[BidClub_]
Moonshots · · 133 min

Robotic Deep-Dive: Beating China, the Race Against Tesla, & New Robotics Timeline With Brett Adcock

Peter DiamandisDave BlundinBrett Adcock

RoboticsAI & SoftwareCompany BuildingTechnical
YouTube ↗
TL;DR
  • Figure’s home launch hinges on general intelligence that can sustain useful work for hours, every day, for years. Brett Adcock says Figure 3 already has the physical capability for laundry, dishes, groceries, and cleaning; the constraint is reliably combining those skills in unfamiliar homes. He hopes to demonstrate long-horizon work in “sub-six months,” explicitly without a promise, and would be “very shocked” to exit 2027 without the broader capability; consumer sales still require safety, reliability, and an “Apple-style quality product.”
  • Adcock sharply disputes what today’s robotics demonstrations prove, while Peter frames China as Figure’s closest competitor. Adcock describes Chinese dancing systems as motion-capture-driven RL controllers, useful for demonstrating hardware but pursuing different work from autonomous household manipulation. His Tesla assessment is explicitly observational—“I personally haven’t seen much out of Tesla in two or three years”—while he says China’s progress warrants urgency: “We just have to go beat them.”
  • Figure is committing substantial capital to a scaling thesis whose commercial payoff still depends on generalization. Adcock says Helix 2.5 showed improving execution and a log-linear relationship with additional data; “if this trend holds,” more data and compute should make robots progressively more capable. Figure already spends low nine figures annually on compute, and the hosts describe its Nscale agreement as 100,000 GPUs and more than $3 billion; Adcock estimates the fully expanded arrangement could provide “well over 100x” current compute and conditionally anticipates billion-dollar Helix training runs next summer.
  • Figure 4 is positioned as the company’s largest hardware advance, with affordability and dexterity driving the design. After seven hand iterations, Adcock says a gloved Figure 4 hand and a gloved human hand are “almost indistinguishable,” and predicts the new robot will make Figure 3 “look like a caveman.” Home economics require something like a $300–$500 monthly lease; those are target economics, while Peter’s suggested $30,000 purchase price receives no explicit confirmation.
  • Manufacturing scale remains conditional on solving intelligence, despite expansive demand estimates. Adcock puts existing Baku capacity at perhaps more than 40,000 robots annually across four lines, and sees a difficult path to a million-unit annual facility if current intelligence problems are cracked. He expresses confidence in fully automated manufacturing within five years, with nearer timing “foggy,” and sizes a possible world market at roughly 20 billion humanoids across commercial work and personal ownership.
  • Vertical integration gives Figure control over the product while multiplying the work required to deliver it. Adcock’s purchased elbow actuator arrived roughly three times too large and overweight, illustrating why Figure builds actuators, touch systems, and other components internally. The same quality standard extends to initial maintenance, cybersecurity, and physical and semantic safety; he also describes reducing Chinese component exposure after tariffs briefly reached 270%.
  • Hark adds a faster-scaling digital business and a direct challenge to incumbent consumer interfaces. Adcock reports more than 1.2 million signups and almost five million browser sessions roughly a month after Hark Pro’s launch, and says Hark already exceeds Figure on revenue. Its assistant and undisclosed hardware pursue personalized intelligence with persistent memory; his broader thesis is that new AI devices will displace phones and computers, requiring hardware and models to be designed together.
Digest · the substance, structured for research

1. Figure’s campus previews two very different commercial markets

  • Adcock estimates roughly “1,500 or so” robots on campus, alongside approximately 700–800 people; robots overtook humans four or five months earlier. Visitor robots can converse, open doors, use stairs and elevators, and escort guests. The stretch goal is an entire arrival, check-in, tour, and conference-room handoff without needing a human receptionist.

  • Figure is pursuing commercial deployments and homes simultaneously. Adcock describes commercial work as tightly specified—you can practically write each day’s assignment on paper—but unforgiving about throughput and performance: “We mess up, you get fired.” Homes permit slower movement while introducing different appliances, layouts, objects, and continually changing instructions, making generalization the central difficulty.

  • Adcock says Figure 3 can do most things in a home, including folding laundry, handling dishes, unloading groceries, opening appliances, and cleaning. The missing capability is delivering that work end to end for “a dozen-plus hours per day,” seven days a week, for years. He calls both markets intelligence problems, while acknowledging separate integration and safety problems that remain difficult.

2. A home demonstration and a consumer launch have different milestones

  • Adcock’s launch criterion is an “Apple-style quality product”: a robot enters a home, talks, and performs useful work without teleoperation, repeated failures, or stopping because it cannot decide what to do. He invokes Humane and Rabbit as warnings about releasing AI hardware before the experience works. Children and dogs further raise the required standard.

  • Since Helix 2.5, Figure has been testing in roughly 20–100 new California homes at any given time. Adcock separately describes entering 30 rental homes booked that day or the preceding day and completing tasks such as making beds and tidying living rooms. He calls these “fully zero shot”: neither the homes nor their objects had appeared in training.

  • The hosts’ private Blackbox demonstration used an early Helix 3 checkpoint. Adcock calls it a “baby AGI moment,” likening its competence to a two-year-old inside an adult body: it attempts open-ended instructions, succeeds at many things, and struggles with others. He stresses that this first checkpoint was significantly undertrained; larger models were already being trained and evaluated.

  • Adcock hopes, “I obviously can’t promise,” that within “the next sub-six months,” perhaps sooner, a robot can complete fairly substantial tasks over hours in unfamiliar homes. He would be “very shocked” to exit 2027 without the broader picture demonstrated. Selling consumers a product comes afterward, once delivery, safety, reliability, and faults are addressed; he gives no firm retail launch date.

3. A coffee capsule established the neural-network investment thesis

  • Adcock divides humanoid development into four chapters: excellent functional hardware; an AI recipe for long-horizon, full-body work; scaling intelligence; and scaling robot production. Hardware needs human-relevant speed, torque, and range of motion while remaining affordable and manufacturable. His argument is that each chapter enables the next; a capable body alone cannot produce general robotics.

  • The original coffee demonstration arrived when Figure relied heavily on C++, SLAM, object detection, and a conventional autonomy stack. Adcock could not see that approach supporting a billion broadly useful robots. Having Figure 1 place a K-Cup into a Keurig through learned upper-body manipulation offered a concrete test of whether neural networks could provide the missing route.

  • Adcock wrote “$500 million” on a Post-it attached to a coffee cup, telling the team that demonstrating this internally might enable a half-billion-dollar raise. Figure subsequently raised a little under $700 million. What mattered to him was the mechanism: the demonstration showed that “deep learning works for robotics,” opening a path beyond separately coding every interaction.

  • The hosts probe whether the transition eliminated handwritten control entirely. Adcock clarifies that a C++ controller still maintained balance, while a single neural network handled the upper body and hands. Moving from that arrangement to learned full-body work took about another year—a substantial remaining engineering problem, despite the apparent simplicity of walking between stationary tasks.

4. Full-body manipulation requires a different stack from dancing

  • Adcock originally thought locomotion could be separated from useful work: joystick the robot somewhere, then manipulate objects. The team learned that foot placement, knees, hips, hands, and head must be coordinated together. “Your whole body is there just to get your hands and your head to a certain place,” the discussion emphasizes; all those joint trajectories need a unified treatment.

  • When a host suggests Chinese demonstrations are teleoperation or stitched macros, Adcock gives a more specific account: many dancing robots use deep-learning RL controllers trained from humans wearing motion-capture suits. He characterizes the routines as open-loop playback that does not react meaningfully to the surrounding environment, although some robots use lidar to maintain spacing during coordinated performances.

  • Adcock considers those displays potentially impressive hardware demonstrations, while calling the work “completely orthogonal” to general robotics. His deliberately harsh assessment is that the relevant dancing AI stack can be built by a college senior with one robot. Figure’s own progression was from Helix 1’s upper-body learning to Helix 2’s full-body locomotion and manipulation.

5. Useful robots face dynamics that tabletop model demos avoid

  • Adcock criticizes general-model robotics videos that use fixed cameras, simple tasks, abundant external compute, and playback accelerated 5x, 10x, or 20x. Slowing movement sufficiently makes feedback control easier, he argues. Figure’s logistics requirement is closer to a package in under three seconds, with errors limited to a few percentage points over a shift, alongside balance and environmental reactions.

  • All Helix inference currently runs onboard: planning, reasoning, trajectory generation, and torque-related control. Server-side work handles real-time speech and fleet orchestration. Adcock values the consequence for reliability: losing internet should not make a robot stop halfway through laundry. The company continues experimenting with onboard and offboard architectures, so he does not declare the allocation permanently settled.

  • The timing constraint is physical. Adcock estimates good 5G links at perhaps 20–30 milliseconds per leg, approaching 50 milliseconds round trip—only about 20 hertz before inference, image processing, and other work. Figure needs some control on the “millisecond level.” His categorical point is that certain neural networks must remain on the robot because they run too quickly for remote execution.

  • Cloud coordination manages a different timescale: which robot works, recharges, or heads to maintenance when damaged or unhealthy. Adcock imagines a hundred robots constructing a house while their assignments change as construction progresses. Even the visible conveyor line needs robots substituting for one another, making fleet orchestration a separate requirement from each machine’s immediate control.

6. Scaling laws turn data and compute into the next experiment

  • With Helix 2.5, Adcock says Figure observed what it thinks are the first humanoid data scaling laws: adding data produced a log-linear relationship in model improvement, alongside better execution in unfamiliar environments. His condition matters: “If this trend holds,” additional data and compute should keep making robots more useful. He frames chapters three and four around testing and exploiting that relationship.

  • The hosts describe Figure’s Nscale arrangement as 100,000 GPUs and more than $3 billion. Adcock places the “VR 2200s” in the second half of 2027 and roughly estimates “well over 100x” current compute if the arrangement is fully expanded. Figure was also doubling its existing compute imminently, while already spending low nine figures annually before the larger agreement.

  • Adcock initially says Figure has 10x more data waiting than the hosts’ demonstration used, then revises that upward to “maybe a factor of like 15 or so.” Its Index collection effort captures physical information unavailable from ordinary internet video, including hand orientation and forces. Incoming data still needs auditing, distribution checks, cleaning, annotation, and alignment with the next pretraining checkpoint.

  • One large pretraining run consumed more than half of recent compute, with Adcock describing approximately 50–70% of capacity devoted to that work. Asked about billion-dollar runs, he says, “I expect that next summer we’ll be doing billion-dollar training runs on Helix,” if everything goes well. He lists data, compute, capital, and successfully establishing the clusters as necessary conditions.

7. Figure 4 pairs a dexterity step-change with lower-cost ambitions

  • Peter floats a $30,000 purchase price and approximately $300 monthly leasing. Adcock explicitly says home affordability requires something like a $300–$500 monthly lease; he does not confirm the purchase figure. Leasing also suits rapid improvement, allowing customers to receive newer robots every few years. His analogy is that humanoids are still in the “flip-phone era.”

  • Figure 4 could be the “iPhone 1 moment,” Adcock says, because the generation changes much more than appearance. Cost reduction is a leading requirement: lower prices should induce additional demand. He also expects progressively greater dexterity, safer interaction, and a better overall experience, with Figure 4 representing the largest generational advance the company has made.

