine a construction zone
- 2:49where a traffic cone just happens to be sitting upside
- 2:51down on top of a stop sign. A rule based system just doesn't
- 2:55know what to do because no engineer sat down and wrote a
- 2:57rule for that specific absurdity.
- 3:00It completely lacks natural emergent behavior, the ability
- 3:03to just figure out a new solution on the fly.
- 3:05Plus relying on HD maps means you need $100,000 sensor suites
- 3:09just to painstakingly pre map cities, making it financially
- 3:12impossible to quickly scale to every random dirt Rd. out there.
- 3:15That leads us right into Section 2, enter the world model giving
- 3:19cars imagination. This is the core academic
- 3:23concept straight out of recent research that's completely
- 3:26revolutionizing how vehicles process their environment.
- 3:29So let's define what we mean by a world model, which is also
- 3:32sometimes called physical AI. Simply put, it's an AI's ability
- 3:37to visually predict how the physical world is going to
- 3:39change based on its actions. It's exactly like human
- 3:43intuition. When you're driving and you see
- 3:45a bouncy ball roll into the street, your brain instantly
- 3:47goes, OK, a kid might be chasing that and you hit the brakes.
- 3:50You intuitively understand the physics and behavioral patterns
- 3:53of the world. A world model is a generative
- 3:56neural system that lets a vehicle literally rehearse these
- 3:58kinds of features in its mind before it ever actually acts in
- 4:02reality. How crazy is that?
- 4:04Building on that incredible foundation, we have VLA, which
- 4:08stands for Vision Language Action.
- 4:10We are moving entirely away from vehicles that just blindly
- 4:13execute lines of code toward a single, unified brain that
- 4:17genuinely understands physical reality and context.
- 4:21And you might be wondering, wait, why is language involved
- 4:23in driving? Well, it's because the system
- 4:25can process prompts and grasp nuanced context far beyond just
- 4:30looking at visual pixels. It can understand a really
- 4:33complex command like pullover safely near the red brick
- 4:36building up ahead, and seamlessly connect that language
- 4:39to the physical environment it's navigating.
- 4:41OK, let's move on and see how this builds into complex
- 4:44decision making, which happens in three remarkably elegant
- 4:47steps. 1st, the system compresses all that overwhelming
- 4:51raw sensor data into a neat, compact mental representation of
- 4:55the world. Second, it imagines the future
- 4:58by rolling out hypothetical scenarios in a virtual sandbox.
- 5:02In literal milliseconds, the car is continuously generating
- 5:05thousands of possible futures, internally testing out different
- 5:08steering angles and braking pressures.
- 5:10And 3rd, it actually plans its trajectory based on the absolute
- 5:14safest of those imagined outcomes.
- 5:16A really fantastic way to conceptualize this whole thing
- 5:19is calling it Chat GT for cars. We all know that Chat GT doesn't
- 5:23actually know facts like an encyclopedia, right?
- 5:26It just predicts the next best word based on massive patterns
- 5:29it's studied in text. Well, this vehicle AI works the
- 5:33exact same way. Just like language models learn
- 5:35to write by studying human language, this AI predicts the
- 5:38next best steering angle or braking pressure based on
- 5:41billions of human driving patterns.
- 5:43It just naturally figures out the unwritten unspoken rules of
- 5:46the road simply by watching humans drive for millions of
- 5:49hours. Let's jump to Section 3,
- 5:52training in the Matrix simulating edge cases.
- 5:56Because the big question is how do you train an AI for the
- 5:59really dangerous stuff when rare edge case crashes just don't
- 6:03happen often enough in real life to provide good training data?
- 6:07And this brilliantly illustrates the sheer scale of what we are
- 6:10dealing with here. 1.5 petabytes.
- 6:13To put that massive number into perspective, 1.5 petabytes is
- 6:16the equivalent of watching high definition movies continuously
- 6:2024 hours a day for over 3 centuries.
