Physical AI: OS, Sim, Models for Safety-Critical Machines
Applied Intuition's founders detail why physical AI for trucks, drones, and mining rigs requires custom OS, fast simulation, and hardware-optimized models—not just smarter LLMs—prioritizing deployment over intelligence.
Physical AI's Unique Demands Beyond Screen-Based LLMs
Qasar Younis and Peter Ludwig emphasize that physical AI diverges sharply from chat or coding LLMs due to safety-critical stakes. While screen AI tolerates errors—like a wrong podcast summary—deploying intelligence on driverless L4 trucks in Japan demands near-perfect reliability. "Learned systems can make mistakes if you’re asking for... something like, 'Tell me about these podcast hosts'... But you can’t do that obviously when you run... driverless trucks," Qasar explains. Physical machines operate in adversarial environments like mining or defense, where failures risk lives and equipment.
This reliability gap drives Applied Intuition's mission: powering cars, trucks, construction, agriculture, and warships with AI for a "safer, more prosperous world." Unlike consumer apps, physical AI must handle real-time control, sensor fusion, and fail-safes. Peter notes vehicles resemble "phones before Android and iOS," fragmented across proprietary OSes lacking unified middleware for AI deployment. Their solution consolidates this into a true OS layer managing schedulers, memory, latency, and OTA updates—critical since "bricking a car" far exceeds bricking an iPad.
Customers span 18 of the top 20 non-Chinese automakers, plus GM, defense firms, and heavy machinery makers. Revenue comes from licensing full stacks or modular tools, enabling OEMs to build in-house while Applied provides the platform.
Evolution from YC Tooling to $15B Physical AI Platform
Starting as YC alums in 2016, Applied bet on unfashionable developer tooling amid VC skepticism that workflows lacked moats. "Doing a tooling company in 2016, 2017 was not... the thing to do... VCs generally... said toolings are just workflows," Qasar recalls. They served robotaxi pioneers with simulation and data infra, evolving through four tech stack overhauls every two years to match AI advances like end-to-end models and transformers.
Today, three core buckets define their 30+ products:
- Simulation & RL Infrastructure: Virtual testing correlates sim-to-real via neural sims for scalable RL. Peter stresses evals shift from deterministic pass/fail to statistical safety ("how many nines" reliability, mean time between failures). No sim perfectly mirrors reality—hydroplaning, construction chaos demand real-world miles—but fast, cheap neural sims enable billions of RL iterations.
- Vehicle OS: Low-level systems for sensor streaming, networking, and updates. Built after market options disappointed, it's now a major business.
- Autonomy Models & World Understanding: Onboard perception/planning for land/air/sea, plus human-machine teaming (voice, fatigue detection as L2++). Multimodal agents let farmers oversee fleets, intervening only on edge cases.
Unlike Scale AI's services focus, Applied remains a tech provider like NVIDIA (sans silicon), with 83% engineers (1,000+ total, 40+ ex-founders). They recruit hardware-software boundary experts, low-level systems hackers, and production ML deployers—curious Michigan-engineer types shunning consumer flash.
Internal AI adoption accelerates this: Cursor and Claude Code top leaderboards for embedded/safety code, creating "bimodal engineers"—those wielding AI outpace peers. Qasar: "AI tools are changing engineering workflows even in embedded systems and safety-critical software."
Hardware Constraints Trump Model Intelligence
The bottleneck isn't smarter models but deploying them onboard constrained hardware. Offboard data-center LLMs balloon in size/speed; onboard needs millisecond latency, low power, tiny footprints via distillation. "The hard part is deploying models onto real hardware, under safety, latency, power, cost, and reliability constraints," Peter asserts.
Legacy autonomy relied on RTK GPS and hand-coded paths for mining/agriculture—reliable but rigid. Modern needs dynamic perception for visual cues, cause-effect (e.g., hydroplaning physics), and planning where actions alter worlds ("plan mode" for multi-step tasks like robotaxis or defense maneuvers). World models aid but falter on rare events; sim-to-real validation persists.
Public trust lessons from Cruise/Waymo: Failures aren't just technical—Cruise's incidents eroded regulator confidence, raising bars. Waymo sets excellence via statistical validation. Peter: "After nearly a decade... we can look at a robotics demo and predict the next 20 problems the company will hit." Demos dazzle but crumble on the brittle last 1%—humanoids, prizes like DARPA ignore production gaps.
Sensors? LiDAR shines for R&D/data but cameras dominate production; Applied supports customer prefs without manufacturing.
Founder Lessons: Survive to Compound
Qasar advises constraining commercial problems early, avoiding mature-firm mimicry: "Compounding technology only matters if you survive long enough to see it compound." 2014 YC stealth/network plays differ from 2026's capital-flooded AI dynamics—new founders face hype cycles.
Hiring targets OS/autonomy/evals/safety experts curious about "how things work," from General Motors Institute lineage. 2-year tech horizons keep them agile.
"Physical AI is not just LLMs on wheels... the future of autonomy may look... like Android for every moving machine," the hosts summarize their vision.
Key Takeaways
- Build physical AI stacks around simulation (for RL scale), OS (for real-time reliability), and distilled onboard models—prioritize deployment constraints over raw intelligence.
- Validate statistically: Target "nines" reliability via sim-to-real correlation; real-world testing never vanishes.
- Bet on tooling despite VC doubt—AI boom vindicates workflows as moats for industrial AI.
- Recruit hardware-software boundary experts and ex-founders for production deployment in adversarial domains.
- For founders: Constrain problems commercially, survive compounding cycles; ignore demo hype, predict the 20 production pitfalls.
- Use AI coding tools like Cursor/Claude even in safety-critical embedded systems to bimodal-ize engineers.
- Human-machine teaming (voice, state awareness) bridges L2++ to full autonomy across ag/mining/defense.
- Fragmented vehicle software needs consolidation like mobile OS did—unify for AI.
- Evolve stacks every 2 years matching research; publish but prioritize applied production.