Nikhil KamathHumanoids Cost as Much as an SUV Now | Nikhil Kamath x Brett Adcock | WTF Online Ep 2
CHAPTERS
- 0:09 – 1:36
Humanoids in your home: trust, safety, and the emotional reaction
The conversation opens with a provocative question about leaving children with a humanoid robot, immediately framing the discussion around safety, trust, and how unsettling (or exciting) humanoids can feel. Nikhil and Brett set up the tension between sci‑fi instincts and real-world engineering limits.
- •Opening question: would you leave kids with a humanoid?
- •Brett’s immediate stance: not safe enough yet for unsupervised interaction
- •Early signal that public fears vs. real limitations will be a recurring theme
- 1:36 – 3:41
Brett Adcock’s origin story: from farm to serial tech founder
Brett walks through his background, early interest in building software businesses, and his first major exit. This chapter establishes his pattern of switching from software to harder hardware problems.
- •Midwest farm upbringing and early tech entrepreneurship
- •Vettery: AI marketplace for recruiting; acquisition by Adecco
- •Motivation to move from software into deep hardware challenges
- 3:41 – 8:39
Electric flying taxis: what’s real, what’s blocked, and when it scales
Nikhil probes how close eVTOL air taxis are to everyday use, especially in India. Brett explains the real constraints: battery energy density and, more importantly, safety certification and regulation.
- •eVTOLs are already flying; certification is the main gating factor
- •Battery energy density vs. fossil fuels remains a fundamental constraint
- •Near-term outlook: paying passenger flights likely within ~5 years
- •Electric aviation expands from short hops to regional to longer routes over time
- 8:39 – 14:52
From Archer to Figure: why humanoids are the ‘meta problem’
Brett explains how building electric aircraft translated into robotics expertise (motors, batteries, sensors, control software). He outlines Figure’s thesis: the world is built for the human form, making humanoids the most general-purpose machine.
- •Aircraft as ‘flying robots’—shared components and systems thinking
- •Humanoid thesis: world infrastructure is human-shaped
- •Figure’s goal: full-stack humanoids doing both home and industrial work
- •Key challenge: enormous action/state space requires neural nets, not hand-coded logic
- 14:52 – 22:06
Inside a humanoid robot: actuators, batteries, sensors, and onboard compute
Nikhil asks for a breakdown of what’s physically inside the robot. Brett describes the core hardware stack—motors/actuators, battery placement, CPU/GPU compute, wiring, torque/force sensing, and camera-based perception.
- •~40 actuated joints; battery in torso; CPU + GPU in torso
- •All compute currently runs onboard for speed and reliability
- •Closed-loop control requires ~200 Hz inference; cloud can’t meet that latency
- •Force/torque feedback across joints enables safe interaction and balance
- •Camera-based perception (no lidar), similar to vision-first self-driving
- 22:06 – 28:34
Efficiency vs. humans: power usage, long-horizon tasks, and “general usefulness”
The discussion shifts from raw hardware to performance metrics: energy efficiency, endurance, and what “general intelligence” means in a home setting. Brett frames success as reliably completing long, messy, multi-step tasks with minimal human intervention.
- •Humans are currently far more energy efficient than robots
- •Path to improvement: both mechanical efficiency and compute/model efficiency
- •Near-term benchmark: completing voice-prompted, multi-room tasks end-to-end
- •Robots must recover from local failures (missing objects, blocked paths, etc.)
- 28:34 – 35:58
The next AI form factor: voice-first interfaces, context, and new devices
Nikhil asks what replaces phones and computers, and Brett argues current devices are “pre-AI” artifacts. Brett predicts voice becomes the dominant UI for AI—paired with new language devices and humanoids that share context and memory.
- •Voice as the natural UI for both digital agents and humanoids
- •Today’s voice models are not ‘A+’ yet (interruptions, turn-taking, tool use)
- •Future devices need persistent context and memory from daily life
- •Thesis: phones/computers will be rebuilt for AI-native interaction
- 35:58 – 38:35
Fleet learning and the data flywheel: robots improving together
Nikhil challenges whether onboard learning is capped by local compute/memory. Brett explains fleet learning: robots collect real-world data, central training updates models, and improvements transfer across tasks in surprising ways.
- •Learning is centralized: collect data from fleet, train on servers, redeploy models
- •Transfer learning: logistics skills can improve unrelated tasks like laundry
- •Goal: robots learn from mistakes with reward signals and trial/error
- •Key scaling advantage: once one robot learns, all robots inherit the skill
- 38:35 – 41:51
Real-world data vs. text/synthetic: why touching the world matters
Nikhil questions whether physical-world data is too sparse compared to internet text. Brett argues physical interaction is a larger dataset in the long run and essential to intelligence; language models provide semantic grounding, but embodiment provides control and manipulation skill.