  • Hands provide the concrete preview. Figure has spent almost two years on the work, built seven iterations, and explored both tendon-driven and non-tendon-driven designs. Adcock argues that “a Figure 4 with a glove and a human with a glove are almost indistinguishable.” Despite calling Figure 3 the best current humanoid, he says its successor will make it “look like a caveman.”

  • Financing is expected to involve banks and funds rather than necessarily Figure’s own balance sheet. Adcock reports substantial interest in financing commercial contracts and sees no reason similar arrangements cannot serve homes. Asked whether a robot insurance industry has emerged, he says Figure has not seen one yet, while expecting both commercial and consumer deployments to need coverage.

8. Off-the-shelf components failed Figure’s integration requirements

  • Adcock says Figure initially hoped to assemble more of the robot through suppliers. Its purchased elbow actuator appeared suitable on torque, encoders, gearbox, and speed, but arrived roughly three times too large and overweight, at an excessive price. For the required economics, he says components cannot cost hundreds of dollars when they need to reach “tens of dollars at most.”

  • Building the replacement creates a multidisciplinary obligation: firmware, electronics, thermal design, embedded software, structures, motors, gearing, packaging, and dynamic cabling. Better performance can enlarge a component until it clashes with another joint or compromises appearance. Adcock recalls actively searching for alternatives before accepting the work: “This really sucks. We have to do it ourselves.”

  • Touch systems follow the same logic. Adcock says externally impressive technology may fail production reliability, cost, thermal, communication, or packaging requirements, while slowing iteration through another organization. Figure therefore develops these systems internally and initially plans to retain maintenance and repair itself. He expects a broader service ecosystem eventually, but wants tight control over the early home experience.

9. Companionship requires memory and expression beyond a black faceplate

  • Adcock acknowledges that the robot’s dark face lacks much of human facial communication. Figure has an HMI team working on the head, face, gestures, nods, and bodily reactions, with further Figure 3 changes planned. He rejects simply adding “googly eyes” to suggest friendliness or safety, calling that a disservice to the complexity and actual capabilities of the machine.

  • His consumer vision combines labor and companionship: a robot knows the household, remembers conversations, and can notice a child returning from school upset. Persistent memory already supports internal experiments in recognizing people and recalling prior exchanges. Adcock also describes the mundane other half of the proposition—unloading his own dishwasher that morning and wondering why his robot was not doing it.

  • Memory and preferences “definitely” do not belong permanently to a particular body. They should transfer between devices, and Figure is experimenting with a conversation following someone across successive campus robots. The hosts find that shared continuity both striking and potentially uncanny. Physical and cognitive personalization are also planned through apparel, accessories, voice, and personality choices.

10. Safety combines physical reliability, judgment, and cybersecurity

  • Adcock separates physical safety from “semantic safety.” A robot must avoid actuator failures and falls, but also understand why knocking over a candle or spilling hot water is dangerous. With more than 40 actuators onboard and children or pets nearby, he says the work reaches from low-level rules in nonvolatile memory through the entire system safety architecture.

  • Figure has spent multiple generations refining its real-time architecture. Adcock points to turbofan engines and self-driving vehicles as precedents for highly complex systems pursuing demanding reliability standards. He sees a path toward safety adequate for millions or billions of robots, while stressing repeated cycles of building, testing, learning, rebuilding, and establishing a track record.

  • Asked about hacking and a physical master off switch, Adcock describes a dedicated cybersecurity team covering both corporate systems and the product. He does not detail a particular switch. His answer is layered and ongoing: “There’s a tremendous amount of work,” with safeguards extending into system engineering and fundamental behavioral rules.

  • Supply-chain exposure is another area of work. Adcock says one week’s tariffs reached 270%, initially prompting the team to expect the disruption would soon end. Rising geopolitical tension reinforced the decision to rotate components out of China. He describes substantial progress and the current position as being in good shape, without claiming every component has become domestic.

11. Space interests Figure, while defense remains outside its mission

  • Space was part of Adcock’s original long-term vision, and Figure had held several exploratory conversations in the preceding six months. One group estimated that putting a human in space to do work costs tens of millions of dollars per hour. Yet discussions about environmental ratings and vacuum-compatible hardware exposed a major gap: “I’ll be candid: we are not ready.”

  • Adcock reaffirms that Figure will not pursue defense work, arguing that selling military humanoids alongside household companions creates a difficult combination. He considers civilian applications vastly larger and more aligned with the team’s goals. The same stance extends to hostile civilians: robots will not be trained to defend themselves against humans, even if some people damage them.

12. Lights-out manufacturing looks plausible before its timing becomes clear

  • Dave frames recursive physical improvement as robots controlling machines that manufacture robots, rather than hands personally fabricating every component. Adcock says quarterly progress has exceeded the lifelong, 30–50-year project he originally imagined. Deep learning began working much sooner than expected, and hardware, models, data, compute, and several billion dollars of cash are now advancing together.

  • A recent warehouse deployment changed his expectations. Figure took a robot to an unfamiliar client site and, Adcock says, it “zero-shot the whole thing.” Previously, changed lighting, conveyors, or package mixes could confuse the model. Asked a year earlier, he would have regarded that result within another year as extraordinary, while expecting it was achievable within five.

  • Adcock feels confident that full lights-out manufacturing can arrive within five years. Whether it takes six months, one year, two, or three remains “foggy.” Figure 4 incorporates design changes intended to remove manufacturing obstacles present in Figure 3: “We think this is the robot that will build itself.” He presents that as an expectation, not a completed production system.

  • Asked what physical manufacturing tasks remain beyond the forthcoming hand, Adcock struggles to name one. He cites ladder-climbing work, forceful interactions, unloading trucks, and carrying 50-pound boxes with neural-network control. The hosts distinguish demonstrating every individual assembly step from generalizing to a completely unfamiliar job; Adcock keeps returning to the latter as the intelligence problem still being solved.

13. Broad physical knowledge should reduce the demonstrations each job needs

  • Figure has spent roughly two years mastering narrow post-training applications, Adcock says. His next objective is a robot entering Baku with training resembling a person’s first days: instructions, documents, slides, and perhaps shadowing. A capable general foundation should let it absorb context quickly, instead of requiring a separately rebuilt policy for every new assignment.

  • Data diversity is producing an unexpected benefit. Adcock describes collecting activities such as washing a dog with a hose and painting a house, then observing better laundry folding. “They love to learn,” he says of neural networks. The connection is broader representation learning: varied physical experiences can improve skills beyond the activities directly demonstrated.

  • Peter extends the argument to prospective humanoid surgeons that share knowledge, avoid jitters, and could provide labor at around a dollar an hour. Adcock’s contribution is the generalist-specialist relationship: a robot might need specific surgical demonstrations while already understanding ordinary physical work. “The surgery bot coming in should be able to go grab you a coffee really quick without any training.”

14. Production capacity and demand estimates sit on different scales

  • Adcock puts current Baku capacity at perhaps more than 40,000 robots annually across four lines at full rate. This is capacity, not a claim that Figure is already producing at that annual pace. A larger manufacturing effort is underway but unannounced, and Figure 4 is intended to help with the company’s manufacturing scale-up.

  • If Figure cracks the current intelligence steps, Adcock sees a difficult but plausible route to a facility producing one million units annually “fairly quickly.” His formulation remains conditional: the company must solve chapter-three problems before that scale makes sense. He describes the resulting demand as feeling “potentially unlimited,” rather than supplying a contracted order figure.

  • Asked about robots eventually outnumbering people, Adcock says he is “fairly confident.” His rough sizing starts with approximately 3.2 billion humans in commercial work, imagines commercial robot demand reaching five to eight billion, perhaps ten, then adds roughly eight billion personal robots. That produces an order-of-magnitude estimate around 20 billion humanoids, without a firm adoption schedule.

  • The intermediate bottleneck is the surrounding human economy. A robot can automate a package line but exhaust its work when trucks stop arriving or upstream employees end their shifts. Adcock distinguishes that phase from whole logistics, supply-chain, healthcare, or data-center operations running continuously: eventually, “entire economies that run 24/7,” rather than isolated machines working longer hours.

15. Adcock sees physical robots as necessary counterparts to advanced digital AI

  • Peter links inexpensive robotic work to universal high income and sharply lower costs for services such as homebuilding. Adcock expects robots to create abundance. Their proposed mechanism is labor becoming cheap and continuously available, with later gains coming from entire operations being automated rather than individual tasks being substituted one at a time.

  • Adcock’s darker alternative is advanced digital intelligence paying humans to execute physical tasks it cannot perform itself. He points to agent-oriented services already hiring people for errands and other work. Without robots, he thinks that relationship could lead somewhere bad: humanoids should perform the physical work that increasingly capable AI systems request, avoiding machines “bossing humans around.”

  • Dave offers a separate manufacturing argument: fixed industrial automation makes enormous quantities of nearly identical products, while AI can help design devices that lack an economical production route. Robots capable of assembling themselves could enable much greater physical variety. Adcock agrees with the broader vision of automated, inexpensive production extending beyond today’s standardized assembly lines.

16. Compute scarcity is already changing procurement and chip strategy

  • Adcock says a compute search six or seven months earlier produced four competitive offers, allowing Figure to negotiate attractive GPU-hour pricing. A later effort to double capacity initially returned no bids, then one expensive offer. He had predicted scarcity in 2027; for Figure, he says it arrived in September 2026: “We can’t get H100s,” including older equipment.

  • Memory compounds the constraint. Adcock recounts a major memory supplier saying existing capacity was committed to phones, servers, and other customers; humanoids were absent from some suppliers’ forecasts. New fabs take three or four years, he says. Figure has arranged memory partnerships that support confidence in near-term robot scaling, while regarding long-term silicon access as a major unresolved requirement.

  • Custom silicon is a possible eventual response, but Adcock highlights the design risk. Chips take two or three-plus years, while inference architectures are changing quickly; a specification written three years earlier would have been wrong for today’s robot. Current chips remain too expensive and power-hungry for the desired affordability, yet locking into an architecture prematurely could constrain the models.

  • Dave suggests specialized chips on available processes around six nanometers as an intermediate route, naming Positron, Etched, and Groq. Adcock does not announce such a project. His broader question is whether existing suppliers can let Figure control its destiny over five to ten years, given the “unprecedented amount of silicon” a sufficiently large robot population would require.

17. China is the competitive threat Adcock watches most closely

  • Asked whether the contest is Tesla versus China, Adcock says, “I personally haven’t seen much out of Tesla in two or three years,” beyond some teleoperated robots. This is his assessment of what he has observed. Peter frames China as Figure’s closest competitor; Adcock says it is moving fast and that Figure intends to beat it.

  • Adcock criticizes current Chinese hardware and policies while emphasizing that annual progress is strong. The hosts mention roughly 150 humanoid companies. He expects Chinese groups eventually to build very capable machines: the current direction differs from Figure’s, but “We just have to beat them.” He declines to single out one Chinese company as clearly ahead.

  • Adcock uses Unitree as an example in his criticism: small hardware, no useful hands, inadequate onboard compute, weak sensing, and poor manufacturing. Researchers add compute backpacks to run substantive models, he says. His preferred competitive test is shipping a home robot with “zero-shot, long-horizon reasoning,” rather than demonstrating increasingly elaborate gymnastics.

  • On regulation, Adcock says an FCC ban on Chinese humanoids surprised him and that he would prefer a level competitive field. He reports little other regulator contact or public discussion reaching Figure recently. His response is to ship incrementally and safely while continuing to build; he says he does not spend every week in Washington trying to shape the debate.

18. Bay Area talent and private capital shape Figure’s operating choices

  • Adcock’s path ran from Illinois and college in Florida to New York, where he sold recruiting marketplace Vettery in 2017 for a little over $100 million. He started Archer the following year and initially considered Los Angeles. Realizing its electric motors, batteries, controls, and sensors resembled a flying robot drew him toward the Bay Area’s electrification and robotics talent.

  • Robotics requires roughly a dozen specialist disciplines, Adcock argues, making location unusually consequential. He considers the Bay Area the strongest ecosystem and Shenzhen the other plausible center for solving general robotics. Recruiting an entire set of teams elsewhere is much harder: the Bay Area is “the most concentrated place in the world for talent” and offers people already inclined to join ambitious technical companies.

  • Archer went public because private markets could not supply the billions required to reach certification, Adcock says. Raising a little under $1 billion was appropriate under those constraints, although the company was too young. For Figure, he wants to achieve general robotics privately before the larger manufacturing, inference, working-capital, and capex demands arrive. Peter discloses that he is an early Figure investor.

19. Hark provides Figure’s speech interface and a separate digital platform

  • Hark operates about 100 feet away and develops personalized intelligence through computer use, speech, memory, vision, models, and hardware. Adcock says Hark Pro launched roughly a month earlier and attracted more than 1.2 million signups and almost five million browser sessions. Usage grew enough that the company restricted some countries, while its hardware remained under development.

  • Figure’s robots use Hark’s internally developed real-time speech-to-speech model. Adcock considers speech the default interface for robotics and AI hardware, while criticizing existing products for weak latency, interruptions, tool calling, and unsuitable devices. “We’re all in on speech,” he says; the objective is a continuously useful interface rather than an occasional voice feature.

  • Figure was Hark’s first beta user and had used it for about five months. Employee adoption is optional, but Adcock describes strong internal usage. The companies also collaborate around larger-scale multimodal pretraining and robotics action prediction. He presents the relationship as overlapping research and product needs, without announcing an acquisition or a formal combination.

20. Personalized intelligence depends on access that users can trust

  • Adcock says Hark already generates more revenue than Figure. He separately guesses it could become “bigger than Figure” in “the next sub-six months,” “maybe sub-90 days,” and potentially even next month, before Figure might overtake it again later. Those broader size predictions are speculative; his explanation is that digital distribution can scale on a substantially different slope from manufacturing robots.

  • The product motivation was dissatisfaction with chatbots functioning as search replacements. Adcock wants an assistant that knows him, accesses his accounts with permission, pays property taxes, books travel, anticipates reminders, and executes longer assignments. For Figure, the parallel is explicit: “The hardware is just the distribution for the AI,” carrying intelligence into the physical world.

  • The hosts identify trust as the common dependency. Adcock likens the information involved to a person’s diary accumulated over time and calls for “a new level of security and privacy.” He says protecting consumers limits advertising and other practices available to competing businesses. He also revises his own timing assessment: after worrying Hark was late, he now believes it is early.

21. Adcock expects AI devices to replace the consumer centerpiece

  • Adcock separates billion-unit centerpiece platforms—today’s phones and computers—from accessories such as earbuds, watches, and glasses. He cites AirPods at approximately 3% of Apple revenue to illustrate the distinction. His categorical thesis is that new AI devices “will kill the devices in the middle,” creating new central platforms with their own peripheral ecosystems. Hark already has undisclosed hardware he says he uses daily.

  • He attacks current Meta glasses and Spectacles as poor products, arguing that accessories cannot work well when the underlying phone platform is unsuitable for them. He declines to disclose Hark’s form factor. Longer term, he is interested in BCI, but describes a nearer “language device phase” in which speech, vision, computer use, and persistent memory become continuously accessible.

  • Asked about OpenAI, Adcock says Figure ended the collaboration about two years earlier after roughly a year of work following the Series B. He alleges that OpenAI staff rarely came to Figure and progress became heavily lopsided while Figure shared its research. He also says OpenAI’s renewed robotics interest created tension. These are his account of the partnership and its breakdown.

  • Adcock predicts traditional phones disappear within ten years and argues Apple has lost substantial AI momentum. He had expected a major GPU purchase and aggressive internal pretraining push; instead, he thinks the gap has become difficult to reverse. His design principle is that models and devices must be co-designed: the hardware is an interface to personalized intelligence, not something onto which that intelligence can simply be attached.

22. Adcock’s optimism about entrepreneurship comes with an unusually intense workload

  • Asked whether a younger founder can still follow this path, Adcock describes three children, three companies, and working at the office every night for a year: “It’s the best time of my life, and I never want to do it again.” He does not portray parallel company-building as effortless. He nevertheless says the work itself is what he and the team want to spend their time doing.

  • Figure’s operating culture requires showing up, including nights and weekends; Adcock rejects work-from-home arrangements in a business whose manufacturing staff must attend the line. His competitive premise is that “you can literally always move faster.” More effort can increase annual progress, and he views speed as necessary against the collective resources and improving capabilities of Chinese competitors.

  • For future founders, Adcock expects AI-native development and robots to make physical products increasingly accessible through instructions. Imagination becomes the limiting factor, and he says AI currently has little of it. He still keeps a notebook of business ideas, sees capital and startup infrastructure as more accessible than when he graduated, and expects an “unbounded opportunity” for the entrepreneurs coming after him.

Full transcript
Peter Diamandis

As you think about competition, is it Tesla and China? Is that how you see it in the world right now?

Brett Adcock

I personally haven’t seen much out of Tesla in 2 or 3 years. I’ve seen China do a lot of dancing. We care about something very different from what we think China cares about. We just have to go beat them. It’s a real race here. The work isn’t really practical. The technology stack needed to get there isn’t practical, and the AI stack to do it can be built by a college senior.

My view is that there are roughly 4 stages in this humanoid race. In chapter 1 of my book, you have to build really good functional hardware. In chapter 2, you have to be able to build an AI recipe. Chapters 3 and 4, for me, are about scaling intelligence and scaling robots.

Dave Blundin

What might we see with Figure 4? What areas are you focusing on improving?

Brett Adcock

We’re in the flip-phone era for humanoids. Figure 4 will probably be our iPhone 1 moment, and it’s coming soon. We need the robot to get cheaper. We’re trying to make the robot more capable. A little sneak peek: we’ve been working on the hands. A Figure 4 wearing a glove and a human wearing a glove are almost indistinguishable.

Peter Diamandis

That’s like an iPhone moment. Figure 3 is probably the best humanoid on the market by far, but Figure 4 will make Figure 3 look like a caveman.

Dave Blundin

Now, that’s a moonshot, ladies and gentlemen.

1. Inside Figure AI's Robot-Filled Headquarters

Peter Diamandis

So, Brett, it’s good to be here for our annual catch-up. It’s been about a year since Dave and I were here to see Figure, and what a difference. You walk in—you just parked in the parking lot over there—and there are about 100 robots outside to greet you.

Peter Diamandis

So my question is: what do you call a group of robots? I’m calling them a flock of figures. Do you have a name for them?

Brett Adcock

Yeah, those are our visitor bots. They walk around the campus. They can open doors, walk up any stairs, and use the elevators. They’re there to greet people and talk. Right now, we’re working on getting them to, you know, get coffee for folks. We have a stretch goal: it’d be great if you could come to campus and never talk to a human as you get set up. You’d get fully checked in, walk to the conference room, do a tour, and go to the bathroom—all through a humanoid.

Speaker 3

Let me show you where I was built. Let me show you what I’m doing today.

Brett Adcock

Yeah, no, exactly. We do that all the time, so why not? We should have a humanoid.

Peter Diamandis

And you said something a minute ago that’s pretty profound: you have more Figure robots than humans now.

Brett Adcock

Yeah, we have more Figure robots on campus.

Peter Diamandis

When did that happen?

Brett Adcock

Probably 4 or 5 months ago. We crossed the number of people here—I think we have 700 or 800 people. So we crossed that number. We have more robots now, many more than humans.

Dave Blundin

And how many have you produced at this point?

Brett Adcock

I don’t know, around 1,500 or so on campus, something like that. We have well over 1,000 on campus, and we have robots in other places around the world as well.

Dave Blundin

Oh, nice. We’re seeing Figure 3 around here. Last time we were here, you were just transitioning to Figure 3. How long before voice? How long before I’m talking to the Figure?

Brett Adcock

You can talk to any robot here.

Dave Blundin

Yeah?

Brett Adcock

Yeah. You can stop it and just ask a question.

Dave Blundin

You can talk to that one?

Brett Adcock

It has real-time speech-to-speech. Every robot here is equipped with a speaker and microphone. We actually use HARK for our voice AI, and every robot here has full real-time speech-to-speech.

Dave Blundin

Incredible.

Brett Adcock

Yeah.

Peter Diamandis

Incredible.

Dave Blundin

A lot of people are asking me, “When am I going to be able to get my…” Let’s jump into some timelines and applications to kick things off. I think that’s the most interesting. When can I order a Figure for the home? I think most people want to know the answer to that question.

Brett Adcock

At a high level, we have 2 tracks. We want to ship robots commercially, and we want to ship robots in the home.

Dave Blundin

Yeah.

Brett Adcock

They’re really different spaces. The commercial market is easier in some ways, where the tasks are very well specified. You can almost write on a piece of paper exactly what the robot does every day. It doesn’t really deviate from that. But it’s really hard in the sense that we have a lot of speed and performance goals to hit, and if we mess up, you get fired. Not great.

The home is a little bit the opposite. We don’t need the speeds we would need in the commercial market, but there isn’t really a specified task. Every home is different. I bet if we asked what kinds of appliances the 3 of us have, they’d all be different. Our homes and their setups are different, and the things we’ll want the robot to do will change quite dynamically.

We’re working on both in parallel. We think they’re both intelligence problems. Figure 3 is really capable of doing most things in a home, so we’re working hard on building larger pretrained models and doing a lot more post-training in RL to crack this problem.

I think you’ll get a different opinion if you talk to other companies in the space, but our view here is that Figure will ship a robot when it can ship a really high-quality product for the home. It needs to be able to walk into the home, talk to you, and do real work. We don’t want to be teleoperating it in the home. We don’t want the robot to have issues or failures, or to stop because it’s unsure what to do. It needs to be a high-quality product when it enters the home—an Apple-style quality product.

Dave Blundin

You have the kids and the dog to worry about, too.

Brett Adcock

Yeah, exactly. It needs to work. It needs to be able to go in there and create this magical experience of having another person in the home. That’s our launch criteria. We’re not going to launch before then.

I think you’ll see others do that, and in my mind it just won’t work. It’ll be like what you’ve seen with AI devices—with Humane, Rabbit, and these other groups in the AI device space. It’s not going to work until you have this really magical product. For us, we’re trying to crack that.

The key constraint is intelligence. Figure 3 robots have the ability to fold laundry, do dishes, unload groceries, open the refrigerator, open the oven, and do real work in the home. They can clean the house up. We’ve shown some of that, but we need to be able to do it end to end for a dozen-plus hours per day, every single day, 7 days a week, for years.

We’re pushing on that. I’d say the biggest constraint is intelligence. How do you drop a robot in any home, just by talking to it, and have it do useful work? There are other things: an integration problem, a safety problem, and other things that are really hard. But by and large, it’s an intelligence problem we’re trying to crack.

Without making a promise—and I won’t tell anybody—we’re doing testing now. At any given time, we’re in somewhere between 20 and 100 new homes in California that we haven’t been in before.

Dave Blundin

Okay.

Brett Adcock

We’ve been doing this since the launch of Helix 2.5. We’re doing it mostly for evaluations of our models and a ton of hardcore testing. My forecast is that the first thing we need to prove to ourselves is that we can do end-to-end work in a zero-shot environment: a home we’ve never been in, where we haven’t trained and haven’t seen any of the objects before. Just through speech, we can take commands—a prompt—and basically do end-to-end work.

We’re doing that really well today at the micro level. If you ask it to do one thing, like make the bed or unload dishes—

Dave Blundin

I literally just did that, actually. A totally open-ended prompt. There was no preconceived idea of what I was going to ask you to do.

Brett Adcock

Yeah.

Dave Blundin

And it takes a really good shot at any task right out of a prompt.

Brett Adcock

Yeah. Maybe I’m being a little verbose, but we need to be able to do that fully end to end over long horizons. We’re training larger and larger models now. What you saw there in Blackbox today was Helix 3. You’re probably the only people in the world who have seen this outside this office.

I don’t know what you guys were feeling, but for me it feels like a baby AGI moment. I’d characterize it as an adult body—an adult human, a robot, but a really dumb adult human, like a 2-year-old in an adult body.

Brett Adcock

You can basically ask it whatever you want. It can do a lot of things, and some things it just can’t do well yet. But it was significantly undertrained. It was our first model checkpoint coming out of Helix 3. We now have much larger models being trained, and we’re evaluating them even this week.

My bet is—what date? What year?

Peter Diamandis

What year?

Brett Adcock

My bet is, within the next—I hope to see, and I obviously can’t promise, but I hope to see in the next sub-six months, perhaps even sooner, we can put a robot into any home and have it do fairly large, long-horizon tasks end to end.

Peter Diamandis

Over hours and hours. Then the question is when you would actually sell them to consumers.

Brett Adcock

Yeah.

Dave Blundin

Yeah. Then we’ve got to figure out how to get them to consumers and make sure they’re really safe. Reliability has to be really high. There can’t be issues and faults.

Dave Blundin

Do you think you’re going to keep robot maintenance and repair as a core in-house capability, or approve outside vendors to do that all over the world?

Brett Adcock

For now, we’re going to do it ourselves.

Dave Blundin

Yeah.

Brett Adcock

We really want to make sure the robot is the highest-quality experience in the home, and we know the robots better than anybody else. We built them, so we’re going to do that ourselves.

Peter Diamandis

You can just crawl into a box and ship itself back.

Brett Adcock

Yep, totally. There are—

Peter Diamandis

The label on itself.

Brett Adcock

The hand comes out, slaps it on, and closes the box. There are many ways to solve this problem. The end state is for sure us not doing everything. But in the beginning, as we ramp the product up, you really want this incredible experience in the home, where it’s basically just like another human is in your house. We’re only going to get that with the highest level of care and attention. We’re putting everything—our blood, sweat, and tears—into this to make sure it comes out really rocking.

The flip side is, if it comes out and doesn’t work really well, you’re kind of dead.

Peter Diamandis

So we really want to make sure it’s really high quality. You were a Steve Jobs fan as a kid?

Brett Adcock

Of course. Yeah.

Peter Diamandis

All right. The parallels are unbelievable. I was with Mike Rowe and Joe Polish at an event recently. Mike Rowe was the guy who did the Dirty Jobs show.

Brett Adcock

Yeah.

Peter Diamandis

I told him I was going to be coming here, and he asked when we’re going to see Figure robots doing welding and roof repairs, and all the blue-collar work. There’s been an interesting conversation that AI is coming for white-collar jobs right now, and it is—from finance and legal to diagnostics. But the question is, when will these robots start coming in and doing traditional bricklaying and all the manual labor out there?

2. The Four Stages of Humanoid Robotics

Brett Adcock

Yeah. My thesis here—let me lay it out. I feel like it may be helpful to give you guys my view of robotics. At a high level, there are roughly 4 stages in this humanoid race.

My chapter 1 is that you have to build really good functional hardware. The robot hardware has to be incredible. You can’t build AGI on a robot without really good hardware. We’re doing that now, obviously, with Figure 01, 02, and 03. We have 04 coming out at some point. We think we’re—anyway, the hardware needs to be able to produce the speed, torques, and range of motion necessary for a human, roughly speaking. It has to be affordable, and you have to make a lot of them. That’s one.

Two is that you have to build an AI recipe that can run end-to-end, long-horizon work: full locomotion and manipulation, meaning walking the body and moving the hands. Besides Figure, we really haven’t seen any of that in the world. You’ve seen groups do tabletop work with off-the-shelf robots that are bolted to a table and overfit to a certain environment. You haven’t seen robots go off autonomously, walk around, and do stuff. We started doing that with Helix 2, full room-wide.

Peter Diamandis

I have the memory of being here when your very first robot’s big deal was putting a K-Cup into the coffee maker.

Brett Adcock

That was—

Peter Diamandis

Forever ago.

Brett Adcock

Arguably, maybe one of the most important turning points we’ve ever had.

Speaker 1

Getting coffee. Yeah, that’s a turning point for everybody.

Brett Adcock

Honestly, it sounds funny and ridiculous, but back then—I think that was 2 or 2½ years ago—everything was coded. Everything on the robots was C++ code. We were using a kind of autonomy 1.0 stack for SLAM, object detection, and everything else.

There was really no path to making these work at scale. The robots just wouldn’t. There was really no way to build a billion units and make them all work across the world. In that demo, we had Figure 01 put a K-Cup into a Keurig. I was sitting there with the team, and I said, “Listen, if we’re able to pull off full upper-body dexterous manipulation with a neural network like this, it’s the first time in human robotics history, I believe, that anybody’s ever done this. But it paves the way for all this to be done through a neural network, with no code.”

Peter Diamandis

Yeah.

Brett Adcock

This is what we really need. The entropy is too high in the space to be able to go out into the world and do everything we want to talk about, like go into all your homes and throughout the world and be able to do stuff. Think about the blue-collar work you just mentioned, too. That involves really forceful interactions with the environment. You’re grabbing objects, making impact. It’s really quite difficult work.

I told the team, “I think if you do this, it’ll be so historic that we’ll probably raise $500 million.” I put $500 million on a Post-it note and threw it on a coffee cup. I said, “If you’re able to actually do this demonstration and show us internally that these neural networks really work, we’ll probably raise half a billion dollars.”

We did it, and I think we raised a little under $700 million shortly thereafter. It was a huge turning point. It showed that deep learning works for robotics. That entered, for me, chapter 2: you’ve got to have an AI recipe that really works well, or you have no hope of making the story work.

Peter Diamandis

People don’t realize you jumped into robotics before neural nets were really a thing. Self-driving cars, too: they were 90% C++ and 10% neural net.

Brett Adcock

And now it’s completely inverted. You’ve got to keep the C++; the brake is connected fly-by-wire, so you need some deterministic code in there. But basically all the decision-making is neural net.

Peter Diamandis

So for the Keurig cup, was it literally 100% converted to neural net? Not a single line of handwritten code?

Brett Adcock

We were sending in a controller that was in C++, but the whole upper body—all the hands and all—was a single neural net. We were basically maintaining balance, but everything above that was a neural network.

Peter Diamandis

And then, to be frank, it took us about a year to go from a tabletop like that to all neural nets all the way down the stack, moving around.

Brett Adcock

A year. Helix 1 was upper-body neural nets, and Helix 2 was full-body neural nets.

Peter Diamandis

Yeah. That took us a year. I was in a stand-up right behind us in the lab here for a year, trying to figure out how to move the robot off a table to do work. We always thought it was really trivial. I should be able to joystick a robot somewhere and then do work again, because you’re usually pretty static when you’re doing work. You’re not walking around making coffee. You’re just standing there.

Brett Adcock

But we realized the hard truth was that this was a full-body experience. Your feet are involved; almost your whole body is there just to get your hands and your head to a certain place in the world and stay balanced.

Peter Diamandis

Yeah.

Brett Adcock

What really mattered was that the body actually mattered more than the hands and the head. What we learned is that all of those joint trajectories needed to be done as a single, almost unified experience. You need to understand what the knee, hip, and feet are doing in relation to the fingers and hands.

Peter Diamandis

At the same exact time.

Brett Adcock

Yeah.

Peter Diamandis

So it needs to be all trained together. I think a lot of people don’t appreciate that. If you watch YouTube all day, you think the Chinese have these insanely dexterous robots, but under the covers—

Brett Adcock

There’s—

Peter Diamandis

It’s teleoperated, or you build these macros: 10 movements that are preprogrammed. Then, when you stitch together the video, you think, “Wow, China’s miles ahead of us in robotics.”

Brett Adcock

The tech stack underneath it is totally contrived. It’s not a real neural net that’s learning on the fly.

3. China's Robotics Progress and Limitations

Brett Adcock

Let me just be clear about what we see out of China. We see robots dancing, and these are deep learning RL controllers.

Speaker 1

Yeah.

Brett Adcock

One of the best forms of it is training through DeepMimic. These are humans wearing a motion capture suit, like a Hollywood film suit. They’re running around doing ninja moves, and then you’re basically training an RL controller to do this and replay it back on the robot. They’re all done open loop, meaning you can just push a button. They don’t react to the environment at all.

A lot of them have little LiDARs in their heads, so they localize when they’re in parades together and don’t run into each other. All of this work is completely orthogonal and different from what you need to solve general robotics.

While it’s impressive to see, maybe for the hardware demonstrations you’re doing, none of it is practical for what we need robots to do at scale. The work isn’t really practical. The technology stack needed to get there isn’t practical, and the AI stack to do it can be built by a college senior with one robot. This is very easy stuff. We’ve done this work before, so we understand it.

One of the coolest things about the investment community in Silicon Valley is that they get the difference, because it’s very subtle to the outside viewer. It’s the difference between the two things. But if you watched ChatGPT become a thing while expert systems just kind of did nothing—

Speaker 1

Yeah.

Brett Adcock

—and you look at early on, it’s giving, “I can’t spell strawberry. I can’t add 4 and 4.” But you knew it was on this trajectory where it was going to punch through. The way my friend John Le, our mutual friend John Lenny, describes it is like building a submarine underwater. You don’t see anything, and then suddenly a periscope pops up. You’re like, “Wait, that [unclear] is the periscope. That means there’s an entire submarine that’s about to emerge.”

This periscope is, we think, starting chapter two, which is robotics work. We ended chapter two with doing kind of a 200-hour livestream with logistics work. You’re seeing it right now. Those robots on that line haven’t stopped since we did that four months ago. They run practically every day.

Speaker 1

Somewhere in the AI, they go themselves, “Could you please let me do something else?” [Laughter]

Brett Adcock

You’re hitting an important part. Chapters three and four for me are scaling intelligence and scaling robots. What we really want to see is what you’re mentioning here with the submarine. As you’re putting more input in, you’re getting better output. For us, our input is data and compute, and our output is getting robots to do more useful work.

Peter Diamandis

Well, you’re blowing through. So stage three is data, data, data.

Brett Adcock

It’s data and compute. It’s scale. It’s AI scaling laws.

Peter Diamandis

When you stood up Helix, the models that existed back then weren’t interfacing with physical systems. Recently, you saw GPT Astra actually be able to operate robots. Talk to us about that. Is Helix a general model that can operate a physical system?

Brett Adcock

One thing I’ll say about these other types of language models and even multimodal models commanding robots: they’re all being used for very basic tasks on a fixed camera frame, very slowly.

If you slow robotics down enough, none of it becomes a really difficult autonomy robotics problem. These videos are shot at 5, 10, or 20 times speed. People realize they’re sped up. They’re sped up because they slowed everything down for latency. These become almost like closed-loop control, but done really slowly.

You need that feedback, but the models aren’t good enough to output this really quickly, and they’re in very low-dimensional spaces on very basic tasks. We have a different problem to solve. We need to move at very different dynamics and frequencies from what we’re seeing in these videos.

We need a robot to walk around and react to the environment really quickly. It needs to move a package over there in under 3 seconds per package. It can’t mess up very much, only a few percentage points over a whole shift.

We need robots to go into a home and do very delicate, fine-grained manipulation: grab cups and dishware and put them away, open doors and cabinets, and navigate through tight spaces like a bedroom or kitchen. We have to do that on some of the most complex mechanical machinery ever built, without messing up. You have to balance on basically a quarter-sized point in your foot above the center of mass in your torso.

It becomes a very different problem when you attack it like this than what you see on tabletop college setups. You’re seeing that now with Astra and the rest of the groups. They’re not real models for robotics. If they were, we could run them.

Peter Diamandis

Sure. So, I think—

Dave Blundin

There’s another subtlety in there, which is that your entire AI stack is totally owned by either Figure or [HARKC?]. When you do a university tabletop setup, it’s on top of at least one layer of a third-party neural net LLM.

Brett Adcock

It’s just not real. You can’t put them into really high-dynamic robots like this and make them work.

Peter Diamandis

Yeah.

Brett Adcock

And you can’t scale them. There’s no path where you can go from that [unclear] to full locomotion and manipulation, then to what we did with Helix 2.5 recently. The architecture needs to change dramatically.

These setups have basically unlimited compute. They’re plugged in with really low-latency communications through a network. They have fixed cameras. They’re generally overfitting the tables for data collection. They’re not scenarios that are realistic for what we’ll see.

4. Helix AI and the Future of Robot Intelligence

We had to build Helix from the ground up to figure out how to build the right brain for the robot. Right now, I’d say we arguably understand how the brain works really well. What the brain is starved of today is data and compute.

We don’t have enough world understanding of what forces feel like, joints, and observations to get it into every home in the world. But we do see a path now. With Helix 2.5, we went into 30 rental homes that we rented the day before or the day of, and we were able to do work like make the bed and tidy the living room without ever seeing the house or objects before. It was fully zero-shot. They call it zero-shot in the community, where none of this was in training. It was the first time these robots had seen this.

5. Scaling Production to Millions of Robots

Think of this as a bunch of stair steps we’ve been taking over the last 3 years, all building toward how we get from there to dropping into any home for 10 hours, where the robot can do work. That’s the trajectory we’re on now. That gets to the heart of your question about when we’re going to see that. We’ve arguably seen some of it this year, in 2026. I think you’ll see more from us this year. I’d be very shocked if we exit 2027 without having seen this whole picture.

Peter Diamandis

How much of the compute is running on the Figure robot versus on Wi-Fi?

Dave Blundin

Also, training versus inference. How much is training? How much is inference?

Brett Adcock

Most of our compute today is in training. The dollars are going toward compute. We’re doubling our compute internally next week. We also did a big deal with Nscale for 2027.

Peter Diamandis

Was that deal 100,000 GPUs, over $3 billion? When does that come online?

Brett Adcock

Those are Vera Rubins.

Peter Diamandis

Yeah.

Brett Adcock

Those are the second half of 2027. Those are VR 2200s, the best computers in the world.

Peter Diamandis

Yeah.

Dave Blundin

Wait, what’s the ratio of training compute you’ll have that day versus, say, a month ago?

Brett Adcock

Let me do the calculation. They’re different types of GPU chips from what we have now.

Peter Diamandis

Figure’s Vera Rubins are 2 times the raw compute. He’s using B300s today or something like that.

Brett Adcock

Yes. They’re roughly on the order of 4 times—

Peter Diamandis

4 times.

Brett Adcock

Yeah. If we took the whole deal with Nscale and sized it all the way up, it would be well over 100 times what we have.

Peter Diamandis

Over 100 times.

Brett Adcock

Yeah.

Peter Diamandis

So, going back to the question of—

Brett Adcock

Yeah.

Brett Adcock

I mean, we also spend nine figures a year on compute now. Before this, we were already spending a large amount internally on compute.

Peter Diamandis

Intimidating to everybody else. Yeah. Inference on the robot, on Wi-Fi?

Brett Adcock

We do all inference on board today, on the robots themselves.

Peter Diamandis

So what have you got on board in terms of compute?

Brett Adcock

All of Helix is fully on board the robot. Everything from planning and reasoning to the low-level work of sending out trajectories and torques is done on the robot today. The only thing we do today is real-time speech-to-speech through servers.

We use servers to help think about the robots as agents, as physical agents. If I had 100 robots building a home or doing construction, how do we command that whole fleet minute to minute, hour to hour? As the home gets built, the job changes.

Peter Diamandis

Yeah, exactly.

Brett Adcock

That’s all done by our agent orchestration in the cloud today, which is the only way to do it. In this case, we have multiple robots on this conveyor subbing out and recharging. If a robot gets damaged or doesn’t feel healthy, it walks to maintenance and another robot comes in. It flags the fleet, so all of this is being worked together. That has to be done in the cloud, so that will always be done in the cloud.

We’re experimenting with onboard and offboard compute. The struggle is that we can’t get around the physics of having very high dynamics on the robot. We need to talk to all the joints and sensors, then communicate back the new plans really fast. We don’t have seconds to wait, or even half a second. We need to get it done at the millisecond level. There’s no getting around that as we work with the very fast dynamics of the world.

In the best-case scenario, you have a big brain living in the cloud and a small brain living on board. In the extreme, you have everything living on board, which is what we’re doing now. The jury’s still out. We’re still doing a lot of experimentation, but all the robots right now are working on board. I think it’s nice because we can lose internet and the robots still work. Even at home, if you have bad 5G service, you don’t want your robot stopping mid-laundry. You want it to continue working. There are a lot of engineering trade-offs we have to work through.

Peter Diamandis

I totally understand this better. So, milliseconds—is that even viable on the cloud?

Dave Blundin

The home Wi-Fi is a bottleneck, but then getting the milliseconds—

Brett Adcock

The best 5G networks we’ve seen can get it down to maybe 20 to 30 milliseconds per leg. You’re looking at almost 50 milliseconds for a full round trip. You’re already at 20 hertz just in the network itself, and then you have the whole forward pass to do for inference. There’s a bunch of other work we need to do related to the images and torques coming back. It’s not as easy as pinging a cell tower and getting it back. We have a lot of other work to do.

Peter Diamandis

Do you think every home is going to have some kind of brain box in the basement, where the cable boxes are?

Brett Adcock

Our hope is that everything’s done on board, and we have the right architecture now.

Dave Blundin

But if every home has 40 of them, do you really want to—

Peter Diamandis

That’s a good market projection. Forty per—

Brett Adcock

No, I mean everything’s done on board, like we do have on board right now. The robots still need—

Dave Blundin

Chips are scarce, though. You could share it, because you might have a roof or gutter-cleaning bot up there.

Brett Adcock

I don’t think so. There’ll be some neural nets that, no matter what, will have to live on the robot. They’re just running too fast.

Peter Diamandis

Sure. It’s all about training world models versus physical models, and what you’ve been doing in terms of task video recording. What’s your vision for training to get to that level of fidelity?

Brett Adcock

We took a pretty big bet on Helix 3. We talked a little bit about it at the all-hands, but we haven’t released anything publicly yet. Before that, we released Helix 2.5 a few months ago, where we showed zero-shot reasoning in 30 unseen homes. We also showed what we think are the first humanoid scaling laws for data. As we increased the data, we saw a very log-linear relationship: the model losses came down, and the robots got better at executing in these environments with more data than with less data.

That was unbelievable to see. We’ve seen this in LLMs, but we hadn’t seen it in robotics before. If this trend holds, then as we add more compute and data to the problem, the robots are literally going to get smarter. They’re going to get better at tasks every single day.

Brett Adcock

We’ve been doing new work on Helix 3. I think we gave you guys a little preview today. I’d largely characterize this as little baby AGI. We have a dude in a suit in the office. We have a secret lab called Blackbox that you guys were in, which is like a little home.

Speaker 1

You said we’re the only outsiders who have ever been in there.

Speaker 2

Yeah. Wow, that’s great. It’s a mind-blowing experience. When the LLMs were improving, they didn’t really look anything like a child. They were doing complex math, but they were also hallucinating. It’s nothing like a kid.

6. Robot Pricing and the Figure 4 Breakthrough

What we just saw in the Blackbox looks exactly like a kid learning how to do things. Nothing is pre-programmed; it’s all neural net. You ask it to do a task it’s never heard before, and it looks exactly like a kid who’s learning how to do it, experimenting and trying. It’ll pick up the chair, and it’s not going to work. You can visualize how this is evolving, just like a kid. It’s so much cooler than the way the LLMs evolved.

So, your production rate right now—I mean, the demand for these robots in homes. I’m asking people all the time: if these things cost $30,000, are you still tracking toward that price tag, or a $300-a-month lease rate?

Brett Adcock

Yeah, in the home, we’ll have to make the robot really affordable. You’ll need some sort of $300 to $500 a month lease.

Speaker 2

Yeah, $10 a day. I mean, it’s under a buck an hour. I ask people when I’m on stage, how many of these would you own if it was that? The average individual says at least 2.

Brett Adcock

Yeah.

Speaker 2

Yeah, you’ve got to either lease it or buy it. There are only 2 ways. We’ve got to get the robot out. Do you have a preference?

Brett Adcock

Not really. I think through a lease, you can bring the cost down quite a lot and advertise that better for the customer. I don’t really know. I think the technology would be radically improving. My view is that we’re in the flip-phone era for humanoids.

Speaker 2

I think Figure 4 is probably the iPhone 1 moment we’re going to have here soon.

Brett Adcock

It’s such a significant step up from where we’re at with Figure 3, but we’re really early. With the iPhone, from generation 1 to generation 8 or 10, every year the hardware got better. The cameras got slightly better.

Speaker 2

Slightly better camera.

Speaker 1

Now you can’t really tell. It’s all asymptotically approaching some perfect iPhone vision for them.

Brett Adcock

For us, we’re really early. I think the robots will continue to get much cheaper, more dexterous, and safer to be around. The way you interact with the humanoid will really improve. I suspect that with a lease model, you can get the newest version every few years, which will be good.

Speaker 1

It feels like your next generation has been on an 18-month cadence.

Brett Adcock

Yeah, we really want to do our next generation at significantly high volumes. We spend a lot of time on it, and we’ll spend about a year between programs right now. We’ll spend more than that on—

Speaker 1

What might we see with Figure 4? What areas are you focusing on improving?

Brett Adcock

Oh, man.

Brett Adcock

The big drivers at a high level, at the requirement level, are that we need the robot to get cheaper. That’s super important. We’ll induce more demand. We’re trying to reduce cost as much as we possibly can.

We’re trying to make the robot more capable. You want it to be able to do more things in the world than it has before. A little sneak peek at that is we’ve been working on the hands for almost 2 years now. We’re probably the only group that’s built both tendon-driven and non-tendon-driven hands. We’ve done 7 iterations of our hand, and I think we really understand hands. One of the biggest step-ups we’ve ever made in the hands is coming. I’d make the argument that a Figure 4 with a glove and a human with a glove are almost indistinguishable.

Speaker 1

Really? Are you serious?

Brett Adcock

Yeah.

Speaker 1

That’s like an iPhone moment. Everything I’ve seen so far is functional, but even the way they grip things—you know, when it was picking up the cups, it was reaching into the cups rather than grabbing them on the outside. I hate to say this, because I think Figure 3 is probably the best humanoid on the market by far, but it’ll make Figure 3 look like a caveman.

Speaker 2

Really?

Speaker 1

Yeah. So where are you in terms of—

Brett Adcock

It’s funny because the iPhone analogy is really interesting. The iPhone 3 was the first really big hit, but it had no camera. The camera came later, and that was a step-function change in that trajectory. I’ll bet the hand is the equivalent of the camera. You throw that hand on there and it has human-level dexterity.

If you look at all 3 robots, they look different from each other. We’ve obviously changed a lot across the 3. We changed from 1 to 2. We changed the architecture of the structures from an endoskeleton to an exoskeleton. All the loads on Figure 2 are on the exterior, like your aircraft does, right?

Speaker 1

And how we build aircraft at Archer.

Brett Adcock

It’s a very different type of architecture. We went that way as we designed the robots. You’ll line all these up, and Figure 4 doesn’t even look like the rest.

Speaker 1

Mm-hmm.

Brett Adcock

It’s just like, “What happened here?” It was a full deviation from our lineup. It kind of looks like it’s maturing, and there’s just a fork in the road.

That’s what I mean by it being such a significant robot. We had a couple of hunches about where things should be headed. We ended up doing a bunch of work on those, and they panned out really well. We took a couple of pretty aggressive bets on complexity to get there. Fair to say, I think Figure 4 will be the largest step-up we’ve ever made between generations.

Speaker 1

That’s pretty extraordinary, because 2 to 3 was a huge step-up.

Brett Adcock

You’ve never seen it like this before. The largest step-up is in Figure 4.

Speaker 1

Maybe you’ll sneak us a peek later after the pod. The hand touch is super critical, right? Being able to get that data. A friend of mine has a company called Touchlab that’s built out the most advanced touch systems for gathering data. Are you doing that all in-house right now?

Brett Adcock

We do it all in-house. The problem with a lot of these external groups is we have to be production-ready. They have to work at really high reliability. It’s got to work perfectly for our system, and it’s got to be low-cost. We need to revise and fix it if we understand it. We need to understand the firmware and embedded software around it. We need to understand the thermals around it.

It might seem like we can just go buy something, toss it in the robot, and make this magical experience. It’s never the case. It never fits. You can’t even make it fit in the CAD, let alone get the thermals to work. Maybe it’s not the right bus, the power loads are too high, the cost is unfeasible, or it doesn’t work past 5 or 6 times and it breaks. None of this stuff is easy. And we have some of the best engineers in the world here who want to move really fast.

Speaker 1

Same thing with how Elon concluded: you need to vertically integrate everything.

Brett Adcock

It never fits. You can’t even make it fit in the CAD, let alone get the thermals to work, or have the right bus connections, or get the power loads right. The cost might be unfeasible, or it might break after 5 or 6 uses. None of this stuff is easy. And we have some of the best engineers in the world here who want to move really fast.

So when we’re working with another group that isn’t beholden to us to do that work, it becomes really tough. That’s what happened with Figure 1. We went out hoping to get the stuff and pull it all in on the supply-chain side, and it all broke down really fast.

We bought the first actuator and brought it in. I think we still have it here. We thought, “Okay, we’re going to use this for an elbow. This is an elbow actuator.” It had all the right stuff: a torque cell, the right encoders and gearboxes, and the actuator’s speed and torque profiles were in the right place. We got it, and it was just this big.

Speaker 1

It was literally this giant block, and it’s for your elbow.

Speaker 2

In China, probably.

Brett Adcock

And we were like, “What do we do with this?” It was the best one on the market, and it was literally 3 times too big and 3 times overweight. It was really expensive. We can’t have it cost hundreds of dollars. It needs to be tens of dollars at most. So we had to go build everything ourselves.

Speaker 1

Now somebody in that room says, “Okay, Brett, we’re going to wind our own actuator.” That’s not a small step.

Brett Adcock

No. We sat there and said, “This really sucks. We have to do it ourselves. Is there any way we’re wrong about this? We don’t want to have to do this.” We spent a while trying to figure out every other possible thing to do besides build it.

If you build it yourself, it really sucks. You have to have a firmware person, an electrical person, a thermals person, probably someone doing embedded software, and someone working on mechanical structures. You have to work on the motor—the rotor and stator—and the gearbox. You probably need a gear designer at that point.

Then you need somebody to put it all together in a really small package, because our design team wants it to look really good. It needs to move around, which means the wire, cabling, harnessing, and connectors all have to be really dynamic and never mess up. Then you have to figure out how to get the heat out of the system, where to put the actuators, and how to daisy-chain them together to build a functional actuator inside a shoulder, leg, hip, or wherever you’re going to put it.

It becomes clear that this isn’t a “go get it and put it together” plan. It has to be done from scratch. Every time you make something better on the engineering side, it could get bigger. If it gets bigger, your elbow or knee might be too big, and sometimes it clashes with the rest of the robot. Then it doesn’t look great. You really want a high-aesthetic product. It becomes a very difficult conversation to have across 10 people.

Do you think that a year from now, when you think about the supply chain coming into this, it’s just raw molten metal and bare wire that you’re starting from? If you’re winding your own actuators, you’re literally winding coils. But at the end of the day, you’re not going to make your own CNC milling machine. You’re not going to make your own die casting—or maybe you do your own die casting, I don’t even know. What does it look like?

Dave Blundin

You’re referring to the conversation we had at the Gigafactory, where raw aluminum was coming in at one end and Cybercabs were coming out the other end.

Peter Diamandis

Buckets of molten metal just pouring. You’re like, “You’re kidding me.” It’s right out of a sci-fi movie.

Brett Adcock

I mean, that’s it. To get costs down, sometimes some of that really has to happen.

Dave Blundin

I can just imagine you saying to your team, “More robots. We want more robots walking around.” It’s so much fun.

Peter Diamandis

It’s nice, though, right? It’s amazing. You’re living in the future.

Brett Adcock

Even the weekends, they’re all running. It’s a bummer when they’re not running. You just come in here and it’s like—

Dave Blundin

How many do you have at home? And how are your kids dealing with it all? Are they enjoying it? You have 2 or 3 young kids, right?

Brett Adcock

Three kids.

Dave Blundin

They’re really young, though, and they just want to run. They run over and try to jump on the robot.

Brett Adcock

Yeah. I’m like, “You can’t really jump on these robots.” We’re not there yet. You can’t jump on them. They come to the office every few weeks.

Brett Adcock

They love being here, all the robots and stuff. It’s great.

7. Designing Robots That Feel Human

Dave Blundin

Well, also, when you walk in now, there are a hundred of them just outside the door, and you don’t feel like you can’t walk through them. You just walk right up to them. They’re moving smoothly enough that you don’t feel like they’re going to slap you in the face, which doesn’t mean you can have conversations with them on the way out.

Let’s talk about the Figure face, because that’s an interesting set of choices. There’s the uncanny valley of making something look too humanoid. Right now, you’ve got a black screen. What’s been the trade-off in your internal conversations about what the face of Figure should look like?

Brett Adcock

Yeah, there have been a lot of conversations about that. We still have conversations about it. The face is really tough because we’re missing a lot of the facial movements of a human.

When you talk to the robot, it has lights and a screen, and it has the gestures it can make with its neck and its whole body. But there’s still something missing to give it a humanlike experience. We’re thinking about making the face more humanlike, because right now it has a black face.

Dave Blundin

It feels friendly, but it’s a little ominous to look at a shiny black face.

Brett Adcock

Yeah, we’re working on the face. We have some big plans for Figure 3’s head and face, and we’re working through this. It’s a hard problem.

Dave Blundin

It’s a human psychology problem.

Brett Adcock

Yeah. We’re talking to the robots more and spending time with them. The robots have memory now. We’ve done work where I can go up to a robot and say, “Hey, Brad, good to see you again,” and it remembers our last conversation.

You’re getting to the point where you might want it to be almost like a companion. You want to talk to it, especially in the home, when it’s around your family and you’ve had it for a long time. We call this HMI, or human-machine interaction. Getting that really well dialed in is something we have a whole team focused on. We have an HMI lead. Every week, we meet, plan, and do demonstrations. It’s really hard.

A lot of it comes down not just to the face, but also to how the body reacts. We’re pretty expressive with our faces and arms as we talk, making small gestures and nodding. All of that helps make something more natural to interact with. We’re doing a lot of work there. The face is a tough one. The thing we’ve really stayed away from is putting googly eyes on the robot.

I think this is one of the most complex pieces of technology on the planet. If we put googly eyes on a robot and say, “Oh, it’s friendly; it’s safe,” that’s a huge disservice. It’s a con. The robot has googly eyes, but it can walk upstairs and open doors. It’s one of the most intelligent things we’ve ever built.

We’re trying to hold a certain standard for design and show what the technology is really capable of. We’re working through this problem now. It’s really hard. We have to get in a room with a lot of people and discuss where the future’s headed. Then we have to do a lot of product iteration. It’s something you can test through countless iterations, seeing how kids and adults interact with it.

Dave Blundin

But the end goal, at least on the consumer side, is to have a home companion that you trust and that’s part of the family.

Brett Adcock

Yes, 100%.

8. When Will Humanoid Robots Enter Our Homes?

You want to be able to talk to it like a human, and you want it to have persistent memory. You might want to spend a greater percentage of your time hanging out with the robot than having to do work. My view is that, in the home, you really want a companion that knows everyone and can be really helpful. A kid comes home from school sad, asking why. And then you want something that can do work, too. We all do. This morning, I was getting stuff out of the dishwasher and thinking, “I can’t believe I don’t have my robot doing that right now.” Hopefully, by the end of the year, we’ll be doing that pretty repeatedly. We’re working through all those things, but we want to be able to connect with real humans every day.

Peter Diamandis

On persistent memory, if you have a pet, like a dog, and it passes away, you get a new dog. But if you have an iPhone, when you get a new one, you just port everything over, and the personality moves to the new device.

Here comes your robot. You’ve fallen in love with it. It has persistent memory. You’ve done a ton of tuning, saying, “I don’t like my laundry folded that way,” or, “You shouldn’t be doing that. You should be doing this in the morning.” All of that state is now part of your bond with this device. Does that live with the device, or does it become like your iPhone, where you port the personality to the new one and the other one just—

Brett Adcock

It definitely does not live with the device. You’ll be able to have that pass through to any device.

Peter Diamandis

And what if you have two of them? Can you swap personalities midday?

Brett Adcock

Yeah, we’re doing some testing here. What if you’re walking by these hundreds of robots and they’re all talking to you?

Dave Blundin

And you’re walking by all the way through?

Brett Adcock

Yeah.

Dave Blundin

All the way through campus, having a singular conversation.

Peter Diamandis

That’s always mind-blowing. When you see them all lined up and they start doing things in synchronous lockstep, that’s mind-blowing for anyone who sees it, because you think of them as people. Then you see them coordinated in a way that people could never be coordinated.

The equivalent, when you’re walking through, is if you’re having a conversation and the conversation moves from robot to robot as you walk down the line. You’re like, “Wait a minute.” That could be good or bad, because of the uncanny—

Dave Blundin

The shared mind.

Peter Diamandis

Yeah.

When I think of these robots, I’d want to personalize them, both cognitively and physically. Right now, they’re all wearing the same outfit and costume. I assume you’ll be able to design your robots.

Dave Blundin

It’s like your iPhone case.

Peter Diamandis

Yeah, the outfit.

Dave Blundin

You’re going to want to design your robot.

Brett Adcock

Yeah, for sure. These robots can all be outfitted with different soft goods and apparel. We have some here we can show you on the way out. You’ll be able to fully accessorize your robot and give it whatever personality you want. Change its voice.

Peter Diamandis

What about tools? Are there going to be Figure-branded tools?

Brett Adcock

What do you mean by tools? Like a hammer?

Peter Diamandis

Well, like a hammer, but a mixing machine. I’m making a cake, and it’s like, okay, well—

Brett Adcock

That’d be pretty cool. Appliances.

Peter Diamandis

Appliances. I spend half my life in the longevity business. My favorite example is: If I’ve got a bunch of sensors in my body measuring my protein levels, hydration levels, peptide levels, and so forth, my robot in the kitchen should be cooking me a meal specifically for what my body needs at the moment.

9. Robot Safety, Security, and Trust

Brett Adcock

Yeah. The thing about tools in general, though, is that right now we think, “Oh, I’ve got all this human-made stuff. My humanoid hand controls it.” But now, for the first time ever, we have the ability to snap on an entire function. Just snap it in. Or I can mentally control it. If the device has your IQ from Hark or Helix built into it, I can just mentally control the device. I don’t actually have to walk across the room to mix the cake. I can just brain it over, because I can—

Peter Diamandis

Turn the oven on without even going to it.

Brett Adcock

Yeah, exactly. It becomes totally integrated into your whole life.

Peter Diamandis

Let’s talk about safety, because one of your goals for getting this out, as you said at the beginning, is that it’s got to be safe. Safety and insurance. How do you think about building the right safeguards into these robots? They’re very capable and powerful, so they can do what they need to do. What’s your philosophy there? What are you building into the system?

Brett Adcock

To make it really safe, you have to work all the way down to the substrate level, almost into nonvolatile memory, really understanding what those rules and affordances look like for the robot. Then you have to work all the way up to the system architecture design, a system safety architecture.

Brett Adcock

It’s really tough because you have both physical safety and what I’d call semantic safety. The robot’s intelligence needs to be pretty high, too. You don’t want a robot walking around and knocking over a candle or spilling hot water. You need to understand the risks in those areas and how to deal with them.

Separately, we have 40-plus actuators on board, and you don’t want any of them failing. You don’t want the robot falling, especially around kids or pets. We’ve spent a lot of time over the last 2 or 3 generations thinking about the right architecture for building a real-time system that doesn’t have faults and issues. We’ve seen this in the world: turbofan engines are rated for 1 × 10⁹ hours before a catastrophic event. There are very complex systems out there that have reached really high safety standards. We need to approach this like self-driving cars and other systems like them.

We see a clear path to designing a system that’s safe enough to scale to millions and billions of units. Beyond the system architecture, there needs to be a lot of iteration: build, test, learn, rebuild, and retest. Then you need to build a track record.

Peter Diamandis

Let’s talk about software. These robots run software, and software can get hacked. Do you have a master off switch that can be physically actuated? We’ve seen agents escaping the lab; it’s been all over the conversation on the podcast.

Dave Blundin

I was with Daniela Rus, who runs CSAIL at MIT, yesterday. They have a whole bunch of Chinese robots they bought because they’re really cool in the lobby, and they’re all transmitting data back to China. They’re recording video and audio.

Peter Diamandis

So I’m not worried about that for Brett, but—

Dave Blundin

It’s actually an advantage, in a sense.

Peter Diamandis

I am thinking about—

Brett Adcock

We buy all our competitors’ Chinese robots.

Peter Diamandis

Having a robot hacked in the home by some outside person. How do you put in cybersecurity or a physical off command, if needed? Is that something you’re thinking through?

Brett Adcock

We have a whole cybersecurity team. That’s all they do.

Peter Diamandis

Yeah.

Brett Adcock

Both product and corporate security. It’s actually a pretty large team, relative to our headcount here internally. Self-driving cars have the same problem: you don’t want all the cars on the street to be hacked and driven into pedestrians. It’s the same with us. We have safeguards in place, and we’re continuing to work on this. It’s an evolving thing. We don’t have a silver bullet where we push one line of code and say, “Yep, that will solve everything,” because it won’t. There’s a tremendous amount of work happening in cybersecurity and system safety engineering, all the way down to the fundamental rules we need to put into nonvolatile memory so the robot behaves properly in the world.

Peter Diamandis

When we met 9 months ago, you mentioned you were well down the road to having no Chinese or third-party components from outside the country. Have you made a lot of progress on that since then?

Brett Adcock

Yeah, we made a lot of progress over the last year or two on this. It happened because tariffs got so expensive.

Peter Diamandis

Yeah.

Brett Adcock

At one point last year, we were paying 270% tariffs for one week.

Peter Diamandis

It was during all that negotiation. I thought it was going to end soon.

Dave Blundin

That’s an incentive.

Brett Adcock

But it was not good. Then, obviously, the geopolitical situation really flared up. We’ve been rotating fully out of different piece parts we had from China. At this point, we’re in really good shape.

Peter Diamandis

Have you seen a robot insurance industry pop up yet? I’m assuming there will need to be insurance models wrapped around robots.

Brett Adcock

Yeah, it’s interesting. We haven’t seen any of that. Both commercial and consumer users will need this.

Peter Diamandis

In terms of financing the robots, are you going to provide financing and have outside financiers come in? Instead of buying one for $30,000, could I get a lease package?

10. Humanoid Robots in Space

Brett Adcock

Yeah, we currently have a lot of demand from people who want to finance robots for us. Right now, a lot of it is focused on commercial contracts. There’s no reason why this won’t happen in the home, too. This is a multitrillion-dollar space, like credit card receivables and lease financing for automotive, and we’re working with the biggest companies in the world on this. We’ll definitely work with different banks and funds to do it. I don’t know if we’ll finance them ourselves.

Peter Diamandis

Yeah.

Dave Blundin

Do you think about space at all? I know it seems like a stretch because we’re in the kitchen right now, but there are only 3 or maybe 4 people on the planet with the wherewithal to train a $3 billion foundation model for the physical world.

Peter Diamandis

Yeah.

Dave Blundin

Meanwhile, things are starting to go into space, not just on Elon’s rockets, but on other rockets, too. If you think about Eric Schmidt launching things into space that need to be maintained or changed, he doesn’t have that foundation model for the physical world.

Peter Diamandis

So there are probably these guys in space, but there’s also my model controlling something else.

Dave Blundin

Yeah.

Peter Diamandis

But that physical knowledge model is an insanely valuable centerpiece.

Brett Adcock

Yeah. It’s funny you asked. This AI stuff really happened fast. When I started Figure, we called it Figure AI. Our mission was to advance AI for robotics. My whole master plan was that there would be a convergence of AI and robotics, and it would be really powerful. I thought maybe that would happen 10 or 20 years later. It happened in the next year.

Dave Blundin

And deep learning started working. We also talked about a key focus on trying to get to space.

Brett Adcock

It’s an area we’re really interested in internally. We’ve had a couple of conversations about it over the last 6 months. One group told us it costs tens of millions of dollars per hour to put a human in space to do work.

Peter Diamandis

Yeah.

Brett Adcock

We were like, “Wow, that’s crazy.”

Peter Diamandis

It’s really bad for PR when things go wrong, too.

Brett Adcock

Totally.

Dave Blundin

I think a humanoid should be great at this.

Brett Adcock

Exactly.

Peter Diamandis

That’s the whole objection to data centers and other things in space: maintenance. Throw a humanoid out there to do it.

Brett Adcock

Yeah, exactly. And how cool would it be to float around?

Dave Blundin

We had a space team here, and they asked, “Are you IP-rated for this and this and this for space?” I was like, “Oh my God, no.”

Peter Diamandis

Vacuum-rated joints.

Brett Adcock

Yeah, exactly. They said, “I’ve got a lot of work.” I’ll be candid: we are not ready.

Peter Diamandis

You very famously said you’re not going to do defense work with Figure. Is that still the case?

Brett Adcock

Yeah, it is.

Peter Diamandis

I respect that. I think it would be difficult to do defense work while also trying to sell robots into people’s homes and into everyday commercial places. I think it just becomes really tough for humanoids.

Brett Adcock

I think you need to pick a route. We’re picking the civilian side, which we think is thousands of times bigger than military applications. It’s also an area that all of us here at the company really want to solve. We want to solve the robot in the home.

Peter Diamandis

I’m curious. When Bird scooters started showing up in my hometown of Santa Monica, people started trashing them, hanging them from trees, and so forth. When we start to see robots walking around the streets, how do you keep them safe from humans?

Brett Adcock

Yeah. I guess we don’t want to be slapping humans, right?

Peter Diamandis

Well, it’s more the question. There’s a significant amount of civil fear, distrust, and anger toward technology. You saw the Waymos burning in L.A. about 18 months ago. I can imagine the robots themselves need to be kept safe from humans.

Brett Adcock

We’re certainly not going to train any robots to defend themselves against humans. To the extent that a human wants to trash our robot, there’s really not much we can do.

Dave Blundin

A lot of cameras on that robot.

Brett Adcock

I know, but we wouldn’t want to try to defend ourselves or stop someone in any way. I think it just is what it is. You’re going to have some of that, but I think it’s probably a very small percentage of what happens. You see this way, too, but I think it’s something we’re going to have to build through.

Dave Blundin

Yeah.

Brett Adcock

We haven’t had it happen since. We’re a little worried: is a blue van going to drive up and snag one of our guys outside?

Dave Blundin

Yeah, but again, take it apart.

Brett Adcock

Yeah, it’s funny.

Peter Diamandis

It’s really hard to imagine anyone getting away with that more than once. There are so many cameras. It’s all connected.

Dave Blundin

Really funny. The robot’s in the back saying, “Listen, man, you don’t want to do this. You don’t want to do this on camera.”

Brett Adcock

I have a 2-kilowatt-hour battery pack. I can light it on fire and take this whole thing down. Don’t test me back here.

Dave Blundin

Brett’s not going to pay a ransom fee.

Brett Adcock

Yeah, he won’t pay. I’ve already asked him before. Don’t do this.

Peter Diamandis

Policy.

11. When Will Robots Build Other Robots?

Dave Blundin

It’d be really funny, just trying to talk him out of it. Can we talk about exponentials for a minute here? Peter’s favorite topic, actually.

Peter Diamandis

Go for it.

Dave Blundin

BotQ—are we allowed to talk about BotQ?

Brett Adcock

Well, what do you want to talk about?

Dave Blundin

I want to know the exponential self-improvement equivalent going on right now. RSI in the physical world is the coolest thing ever. You know, von Neumann probes. It feels like they’re very close to the point where they don’t build themselves by winding the actuator, but they control the machine that’s winding the actuator, and then it’s a closed loop.

Brett Adcock

I’ve always described it this way. My takeaway from the Gigafactory is that it’s not robots building robots. It’s robots controlling machines that build robots that control machines. But it’s the same effect.

Dave Blundin

It’s a closed loop, just exponential takeoff of the volume. You can see it. When we were here before, there were a couple dozen, and now there’s massive scale going on. There are robots everywhere.

Brett Adcock

And you said you slowed it down, too.

Dave Blundin

So when do we hit RSI in robotics? When do we hit robots building robots? What’s the time frame for that? What’s the very first robot where no human labor went into it at all? It was purely—full lights-out manufacturing, man. It feels like we’re so close.

Brett Adcock

We’re still making a lot. We have a lot of robots coming off the line. We’re in the process of scaling back up as we go into December and January.

Dave Blundin

When do we hit RSI and robotics? When do we hit robots building robots? What’s the time frame for that?

Peter Diamandis

What’s the very first robot where no human labor went into it at all? It was purely—

Dave Blundin

Yeah, full lights-out manufacturing. It feels like we’re so close.

Brett Adcock

The thing is, every quarter that goes by, the progress we’re seeing here is astronomical. I have to be real with you guys: when I started Figure, I was so excited about being able to work on the humanoid problem. I thought it was probably one of the most important businesses I could be building. I was going to do this for 30, 40, 50 years—just the rest of my life. I didn’t expect that in 2024 or 2025 we’d see all these breakthroughs and be on a full exponential.

We got really lucky. Deep learning showed up a year later in full force. The robot hardware got really capable. It’s now doing work with neural nets fully end to end, like we’re seeing out there. It’s nuts, and it’s getting better and better.

We just took a robot down to a client site it had never been to before, about three weeks ago. It had never been to that warehouse or anything else, and it zero-shot the whole thing.

Dave Blundin

Yeah.

Brett Adcock

Before, if the lighting changed, the conveyor was different, or the package mix changed, the robot would get confused. It was hard, but it made sense: it had never seen any of that stuff before. We did larger-scale pre-training with Helix 2.5, took that model down to the client site, and it zero-shot the whole thing. If you’d asked me a year ago whether that would be possible a year later, I would have said it would be incredible if we did it, but certainly in the next five years. Everything is improving so fast.

We’re seeing the work we’re doing on the hand for Figure 4 and on Figure 4 itself. We’re seeing the work on Helix 3, the amount of data coming online, and the amount of compute coming online. We have several billion dollars of cash on the balance sheet that we’re working to deploy.

Dave Blundin

Congratulations for that.

Brett Adcock

Thanks. It seems certainly within five years. It still feels certain that within five years we’ll have full lights-out manufacturing. It’s foggy whether it’ll happen in 6 months, 1 year, 2 years, or 3 years, but it’s right there.

Peter Diamandis

Lights-out manufacturing means there are no humans there and no lights are needed?

Brett Adcock

Full lights-out manufacturing, as we call it. We designed Figure 4 to help with this problem. There were a bunch of things in the Figure 3 design that we felt were really tough for full closed-loop, lights-out manufacturing, so we designed them into Figure 4. That’s why we’re saying it’s an iPhone 1 moment. We think this is the robot that will build itself.

Dave Blundin

Let’s flip it around and ask the other way. When the new hand comes along, which is imminent, is there anything in the current manufacturing process—someone controlling a milling machine, or whatever component—that you can think of that won’t be able to be done by that new hand? What’s the last blue-collar job opportunity that somebody needs?

Brett Adcock

Not really. I might have said ladders and some of this forceful interaction stuff would be hard, but the robot’s climbing a ladder next. It’ll just climb over and over every day. We’re doing forceful interaction work now and carrying big boxes around—50 pounds—with the neural network. We’re unloading things in a truck. We’re doing things in a home, like dishes and folding laundry. All the puzzle pieces are right there.

Peter Diamandis

The ultimate functionality for me is when a robot is doing surgery. When is a humanoid robot the best surgeon on the planet? The timeline is probably lights-out improvement to that moment.

I mean, we’ve seen AI become the best diagnostician. The numbers are blindingly obvious: AI versus humans, AI wins every time right now. The next question is when we have a robotic surgeon. With federated learning, every robot knows every surgical case ever done. It doesn’t have any jitters. It didn’t have a fight with its boyfriend or girlfriend the night before. It’s there and able to do surgery. What’s fascinating is it’s a dollar an hour for the surgery.

Dave Blundin

Right. Any place on the planet.

Peter Diamandis

I also feel like half the time I’ve visited a doctor in my life, I’ve been misdiagnosed.

Dave Blundin

Yeah.

Peter Diamandis

Having a network of the smartest thing in the world with you, there in a room checking you out, would be incredible. It’d be the world’s best doctor for every country in the world.

Dave Blundin

Yeah.

Brett Adcock

We’re there now. The question is when these robots are the best surgeons in the world—when you need something and say, “I do not want that doctor touching me.”

Peter Diamandis

I want the robot to do it perfectly every time.

Brett Adcock

Yeah.

Peter Diamandis

I think I want to capture this, though. That hand comes out imminently, and we can demonstrate every single step of the manufacturing of the Figure 4 being done with that hand. That’s a big moment in human history. I think we might want to document that, literally every step of this entire process. Here’s something that arrives in a box: open, assemble, build.

Brett Adcock

Yeah. I mean, guys, here’s what I see. We see a task and pick out something we need to be able to do. We get data for it, train Helix, and then the robot goes and does it.

Peter Diamandis

Yeah.

Brett Adcock

We’ve done that over and over for the last couple of years, and we’re like, “Man, these robots are really capable.” How do we get them to generalize? That’s the real thing. How do we get one to walk into Baku with a minimal set of instructions, like whatever training a human would get—probably some slides and documents, and maybe shadowing somebody for a few days?

We want the humanoids to be able to go through that same process. That’s really the chapter we’re in right now: how do we scale intelligence? How do we get this thing to do it? From a hardware perspective, the hardware is going to get as dexterous as human hands. This is going to keep improving at a steady rate.

If it has a hard time handling a needle one year, the next year it’ll get better and better at it. This is all happening really fast. This is single-digit years.

Peter Diamandis

I think that’s actually a much higher bar, though, than the moment in history when lights-out manufacturing needs to demonstrate: “Here are all the steps of creating a Figure 4. Pick any one at random. This robot with this hand can do that step.”

That’s a moment in history that’s actually sooner and easier than giving it an open-ended task it’s never even thought of before and having it figure out how to do it. That’s a much higher bar, but it’s probably only six months later.

Brett Adcock

The one thing I’ve learned is that these neural networks love data.

Peter Diamandis

They love to learn.

Brett Adcock

They love to learn. They love the variety of data coming in. I would have said that even on this logistics line we’re seeing behind us now, it’s running and the packages are moving. They’re squishy and moving on the conveyor. In a lot of cases, the robots haven’t seen these types of packages before, and for whatever reason they’re just good at learning these representations during training. I wouldn’t have thought that would be the case.

We’re going out now and doing large-scale data collection, bringing in more diverse data. The robots are getting smarter at things they haven’t even been trained on. We’re going out and getting data—some guy has a hose and is washing down his dog; somebody else is painting a house. We take data like this, and the robot gets better at folding laundry.

Peter Diamandis

Yeah.

Brett Adcock

It’s crazy. This data diversity—we’ve seen this with LLMs and pretraining—is so healthy for the neural net to learn these representations across our world.

Peter Diamandis

Do you think—I don’t know if you can quantify this—how much is general-purpose physics, physical-world knowledge, and how much is, “Okay, if I want to train you to do this particular type of surgery, then I need exact data from just that surgery”? When people automate these things, they think entirely down these narrow swim lanes, but we’ve never had the general-purpose neural net before.

Brett Adcock

Well, we’ve been on the narrow swim lane for two years. That’s what we do. We’ve been doing that all along. We’re arguably probably the best in the world at post-training these narrow use cases.

What you really want from a humanoid is for it to adapt to things really quickly, things it’s never seen before. In the case of, say, BotQ or even surgery, you want very minimal demonstrations for the robot to be able to do the work. That means the prior—the robot coming into it without any surgery training—should be that it has a decent understanding of what it should be doing and how to do it, just with a prompt. Then, with more specifics—“I need you to do this task, then go here and do this, and watch me; this needs to happen this way”—the robot should be able to do the work.

It needs to have a good base understanding of how to do this stuff. A human walks in to get prepared for what their first day at Baku looks like. They’re a really capable human. They know a little bit from verbal prompting or from being shown how to do things, and they understand and latch on to the concept.

For us, it’s really not about these narrow peaks anymore. It’s about the broad-level intelligence we have in LLMs. From there, the robots can be fine-tuned to a point where they’re the best in surgery, logistics, or some specialized task. But they should be generalizable enough to do most things. The surgery bot coming in should be able to go grab you a coffee really quick without any training, but it might need some specific training for that surgery.

Peter Diamandis

Coffee bot—my favorite. When we last heard, you were just gearing up to build your manufacturing facility. What’s the status on that?

Brett Adcock

Yeah, we’re working on that now. We haven’t announced anything publicly. The current Baku capacity, in my estimates, won’t really hold for the future. We can do maybe over 40,000 units at full rate at Baku next door a year, across four lines. If we can crack these next few steps, I think we’re looking at potentially unlimited demand.

Peter Diamandis

I can imagine that. Do you think Figure 4 is going to be the first high-mass-production robot? