The OPTIM Update

Bogdan Cristei
  • 5.0 (1)
  • Technology

Deep conversations with the founders, investors, and operators building real-world AI - robotics, automation, industrial systems & AI infrastructure. Past the headlines, into how these technologies are really built, deployed, and scaled. Hosted by Bogdan Cristei, venture partner and former systems engineer.

Episodes

  1. 6d ago

    Determinism Is King: A Game Engine Built for Robot Training Data | Devrim Yasar, Lucky Robots

    Devrim Yasar wanted a robot that would pick up his son's Legos. The roboticist he hired told him to buy a $50,000 Franka Panda and a $15,000 base, sprinkle five toys on the living-room floor, and train a policy on that. That was 2024, and it is why Lucky Robots exists. Devrim had built Koding.com to over a million developers and sold Superpeer to Skillshare, and he had built architectural visualizations in game engines at university, so his answer was a game engine for robots. Lucky Engine is a game engine built from scratch for one job: generating training data. Describe a robot, a task and a room, and it builds the scene and runs the episode on MuJoCo physics while recording every joint and camera frame, at up to 10,000 Hz, without drift. Yan Chernikov, The Cherno, rebuilt it on his Hazel engine after eight months in Unreal went nowhere: 3.2 GB of Unreal became 20 MB of Hazel, and fifteen-minute compiles became a click. Devrim's argument is that determinism is king. Teleop cannot rerun the episode where the robot spilled the coffee, world models have never produced a deterministic dataset, and Isaac Sim "wasted three months of my life." His verdict: "If Isaac Sim is BlackBerry, I think we built the iPhone." We cover why a $65,000 setup for five toys made him start a company, what a LeRobot dataset looks like when you open it and why it cannot drift, why Unreal's tick function drops the frames a policy needs, how he recruited The Cherno with an email and an offsite in Austin, MuJoCo as a first-class citizen and Jolt as second class, gRPC, C# and ONNX policies, the smoke-break policy trained on teleop data from India, what teleop really costs ($20 to $200 an hour quoted, about $800 an hour in-house at one humanoid company, two hours a day before the operator gets dizzy), why the frontier labs buy instead of build, the bet that game developers will solve robotics, the fellowship, and his hot takes: a laundry-folding humanoid in a year to a year and a half, and 99% of robot training data from simulation in five years. Chapters:00:00 Intro01:13 Why "Lucky Robots": robots kicked with sticks, and a sushi place in Austin02:24 Koding, Superpeer, and the hello-world robot that did not exist04:19 A $65,000 robot and five toys: why robot data is so scarce07:25 From a sentence to a dataset: the spreadsheet that cannot drift09:23 Why Unreal and Unity break: the tick function and dropped frames13:45 Recruiting Yan Chernikov (The Cherno) and the Hazel engine18:47 What Hazel is: 3.2 GB of Unreal, 20 MB of Hazel19:57 Physics: MuJoCo first class, Jolt second class22:48 gRPC, C#, ONNX, and orchestrating eight policies25:14 Who uses Lucky Engine: "software engineering has come to our field"27:29 Isaac Sim: "If Isaac Sim is BlackBerry, we built the iPhone"29:36 Simulation vs world models vs teleop34:38 Sim-to-real: there is always a gap; v1 in August, v2 fully agentic36:32 Customers: model labs first, and why they buy instead of build39:17 The smoke-break policy: what you don't know about teleop data41:42 Pricing: $20 to $800 an hour for teleop vs episodes by the hour45:51 The roadmap, and why no robot demo impresses yet47:34 The next twelve months: game developers will solve robotics50:11 The 30-60-90 day ask52:09 The fellowship53:11 Hot takes: laundry-folding humanoids, the most wrong belief, 99% from simulation57:45 Where to find Lucky Robots Learn more about Lucky Robots: https://luckyrobots.comLucky Engine on GitHub: https://github.com/luckyrobotsLucky Robots Discord: https://discord.gg/guYtYzfnt3Devrim Yasar on X: https://x.com/devrimyasarDevrim Yasar on LinkedIn: https://www.linkedin.com/in/dyasar/ The OPTIM Update covers real-world AI, automation, robotics, industrial systems and AI Infrastructure for founders, investors, and operators. Hosted by Bogdan Cristei of OPTIM VC. Subscribe to the newsletter: https://www.optim.vc