- •Internet text is finite; physical-world interaction is effectively unbounded
- •Language/VLMs provide semantic understanding (e.g., recognizing a ladder)
- •Robotics needs action/control learning that text models don’t provide
- •Figure trains on human demonstrations (navigation + manipulation) to build ‘Helix’
- 41:51 – 46:59
Kids + humanoids: supervision today, design choices, and ‘Westworld’ temptation
They return to the original safety question with specifics: Brett has tested a robot at home but under monitoring. They debate whether humanoids should look human, friendly, or intentionally tool-like, and explore how far realistic faces/voices could go.
- •Brett wouldn’t allow unsupervised robot-kids interaction yet; safety track record needed
- •Design philosophy: useful work tool, not human imitation or ‘googly eyes’
- •Acknowledgment that human-like appearance/behavior is technically feasible
- •Implications of voice/identity cloning and ‘bringing back’ lost loved ones
- 46:59 – 56:59
Dystopia narratives, AI regulation, and China’s robotics reputation
Nikhil asks whether dystopian messaging is used to create regulatory moats. Brett emphasizes optimism with safeguards, then challenges the claim that China is ahead—arguing the bottleneck is general-purpose autonomy, not manufacturing capacity.
- •Brett’s stance: optimistic future, but non-zero risk requires safeguards
- •Regulation-as-moat theory discussed as an industry dynamic
- •China: strong manufacturing capacity, but autonomy/general-purpose robotics is the hard part
- •Humanoids are closer to consumer electronics manufacturing than automotive complexity
- 56:59 – 1:07:54
Industry landscape and the first real jobs humanoids will do
Nikhil asks who else is impressive in the ecosystem (Tesla, Boston Dynamics, software-only players). Brett critiques teleoperation-heavy demos and describes Figure’s real commercial deployment and the evolving view that home use may arrive sooner than expected.
- •Critique: teleoperated demos are misleading; autonomy is the real benchmark
- •Figure’s BMW deployment: daily work on a production line for ~10-hour shifts
- •Shift in thesis: home is becoming near-term with enough data collection
- •Project ‘Go Big’: building an internet-scale robotics pretraining dataset via real homes
- 1:07:54 – 1:10:38
Why Figure ended the OpenAI partnership and went fully in-house on models
Brett explains the decision to stop working with OpenAI: Figure’s internal team was moving faster on embodied, embedded, on-robot performance. He also notes hiring and perception issues that come with appearing to ‘outsource AI’.
- •OpenAI led Figure’s Series B; collaboration lasted ~a year
- •Decision: Figure believed it was ahead internally on robot-embedded model work
- •Embodied AI requires tight hardware/software iteration—not ‘throw it over the fence’
- •High-velocity evaluation culture: testing many model variants daily
- 1:10:38 – 1:17:24
Investment lens: when AI turns into revenue, and where Brett would bet
Nikhil challenges the gap between AI investment and real revenue. Brett argues AI and robotics will compound over 5–20 years, potentially creating the largest company ever via labor automation; he then discusses what he’d invest in outside his own companies.
- •Long adoption curve: habits and society shift over decades, not quarters
- •Robotics targets the biggest TAM: human labor as a major GDP component
- •Vision: billions of robots; GDP expansion as ‘synthetic labor’ scales
- •Outside pick: strong endorsement of Waymo as an underappreciated generational business
- 1:17:24 – 1:28:41
What builders should do (especially in India): components, data, and execution speed
Nikhil asks what Indian entrepreneurs should build given lower labor costs and less risk capital. Brett advises aiming for big problems, then names concrete gaps: robotics supply chain, components, data collection, and practical hustle to find wedges into the ecosystem.
- •Advice: ‘go for it’—big markets attract capital, talent, and motivation
- •Humanoids need help across compute/training, data collection, and components
- •Supply chain reality: many existing motors/sensors are inadequate; vertical integration is common
- •Tactical guidance: prototypes, outreach, conferences, and persistent iteration
- 1:28:41 – 1:36:19
Society after humanoids: collapsing prices, purpose crisis, and a ‘net gray’ future
They close with the macro question: what happens to jobs, capitalism, and meaning when synthetic agents can do most work. Brett forecasts collapsing costs of goods/services and abundance, paired with a human purpose challenge and a radically different society within decades.
- •Abundance thesis: goods/services costs trend toward near-zero as labor is automated
- •Robot economics: costs fall with scale and with robots building robots
- •Human impact: loss of traditional purpose vs. liberation from daily drudgery
- •Brett’s view: outcomes will be mixed (‘net gray’) but transformative in 20 years