- 6:23Tesla's pipeline processes this unbelievable amount of video
- 6:27using 48 different neural networks and it takes 70,000 GPU
- 6:31hours for just a single cycle. The massive, unprecedented
- 6:34computing power required to train these models is truly mind
- 6:37bending, but processing all that data is really only half the
- 6:41battle. The old way of doing virtual
- 6:43simulation was a major bottleneck.
- 6:46You had to rely on traditional graphics engines, where human
- 6:48engineers and 3D artists would literally take months to
- 6:51painstakingly manually build out 3D cities and scenes.
- 6:55If they wanted to simulate a rainstorm, somebody had to go in
- 6:57and code the raindrops. Today, we've entered the era of
- 7:00generative simulators. Here, the AI learns physics
- 7:04directly from video data and builds these massive, complex
- 7:07environments completely automatically in a fraction of
- 7:10the time. Now what's really interesting
- 7:12about this is an insight from Sonia Fiddler, Nvidia's VP of AI
- 7:16Research. With revolutionary projects like
- 7:19ALPA Dreams, developers are basically just learning
- 7:22simulations straight from data. They are creating these
- 7:25interactive real time simulated worlds that aren't just static
- 7:28backgrounds but dynamic environments that adapt
- 7:31instantly and realistically to whatever the AI agent decides to
- 7:34do. The power of these generative
- 7:36simulators is an absolute game changer.
- 7:39They can reconstruct fully interactive 3D scenes from simle
- 7:422D video recordings, they can whip U extreme weather on
- 7:45demand, and they allow for interactive what if testing in
- 7:48100% complete safety and you know this matters directly to
- 7:52you. This is the exact reason why one
- 7:54day you'll be able to sit comfortably in the backseat of a
- 7:56car reading a book while it navigates a massive Blizzard.
- 7:59You will trust it because that AI won't panic.
- 8:02It will have already successfully driven through
- 8:04millions of simulated blizzards before its tires ever touch real
- 8:07snow. Moving to our final part,
- 8:09Section 4 physical AI frontier, the ultimate end game.
- 8:15We need to zoom out here and discuss where all of this is
- 8:17heading and the significant hurdles that are still left to
- 8:20clear before it's everywhere. Because look, the tech isn't
- 8:23perfect yet. Handling unpredictable Black
- 8:26Swan edge cases is still incredibly difficult.
- 8:29A Black Swan event is something totally bizarre, like a massive
- 8:32sinkhole suddenly opening up, or maybe someone in AT Rex costume
- 8:36riding a unicycle down the middle of a highway.
- 8:38Those are inherently hard to train for.
- 8:40Additionally, safety regulators are demanding transparency, and
- 8:44they really struggle to accept the black box nature of AI
- 8:47decisions where you can't explicitly trace the logic.
- 8:50And eventually these models must also learn to integrate touch so
- 8:53they can deeply understand ground dynamics, and audio so
- 8:56they can hear critical things like an approaching emergency
- 8:59siren. So the crucial overarching point
- 9:02is this. We are moving firmly into the
- 9:05era of software 2.0. We are replacing rigid, explicit
- 9:09programming logic with learned behavior.
- 9:11Folks, this isn't just a standard software upgrade.
- 9:14It's a completely different relationship between humans and
- 9:17machines. We're shifting away from
- 9:19dictating every single meticulous instruction.
- 9:22Instead, we are simply teaching a system how to learn.
- 9:25It is a profound shift in how we build and interact with
- 9:28technology. Which leaves us with a truly
- 9:31fascinating question to ponder. When AI completely masters the
- 9:35physical rules of driving, what else is it going to drive?
- 9:39A car? Learning to navigate a busy city
- 9:41is really just step one. Soon, humanoid robots could
- 9:44learn incredibly complex tasks like manufacturing, cooking, or
- 9:48even assisting in surgery just by observing us take a second to
- 9:52let the gravity of that sink in. It is an incredibly exciting
- 9:56shift that will absolutely redefine our relationship with
- 9:58the physical world. Thank you all so much for
- 10:01joining me on this deep dive explainer today.
- 10:03I truly hope you found it useful, that it sparked some
- 10:05curiosity, and that you'll learn something new.
- 10:07Catch you next time.