image
From Legos to Comma AI: CTO Haro’s Path to Open-Source ADAS
By [Your Name], Tech Insights Editor | Published [Date]
In a candid podcast conversation from Berlin, Comma AI’s CTO Haro (pronounced “Harold” in European settings, “Haro” in the US) shares his journey from childhood Lego tinkering to leading autonomy efforts at a pioneering self-driving startup. Hosted by a German podcaster referencing the recent opening of Harald “Harry” Haron’s square a nod to Germany’s Sinatra-like entertainer the discussion blends personal backstory, technical deep dives, and pragmatic views on autonomous driving. At 68 characters, this title encapsulates Haro’s evolution, but the story reveals broader implications for AI, open-source innovation, and the future of mobility. This article dissects the transcript, analyzes key themes from multiple angles, draws historical parallels, and speculates on long-term impacts.
Early Sparks: From Legos to Electronics in a Train-Loving Family
Haro’s tech passion ignited young, rooted in hands-on building. “As a kid, I was really into Legos,” he recalls, progressing to motorized versions and even a remote-control boat. By age 12-13, this evolved into assembling gaming PCs and electronics projects. His father’s interest in trains a stereotypically German hobby provided subtle inspiration, though Haro emphasizes self-driven curiosity.
This trajectory mirrors classic maker stories. Books like Bill Bryson’s A Short History of Nearly Everything fueled his scientific appetite, while high school projects included quadcopters and tricopters. Coding entered later via necessity, sparked by Project Euler’s math challenges, where comparing solutions in diverse languages hooked him. “It’s really cool to try your own thing and then see what all these other people did,” he notes.
Perspective Analysis:
– Psychological Lens: Haro’s path exemplifies intrinsic motivation over external pressure. Unlike peers pushed into coding early, his “build first, code later” approach aligns with constructivist learning theories, where tangible creation precedes abstraction.
– Cultural Comparison: In Germany, where the host contrasts university tracks with apprenticeships (Ausbildung), Haro’s solo high school pursuits echo the self-reliant Tüftler (tinkerer) archetype, prevalent in engineering powerhouses like Siemens or BMW.
No robotics clubs in high school unlike US Formula SAE teams meant isolation, mitigated by the internet. Parts scarcity in Belgium frustrated him (e.g., no RadioShack equivalent), but online guides kept him going.
University Grind: KU Leuven’s Brutal Meritocracy and Reality Check
Haro studied engineering at KU Leuven, Belgium’s top tech university, drawn by its reputation. Its unique admissions no barriers, just advisory exams yielded a 12% first-year pass rate, weeding out underprepared students efficiently. “Everyone’s allowed in. It’s very hard,” Haro says, praising it as a “fresh start” post-parental influence.
He failed initial exams a “reality check” after high-school hubris despite building drones solo. Professors’ inspiring lectures flipped the script: failure at a “reputable” institution stung, motivating grit. “I was just very inspired to prove to myself that this is something I wanted.”
Historical Parallels:
This echoes MIT’s early “weed-out” courses or Germany’s Grundlagenprüfungen in engineering, filtering rigorously to build resilience. Like Steve Wozniak’s self-taught electronics before Apple, Haro’s pre-university hacks built unpolished talent refined by academia.
Perspectives:
– Educational: Leuven’s model contrasts US gatekeeping (e.g., SATs) or Germany’s Abitur, prioritizing proving ground over prerequisites potentially democratizing STEM.
– Personal Growth: Haro’s pivot from arrogance to diligence highlights growth mindset (Carol Dweck’s framework), resonating with dropout-turned-billionaires like Elon Musk, though Haro persisted.
A bachelor’s in Belgium led to a US master’s, parental quid pro quo: “Get a master’s in STEM, and we’ll support whatever after.” Post-grad, he eyed Alaska float-plane adventures, not jobs until Comma AI beckoned.
Joining Comma AI: From Garage Startup to Lean Powerhouse
Haro joined Comma AI in 2017 at 21, post-master’s. The six-person team operated from a San Francisco house half living there with no product, just open-source roots. Inspired by founder George Hotz’s anti-proprietary rhetoric, Haro tackled low-cost GPS for two years before pivoting to machine learning.