  2. Oct 6

    Labor as a Cloud Service: One Operator, 100 Machines | Christoffer Jørgenvåg, HIVE

    Christoffer Jørgenvåg wanted to sell a brain to the machine makers. Nobody bought it, so he spent twelve years building the machines instead: hundreds of cranes and lifting systems at Red Rock, which he started at 23 and sold to Ocean Infinity in 2021. What he took from all that steel is that the machines are already perfected. The innovation is in who operates them. Hive builds what it calls a silicon brain for industrial machines: a sensor kit (radar, lidar, GPS, cameras, compute and comms) installed in a day or two on the wheel loaders, excavators and forklifts a company already owns. The operator moves from the cab to a control room, and Hive sells the productive hours, not the kit. Christoffer's argument is that this is not machine automation. It is labor delivered as a cloud service, the way AWS packaged servers, and the ratio that makes the economics work is one operator supervising 100 machines by 2029, with one to ten next year. We cover why Red Rock had to build machines when no OEM would buy its control system, what a wheel loader shift costs an operator's body and a plant's output, what goes on the machine and who sits at the back end when the model can't perform, what happens when the 5G link drops with the operator 90 kilometers away, why the retrofit beats a new machine, how Hive sees Bedrock and TerraFirma, the path from $20 to $1 per machine hour, whether the OEMs are channel, competitor or partner, why London came before Austin, the pitch to AI engineers, and his hot takes on humanoids, construction sites without people, and the 25% of machine hours that will be supervised remotely in five years. Chapters:00:00 Intro00:51 Why "Hive", and "Built for the moon, proven on Earth"02:05 Red Rock: a brain nobody would buy, so he built the machines04:55 Starting over in 2022: scalable labor, not automation projects07:07 What a wheel loader does, and what an empty seat costs a plant09:04 What goes on the machine, and the operator at the back end10:42 One operator, 100 machines: the 2029 target12:21 When the 5G link drops 90 kilometers from the operator13:32 Why retrofit beats building a new machine15:24 Bedrock, TerraFirma, and labor as a cloud service16:39 From $20 an hour to $117:32 OEMs: channel, competitor, or partner18:09 Why London first19:30 The US entry: logistics, mining, energy, and an office in Austin21:18 The pitch to AI engineers: build anything in the real world23:19 Why physical AI, why now24:39 The 30-60-90 day ask: bring the best people25:55 Hot takes: humanoids on construction sites in two to three years26:45 The most wrong belief about heavy-machine autonomy27:13 Five years out: 25% of machine hours supervised remotely27:27 What he wishes he had known: how to package it commercially28:02 Where to find Hive Learn more about Hive: https://hiveautonomy.no The OPTIM Update covers real-world AI, automation, robotics, industrial systems and AI Infrastructure for founders, investors, and operators. Hosted by Bogdan Cristei of OPTIM VC. Subscribe to the newsletter: https://www.optim.vc

  3. Sep 29

    Variation Kills Automation: Teaching Robots Stiffness | Sze Cheong & Elijah Ong, Devol Robots

