Press play on the question, not just the rounds. A robot arm over a bin of mixed parts can name everything in the bin. It has never done the work the bin is for. This episode asks what teaches it, and what would count as proof that the teaching worked. Two answers, $210M between them. Mecka, $60M Series B announced October 7, 2026, led by Sequoia, pays people to wear body sensors and record everyday tasks, folding laundry, making coffee, repairing vehicles, and turns those recordings into structured human-demonstration data: motion, contact, force, geometry. Its dataset, EgoVerse, is 1,362 hours, 80,000 episodes, 2,087 demonstrators, all company figures. Eka, $150M Series A reported October 20, 2026, at a $1.45B post-money valuation, led by a16z and F-Prime, skips human demonstrations and trains mostly by reinforcement learning inside a physics simulator, with a custom gripper that reports force alongside vision. In a curated demo cell it shows about 1,500 picks per hour of objects up to 1 kilogram, against a human baseline around 600, with success over 95 percent. Its gripper, the company says, costs under $1,000 and is rated for 5 million duty cycles. The episode tests both routes against four criteria, each named once and applied throughout: validity, transfer, scalability, reliability. Validity splits on coverage versus correspondence: Mecka's recordings come from real work but their coverage of a robot's narrow tasks is unshown, while Eka's gripper senses contact in the loop but whether simulated contact corresponds to real contact is unshown. Transfer ends asymmetric: Mecka is missing a documented transaction, Eka is missing a production floor. Scalability ends with two cost curves, both undisclosed: human capture that may not get cheaper, and simulator fidelity that may not hold as tasks harden. Reliability is the thinnest file: one company rating, one unanswered expert question about a full year on an assembly line without breakdown, one unshown shift, no field result from either side. Listen for the verdict the episode actually reaches: unresolved. Neither answer has converging evidence across the four criteria on the current record. What to listen for next: a named customer buying Mecka's data twice, an independent check of recorded hand-pose accuracy, Eka's throughput across a full production shift, and field endurance across sites and conditions. A bigger round is not on that list. A new round changes the price of the bet, not the evidence behind it. Watch: https://www.youtube.com/watch?v=JHNuqkSqMzAX: https://x.com/Forefrontshow_Instagram: https://www.instagram.com/forefrontshow/LinkedIn: https://www.linkedin.com/company/forefrontshowSubstack: https://forefrontshowofficial.substack.com/p/forefront-daily-where-robot-trainingRSS: https://anchor.fm/s/11810b5d0/podcast/rss Sponsor Forefront: https://muse.ai/s/forefront-sponsor-rate-card-si6pxbxpxe8elkTools we use: https://muse.ai/s/forefront-tools-we-use-ui6pxbxpxcxmjqxi/Ti5sxdE86ZzXW6uFgIZKLSni0dP2X7m_Uw20EDMbzUUWork with us: https://muse.ai/s/work-with-aman-advisory-xki6pxbxpxxwxy7xoJob board: https://muse.ai/s/forefront-job-board-xwxp6pxbxpxgo5xrxgInbound form: https://docs.google.com/forms/d/e/1FAIpQLSeyLOAFaUH2ksZd1VkQ94YoyFbKp5CB5RuOt9sLR1YQs-x4TQ/viewformNewsletter: https://forefrontshowofficial.substack.com