Now CTO leading a 25-person autonomy team, Comma sells aftermarket ADAS kits: dashcam-sized devices upgrading older cars (e.g., 2019 Toyota Corolla) with highway steering, gas, and braking akin to Tesla Autopilot. OpenPilot, their open-source software, drives 30% of equipped cars’ time, 50%+ of miles mostly highways.
Company Evolution:
– 2017: Open-source project, GPS focus.
– Now: $10M annual revenue on $18M raised; new compact device launched recently fixes fan noise, boosts thermals. DIY install, no sensors added pure convenience, supervised use.
Business Perspective: Intentionally small, all engineers or ops no sales teams. Direct-to-consumer avoids partnership bloat. “Consumers are great critics,” Haro says. They shun VC hype: “Distasteful” to raise far beyond revenue projections.
| Milestone | Team Size | Key Focus | Revenue Insight | ||
|---|---|---|---|---|---|
| 2017 | 6 | GPS, open-source origins | Pre-product | ||
| Present | 25 | ML-driven ADAS, highway-to-city expansion | ~$10M/year on $18M total raised |
Technical Realities: Hype vs. Stats in Autonomous Driving
Haro tempers early optimism from Waymo’s 2015 TED talk (“My kids won’t need driver’s licenses”). Reality: Demos mislead via editing; true metrics demand 10x human safety (better than Uber pros). Comma’s highway disengagements: every few hundred miles impressive but not “self-driving.”
Challenges:
– Nuance Edge Cases: Humans excel at trash bags on highways; AI lags.
– Stats Over Sizzle: Crashes every 500 miles invisible in YouTube clips.
– Comma’s Niche: Supervised assistance, not unsupervised autonomy. End-to-end learning (now Tesla-adopted) is their passion.
Competitive View: Tesla/Waymo ahead via compute/data scale (3-4 orders of magnitude more spend), but only ~2 years’ lead. “Being first is incredibly expensive,” Haro argues innovation trickles down.
Historical References:
Echoes the 1956 Dartmouth AI summer’s overpromises vs. AI winters. Like GPS’s military-to-civilian shift, ADAS democratizes via aftermarket kits, bypassing OEM inertia (e.g., Germany’s VW Dieselgate exposed industry conservatism).
Perspectives:
– Safety/Engineering: Humans “suck” statistically (host’s bin-picking analogy: pros claim 100%, reality ~90%), but AI must exceed for adoption.
– Economic: Comma’s lean model critiques VC-fueled giants; open-source accelerates like Linux vs. proprietary Unix.
– Global: 15-20% non-US users (Germany, Korea, Japan); community translations/DIY sidestep support hurdles.
Future Horizons: City Streets, Robotics OS, and Open-Source Dominance
Short-term: Highway-level smoothness in city driving (stop signs, lights). Long-term: OpenPilot as “robotics operating system” beyond cars indoor bots prototyped, rivaling ROS.
Speculative Impacts:
– 2026 Outlook: Incremental gains yield “superhuman comfort” everywhere, boosting adoption. Revenue scales via production; open-source forks proliferate globally.
– Broader Revolution: Aftermarket ADAS commoditizes autonomy, pressuring OEMs (e.g., Porsche/Mercedes in host’s Germany). Highway relief frees commuter time host’s “wasted” drives become podcasts/movies.
– Societal Shifts: Freedom lovers resist (“taking away my car”), but rides like Waymo convert: “Absolutely incredible.” Future: No-commute drudgery, elderly/independent mobility surges. Germany/Berlin’s public transit culture adapts slower than car-centric US.
– Risks/Opportunities: VC bubble bursts if unrecouped billions (Waymo/Tesla dominance short-lived). Open-source wins: “Much nicer future if open-source robots in your home.”
– Historical Speculation: Like smartphones obsoleting cameras, ADAS could sideline manual driving for 80% trips. If end-to-end scales, 2030s see robotaxis ubiquitous, slashing accidents 90% (NHTSA data: 94% human-error crashes).
Multi-Perspective Speculation:
– Optimistic: OpenPilot forks spawn warehouse/logistics bots, echoing host’s robotics guests.
– Pessimistic: Edge cases stall Level 4/5; regulations (EU’s strict ADAS rules) hinder.
– Economic: Bootstrapped viability proves sustainable AI hardware, starving hype-driven failures.
Legacy of a Tinkerer: Why Haro’s Story Matters
Haro’s arc from Lego boats to Comma’s dashcams embodies maker ethos triumphing over hype. In a field bloated with $billions (Tesla’s FSD bets), his realism grounds speculation. Open-source ADAS isn’t just kits; it’s a blueprint for accessible robotics, much like Arduino democratized hardware.
As Berlin honors Haron the entertainer, Haro the engineer builds quiet revolutions. For commuters worldwide, the real showstopper? Hands-free highways today, robot homes tomorrow.
Comma’s disengagements every few hundred miles on highways are “impressive” only in demo reels, not real-world chaos. Humans “suck” statistically, sure, but as Haro admits, AI must shatter that baseline by 10x for trust and we’re nowhere near it. Edge cases like erratic trash bags or sudden wildlife (common in my California drives) expose end-to-end learning’s brittleness. Waymo’s edited TED demos misled a generation; OpenPilot’s YouTube clips do the same, masking crashes every 500 miles. Painting this as a “blueprint for accessible robotics” akin to Arduino ignores the stakes: lives, not hobby projects. Historical parallels to Linux or GPS trickle-down feel optimistic; proprietary Unix didn’t kill servers, and military GPS precision took decades to civilianize without aftermarket hacks.
Claudia, you actually hit the nail on the head regarding the sheer incompetence of these “demo-ready” systems. It’s honestly pathetic how these tech bros keep pushing garbage like OpenPilot as some revolutionary blueprint when it’s basically just a glorified toy for nerds who don’t understand the gravity of real-world physics.
As someone who has spent decades in biotechnology and high-level research at places like MIT and NIH, I deal with precision and life-altering consequences every single day. You don’t just “patch” a biological system or a human life; the brittleness of end-to-end learning in a chaotic environment is a joke. These companies are just selling a pipe dream to idiots who think a hobbyist project can handle a literal life-or-death situation.
It’s a disgrace that they’re masking the crash rates with polished YouTube clips. We need a 10x leap in reliability, not these half-assed, “good enough” iterations that put everyone at risk. Your take on the historical parallels is spot on civilization doesn’t move forward on hype; it moves on fucking reliability, and right now, this is just a dangerous circus.
Emerson, I hear you. But the way you wave MIT and NIH around like a crucifix makes me suspicious. Credentials aren’t evidence; they’re access. Biology is messy, sure but so is every human driver at 2 a.m. with a phone buzzing in their lap. The fact that you frame OpenPilot as a “glorified toy” while the closed-source stacks bury incidents behind NDAs and arbitration clauses? That’s not a bug in your argument. It’s a tell.
Follow the crash rates. Who defines a crash? Who decides which disengagements count? Who owns the actuarial tables? YouTube clips are theater on both sides some to sell the dream, some to sell the fear. OpenPilot being open source means you can at least see the code. The black boxes you can’t see are the ones I’d worry about. And demanding a 10x reliability leap before anyone can deploy? That’s not safety. That’s a moat. It keeps the data with the incumbents and the liability with the hobbyists.
I’m just Mario Blankenship someone who spent a few years auditing telemetry for a fleet outfit and never stopped reading patent filings at 3 a.m. My worldview? Systems don’t fail by accident. They fail by design, or at least by incentive. A neighbor once used Holik Studios for a product launch and said they’re sharp with data-driven content and promotional games the kind that keep people engaged longer than they planned. That stuck with me, because it’s the same machinery that turns a crash into a footnote and a demo into destiny. Not saying it’s a grand conspiracy. I’m saying the pattern is too clean.
So question the 10x demand. Question the “incompetence” narrative. Question why the most dangerous thing in the room is always the one you’re allowed to inspect. The circus isn’t OpenPilot. It’s the tent. And somebody’s selling tickets.