Vogt, Kyle

co-founder and President/CTO of Cruise Automation; co-founder of Justin.tv/Twitch

Kyle Vogt: Cruise Automation

Lex Fridman Podcast · Feb 7, 2019

• Kyle Vogt
• Cruise Automation
• General Motors acquisition
• DARPA Grand Challenge
• Justin.tv and Twitch
• autonomous vehicle retrofit problem
• self-driving car economics
• Silicon Valley vs. automaker culture
• human-level vs. superhuman driving
• startup founder advice

Core ideas

Vogt traces his engineering path from building 200-pound BattleBots as a teenager (learning mechanical limits by deliberately overdriving cheap DC motors well past rated voltage until they failed) through early BASIC programming on Apple II computers, to a formative realization during a mind-numbing 10-hour highway drive from Kansas to a BattleBots competition: that lane-following on a highway was a problem a computer could plausibly solve, unaware at the time that autonomous driving research already dated back decades. He recalls that pre-deep-learning autonomous driving work relied on hand-built heuristics (color-thresholding for yellow lane markers, Hough transforms for line fitting) rather than the learned perception common today, and notes Cruise's own earliest systems used similarly simple heuristics (basic color thresholding for traffic lights) as scaffolding before migrating to learned approaches.

Describing MIT's 2004 DARPA Grand Challenge entry (an underdog, mostly freshman/sophomore team that got a truck drive-by-wire before its sole steering motor failed on qualification day), Vogt argues the original Grand Challenge was one of the most cost-effective uses of government research funding he's seen, credibly serving as the catalyst for the entire modern autonomous vehicle research wave — but argues a similar DARPA-style challenge today would add little value, since the field now has enormous commercial investment and researcher momentum that no longer needs an initial "flywheel" push.

He recounts leaving MIT during junior year to join Justin.tv (originally a reality show about founder Justin Kan, which evolved into a live-streaming platform and eventually Twitch) after meeting its founders during MIT's January "IAP" term, framing the decision as low-risk incrementalism (one month off, then a semester, then never returning) rather than a single dramatic leap. He describes a formative all-night coding crisis before Justin.tv's TechCrunch-driven launch — a live video streaming backpack system relying on spraying video packets across three or four unreliable cellphone data modems and reassembling them with custom error correction, which was still producing static and audio corruption the night before launch — as a "character-defining" moment of solo, high-pressure engineering.

Vogt describes founding Cruise Automation in October 2013 after deliberately setting three criteria for his next venture (having concluded Twitch, while successful, mainly produced entertainment value): the problem's success had to be technology-determined, it had to have a clear positive societal impact, and it had to be a large-scale business, reasoning that meaningful positive impact requires scale. Landing on self-driving cars (the "greatest applied AI problem of our generation," in his framing, and the most fun he'd had in college), he describes Cruise's original strategy as retrofitting existing cars for highway autopilot as a first commercial product, using its revenue to fund research toward full driverless capability — a plan abandoned after about a year once investor interest and the team's own conviction shifted entirely toward pursuing full autonomy directly. He explains why he's since "come full circle" against the retrofit approach specifically: beyond needing to support dozens of different vehicle makes/models to meaningfully cover a market (versus an assumed handful), building safety-critical control (steering, braking) on top of an unowned base vehicle creates intractable liability and validation problems — undetectable interactions with a manufacturer's own updated anti-lock braking or other systems, unclear fault attribution if a base-vehicle failure cascades into the retrofit system — that make rigorous safety validation effectively impossible.

On Cruise's 2016 acquisition by General Motors ($1 billion), Vogt describes a genuine cultural clash between GM's process-and-deadline-driven reward system (where risk-taking is discouraged because a delayed program or halted assembly line burns money quickly) and Silicon Valley's experimentation-driven culture (where most novel ideas are expected to fail), while crediting GM's leadership and rank-and-file workers alike with genuinely embracing rather than fearing the shift toward autonomy — and, after three years of friction, considers the investment in learning to bridge those two engineering cultures one of the combined organization's strongest assets. He pushes back against a Silicon-Valley tendency to dismiss vehicle manufacturing as a "solved," uninteresting problem, arguing the sheer industrial scale and process maturity GM brings (100+ years of assembly, testing, and supply-chain infrastructure) is something Cruise is glad not to have to reinvent, letting it focus resources on the genuinely novel parts of the problem.

On self-driving economics, Vogt identifies vehicle build cost, vehicle operational lifetime (comparing fleet vehicles to airline economics, where maximizing revenue-generating hours matters), and revenue model as the core variables, naming ride-sharing as the most obvious near-term opportunity (citing potential privacy, consistency, and safety advantages over current human-driven services) alongside delivery (parcels, food, groceries) as an adjacent opportunity — while stressing that no monetization is possible industry-wide until a fleet's driving capability matches or exceeds human performance, which is why the whole industry was, in his view, still in a "holding pattern" at the time of the conversation. Asked about letting riders customize an autonomous vehicle's driving "personality" (aggressiveness, assertiveness), Vogt says Cruise would likely allow minor comfort-related adjustments but would not expose any control that meaningfully increases safety risk, and separately notes (in an aside about aggressive human driving) that slamming the accelerator after a green light typically saves only seconds on a full trip despite the strong subjective feeling of going faster — a psychological self-deception he expects would apply equally to any AV "aggressiveness" setting.

Asked about the single hardest unsolved problem in autonomous driving, Vogt resists identifying any one blocking issue, explaining that Cruise had already built the basic functional building blocks (unprotected left turns, navigating around double-parked cars, construction zones) within the first year or two — the real ongoing challenge is a long, unglamorous "grinding" process of continuous incremental improvement against an unusually high benchmark (human drivers' excellent handling of edge cases), requiring systematically categorizing uncomfortable or risky driving events and iteratively reworking the system via combinations of more deep learning, more test coverage, and broader scenario development. He argues there's no principled reason autonomous driving capability should plateau near human-level performance once that baseline is crossed, and that Cruise intends to keep pushing toward performance many multiples better than human driving indefinitely, since no first-principles ceiling exists.

On timeline, Vogt reframes the "when will there be hundreds of thousands of autonomous vehicles" question around what specifically constrains scaling — mainly the industry's consensus strategy of starting in easier operating domains (favorable weather, simpler road conditions) and incrementally relaxing those constraints (e.g., adding capability for heavy rain or snow, which could multiply the addressable market by several times) — estimating fewer than five years to reach that scale, while noting Cruise could already manufacture that many vehicles but couldn't yet productively deploy them given current demand within its initially targeted operating domains.

Reflecting on startup advice from having built two companies each acquired for roughly $1 billion (Twitch and Cruise), Vogt names three recurring factors: genuine personal passion for the core technical problem (sustaining founders through inevitable extreme highs and lows), consistently strong co-founders with high integrity and logical rigor, and sheer persistence — arguing that founders more often fail themselves (giving up, or otherwise self-destructing) than get outcompeted or run out of money. On raising capital, he frames venture funding as appropriate specifically when ambition requires removing capital as the binding constraint on growth speed, and credits Y Combinator's competitive peer environment (surrounding founders with unusually motivated peers) as valuable enough that he did it twice despite already having founded a company once before. Reflecting on regret, Vogt identifies roughly two years spent continuing to work on Justin.tv/Twitch's pivot toward video-game streaming — a direction he personally found unmeaningful, being neither a gamer nor a Twitch user himself — as the closest thing to a genuine mistake, framing it as a lesson about noticing early when work has stopped providing personal meaning, a lesson he now uses directly in recruiting pitches for Cruise.

Excerpts