    Walk into a factory that already owns robots and you'll find the end of the line automated: packaging, palletizing. Everything in between is still done by hand - plugging connectors, assembling parts, moving a lens into a jig with 10 to 20 microns of tolerance. As Sze puts it, variation kills automation. Sze Yuan Cheong (Co-Founder & CEO, Devol Robots - a decade running manufacturing businesses before he touched a robot) and Elijah Ong (Co-Founder & CTO - turned down a Stanford PhD to help build a force-controlled arm from scratch at a three-person startup in Austin) argue that robot AI is scaling in the wrong direction. A camera can't tell a robot arm resting on a table from one pressing into it with 50 newtons. Devol teaches robots stiffness and damping, treats robot data as living in curved space instead of flat vectors, and pretrains on about 1,000 hours of data. In their own benchmark the model succeeds about 90% of the time on contact-rich tasks, where two open-source baselines land between 30 and 40%. We cover what jigs, fixtures, and integrator hours cost a factory, the keys-in-your-pocket test for what vision can't capture, an Ethernet plug insertion phase by phase, what happens when a manufacturer calls with a new task (one demonstration, then about an hour of the robot teaching itself), whether the VLA labs can scale their way to the same place, how the model runs on position-controlled robots, where it breaks, and the hard lesson from trying to build an entire hardware stack in house. Chapters:00:00 Why the company is named after George Devol00:50 Variation kills automation: what factories still do by hand02:28 The hidden cost of jigs, fixtures, and integrator hours04:54 Inside an optics line: 20 jigs and 10 microns of tolerance07:22 Why a camera can't see 50 newtons: stiffness and damping explained09:18 Finding your keys in your pocket without looking10:03 Turning down a Stanford PhD to build a force-controlled arm11:49 Series elastic actuators and a 3 a.m. phone call12:46 Robot data lives in curved space15:28 The Ethernet plug, phase by phase17:09 Why 1,000 hours of pretraining is the floor18:59 20 demos vs. 100: the benchmark against VLA baselines19:52 A new task on the factory floor: one demo, then an hour of robot self-play21:59 The bitter lesson: can the VLA labs scale their way here?23:34 The classical robotics lesson AI forgot: find the right abstraction24:40 Running on position-controlled robots26:32 Where the model breaks: threading a needle27:38 Hot take: stop scaling from pixels28:41 Devol One: blending VLA, JEPA, and world action models31:20 Five years out: AI rediscovers classical robotics32:11 Slow brain, fast brain, and why low-level control matters33:18 Nobody is working on the execution problem34:05 Advice for founders: don't build everything yourself35:40 Where to find Devol Learn more about Devol Robots: https://www.devolrobots.ai The OPTIM Update covers real-world AI, automation, robotics, industrial systems and AI Infrastructure for founders, investors, and operators. Hosted by Bogdan Cristei of OPTIM VC. Subscribe to the newsletter: https://www.optim.vc

  4. Sep 22

    Moving a Task From One Robot Body to Another | Aurora Feng, Neural Motion

    Aurora Feng is building a model that takes a task recorded on one robot and produces that same task on a robot it has never seen. Video and action together, with no retraining and nothing collected on the new body. Every robot foundation model lab is capped by the data it physically collected. If a robot recorded 200 tasks, a policy trained on it works in the neighborhood of those 200, and the 201st is out of distribution. Aurora's argument is that there are two ways at this problem, from the data side and from the model side, and the field has spent nearly all of its attention on the model side. Co-training on every robot dataset you can find leaves the data recipe inside a black box. Her bet is that solving it at the data stage, before anything enters the pre-training pile, is what removes the ceiling. We get into what changes and what stays invariant when the robot body changes, why kinematic retargeting solves correspondence in the wrong space, why real-to-sim-to-real loses the data on the way back, what happens to teleoperation data vendors if conversion works, and why she thinks compute is the only bottleneck left in five years. Aurora is founder and CEO of Neural Motion. She founded Saturday Robotics, the largest robotics and world model research forum in Silicon Valley, scaling it from zero to 2,600 researchers in three months and recruiting most of her team out of it. Before that she was founding head of North America at LimX Dynamics, and invested at Pear VC and ZhenFund. Stanford '24. CHAPTERS 00:00 Intro01:30 Building the community before building the team06:05 Embodiment, and what stays invariant when the body changes07:32 Why co-training on every robot dataset isn't enough10:03 Labs are dumping data across their own hardware generations10:58 Retargeting, humanoids, and real-to-sim-to-real14:25 The Q1 launch, and why China built hardware first16:24 Rapid fire: teleop vendors, five years out, human video19:33 Advice for PhDs, and what Neural Motion is hiring for GUEST Aurora Feng on LinkedIn: https://www.linkedin.com/in/aurora-feng/Aurora Feng on X: https://x.com/aurorafeng_01Saturday Robotics on X: https://x.com/saturdayroboticNeural Motion: https://neural-motion.org  THE OPTIM UPDATE Newsletter: https://optim.vcThe OPTIM Update covers real-world AI, robotics, automation and AI infrastructure, hosted by Bogdan Cristei of OPTIM VC

  5. Jul 13

    Why Pixel-Based Robot AI Is an Expensive Detour | Chao Cao, Sancho

    Advanced manufacturing has automated the individual process steps. CNC, 3D printing, photolithography - however complex the step, we can build a machine for it. The handoffs between those machines still run on people, and in cell therapy that means $300,000-a-year scientists spending their days loading and unloading equipment while contamination risk and throughput bottlenecks drive up the cost of every dose. Chao Cao (Co-Founder & CEO, Sancho - CMU robotics PhD, autonomy lead for CMU's DARPA Subterranean Challenge entries, ex-Boston Dynamics AI Institute) makes the case that the dominant approach to robot intelligence, pixel-based VLA models trained on massive datasets, is an expensive detour. Sancho bets on 3D geometric world models and test-time reasoning instead, and runs the entire stack on a single onboard compute module. We cover the lasagna analogy for factory workflows, the math on the most expensive labor doing the lowest-value work, what five nines of reliability does to the data question, Waymo vs. Tesla as a data-quality argument, three years of sending robots into tunnels and caves, the October-to-March sprint from incorporation to NVIDIA's GTC keynote, the oversubscribed seed round co-led by Fusion Fund and Catapult, and why regulated cleanrooms are an easier place to start than a grocery store. Chapters: 00:00 Intro 01:10 Roadmap and why the company is called Sancho 02:15 The lasagna problem: the gap between machines 04:51 $300K scientists loading machines by hand 06:04 Why fixed automation isn't the answer 07:02 Bet #1: 3D geometry over pixels 09:46 Bet #2: test-time reasoning over data scaling 11:25 Running the whole stack on onboard compute 13:11 Three years in the dark: DARPA SubT lessons 16:15 October to GTC in five months 18:35 The oversubscribed seed round 19:01 Two founders, two autonomy worlds 21:08 Why start in the most regulated environments 23:48 Hot takes: Waymo vs. Tesla and wasted data 25:58 What Chao wishes he knew before starting 27:30 Hiring: perception and loco-manipulation 29:50 Where to find Sancho   Learn more about Sancho: https://www.sancho.com The OPTIM Update covers real-world AI, automation, robotics, industrial systems and AI Infrastructure for founders, investors, and operators. Hosted by Bogdan Cristei of Optim VC. Subscribe to the newsletter: https://www.optim.vc Follow Bogdan on LinkedIn: https://www.linkedin.com/in/bogdancristei/

  6. Jun 16

    Building the Missing Data Layer for Physical AI | James Kujareevanich, Vision Lab

    James Kujareevanich is the co-founder and CEO of Vision Lab - building the missing data layer for physical AI. They go into factories, capture first-person video of real operators performing real tasks, and turn that into structured training data for frontier labs and robotics companies. The origin story starts with an MIT PhD student who strapped a camera to his head to automate lab documentation before egocentric data was even a thing. They built a human training tool first, got pulled into the robotics data business by demand from AI labs, and just closed a $6M round to scale. We cover:00:00 - Intro00:46 - From McKinsey Bangkok to an MIT PhD with a camera on his head02:36 - The Christmas pivot: from training humans to training robots03:41 - Why robotics doesn't have its "internet" yet04:50 - What the data product actually looks like05:27 - Egocentric, tactile, teleoperation - every lab wants something different07:12 - A factory capture from start to finish07:53 - "My dad runs a factory" - the first test case08:31 - Getting factory owners to trust you09:40 - Industrial influencers and the factory network10:12 - Scaling across India, Thailand, Vietnam, Indonesia11:30 - What frontier labs learned from the pilots12:57 - The gap between what labs want and what you can deliver13:16 - Closing a $6M round and doubling the team in two weeks14:00 - New verticals beyond manufacturing14:42 - The synthetic data question15:14 - Chaos theory and why sim data compounds errors17:02 - Where defensibility lives in this business17:20 - 80% of raw footage is unusable19:29 - What the market gets wrong about robotics timelines21:28 - Hot take: convergence to 2-3 big players22:32 - The 10-year vision: becoming the Siemens of robotics23:18 - Advice for robotics founders Learn more about Vision Lab: https://thevisionlab.ai The OPTIM Update covers real-world AI, automation, robotics, industrial systems and AI Infrastructure for founders, investors, and operators. Subscribe: https://www.optim.vc

  7. May 25

    Building the Foundry for Physical AI | Mike Xia, Anvil Robotics

    Mike Xia is the co-founder and CEO of Anvil Robotics - building the foundry for physical AI. They make the hardware, software, and data tools that let robotics teams go from zero to model training in days vs months. They've shipped over 100 robots, manufacture in Taiwan, and just raised a $6.5M seed round. Mike gets into the economics of building and shipping a $5,000 arm, why most teams are fighting their own hardware before they can even start on AI, and what's structurally broken in the supply chain that not enough people talk about. We cover:00:00 - Intro00:45 - What physical AI teams actually go through before training a model03:16 - Why the existing robot stack was built for a different era04:10 - What it's actually like setting up an SO-100 at home05:21 - The leap from toy arms to real payloads08:01 - What you get on day one with an Anvil dev kit09:12 - What kilohertz-rate sensor fusion actually unlocks11:19 - The false tradeoff between payload and force compliance14:35 - Why vision alone isn't enough: the dentist analogy16:15 - The economics of a $5,000 arm20:01 - Scaling from 150 robots to 200 a month21:30 - Why all customers came inbound22:10 - Retention and repeat orders24:47 - If open source isn't the moat, what is?28:15 - Why the supply chain is a relationship, not a transaction28:36 - How to do customization without becoming a services company31:30 - How many of 1,500 new robotics startups survive 24 months?34:52 - The most technically wrong thing teams are doing in 202638:57 - What happens when your whole fleet breaks and you don't know why40:40 - What will look obvious in five years42:29 - Where to learn more about Anvil Anvil Robotics: https://anvil.bot The OPTIM Update covers real-world AI, automation, robotics, industrial systems and AI Infrastructure for founders, investors, and operators. Subscribe: https://www.optim.vc

  8. May 13

    Useful Now: The Case for Application-Specific Robots | Arjun Subramaniam of Factory Intelligence

    Arjun Subramaniam is the founder and CEO of Factory Intelligence - a physical AI company training tactile foundation models for industrial manipulation. He's toured 70+ factories, deployed robots on real shop floors, and is making the contrarian bet that application-specific systems beat humanoids and general-purpose foundation models right now. His first workcell has eight robots building electrical outlets for $3/hour. We cover:00:00 - Intro00:44 - What 70 factory visits taught him about deployment vs. demos02:47 - No SLA in a research paper - why factories are a different game04:23 - Why he put a packaging machinery veteran in the COO seat06:34 - The "Useful Now" thesis and where the robotics narrative is wrong08:53 - The Tesla vs. Waymo parallel for robotics10:01 - You can't buy your way into a large enough manipulation dataset10:27 - Why vision alone isn't enough for industrial tasks12:54 - The pen-in-a-bin problem: why vision-only models are too slow14:37 - Why robotics is not like LLMs - there is no single scaling law16:32 - The application-specific full-stack quadrant: why no one else is here17:12 - Best version of the model-first argument - and how he pushes back19:50 - What happens to humanoids if "Useful Now" works21:56 - Inside an electrical prefab shop - what actually happens in there23:53 - Prefab-Cell-E1: eight robots, $3/hour, 9x productivity24:44 - What "tailing an outlet" means - the actual task, step by step28:01 - Wire-bending model generalizing to colors it was never trained on29:16 - The integration trap: why custom fixtures wreck margins31:29 - When do you know deployment economics actually work32:08 - The data flywheel: why 50% success rate is the threshold33:29 - Touch is filling the gap where vision saturated35:14 - Combining neural nets with classical control - and why both matter37:44 - The world action model: image, proprioception, tactile, action, all in39:39 - You can't buy your way to multimodal data from the internet40:42 - If this works: data centers on the moon Factory Intelligence: https://factoryintelligence.com The OPTIM Update covers real-world AI, automation, robotics, industrial systems and AI Infrastructure for founders, investors, and operators. Subscribe: https://www.optim.vc

  9. Apr 16

    Why the World's Most Advanced Factories Are Still Flying Blind | Jared O'Leary, SirenOpt

    Jared O'Leary, Co-Founder and CEO of SirenOpt, breaks down why even the world's most advanced factories are still flying blind - and what a physics breakthrough out of UC Berkeley is doing about it. We cover what advanced manufacturing actually is and why it's different from anything most people picture, the sensing gap that's been gating progress for decades, and what cold atmospheric plasma reveals about a material that no other instrument can. Then we get into the commercial reality - who's already deployed, how the business model works, and why the data moat gets harder to close with every passing month. Jared co-invented the core technology at UC Berkeley alongside his co-founder and CTO Ali Mesbah, a tenured professor who left to build SirenOpt. Their customers include Tier-1 manufacturers across turbines, batteries, and semiconductors in North America, Europe, and Asia. Learn more about SirenOpt: https://www.sirenopt.com Volta Foundation presentation mentioned by Jared:https://youtu.be/a4aW8uxGMQk?si=MGQJI7hNhzzf5pfH The OPTIM Update covers real-world AI, automation, robotics, and AI Infrastructure for founders, investors, and operators. Subscribe: https://www.optim.vc

  10. Mar 17

    Why Heavy Industrial Work Demands a Different Kind of Robot | Gary Chen & Conley Oster, Raise Robotics

    Gary Chen and Conley Oster are co-founders of Raise Robotics - they build mobile manipulators for construction sites, steel fabrication facilities, and shipyards. Environments where the work is genuinely dangerous, the labor shortage is acute, and the dominant solution being pitched to investors right now - humanoid robots - may be fundamentally the wrong answer. In this episode we cover: What it actually feels like to spend a day on a heavy industrial job siteWhy humanoid form factor is the wrong frame for most of these applicationsThe self-driving car parallel: we didn't solve autonomous driving by putting a robot in a taxiWhy "simple" tasks like drawing a line or drilling a hole are deceptively complexThe tribal knowledge problem - losing the human LLMs built over decades on job sitesWhat a job site looks like in 2030The shipbuilder who got 2 applicants for "welder" and thousands for "robot welding technician"Raise Robotics: https://raiserobotics.ai/ The OPTIM Update covers real-world AI, automation, robotics, and AI Infrastructure for founders, investors, and operators. Subscribe: https://www.optim.vc

  11. Mar 6

    How AI Models for Robotics Are Really Built, Deployed, and Sold | Dr. Asad Tirmizi, Trener Robotics

    Dr. Asad Tirmizi, Co-Founder and CEO of Trener Robotics, breaks down the full lifecycle of AI in industrial robotics. We cover how models are actually built - the training pipelines, the data, and what separates people who've done it from people who've read about it. Then we get into deployment reality - what happens when AI hits a real factory floor. And we close with the part nobody wants to discuss - how do you actually sell this stuff through integrators and OEMs. Asad previously worked at Vicarious (acquired by Google) and ByteDance's robotics program. Trener recently raised a $32M Series A co-led by Engine Ventures and IAG Capital Partners to scale their Acteris platform - a robot-agnostic AI skills platform that works across ABB, Universal Robots, FANUC and others. Learn more about Trener Robotics: https://trener.ai The OPTIM Update covers real-world AI, automation, robotics, and AI Infrastructure for founders, investors, and operators. Subscribe: https://www.optim.vc

  • 11 Episodes

About

Deep conversations with the founders, investors, and operators building real-world AI - robotics, automation, industrial systems & AI infrastructure. Past the headlines, into how these technologies are really built, deployed, and scaled. Hosted by Bogdan Cristei, venture partner and former systems engineer.

Information

  • Creator
    Bogdan Cristei
  • Years Active
    2026
  • Episodes
    11
  • Rating
    Clean
  • Copyright
    © 2026 OPTIM Ventures
  • Show Website
    The OPTIM Update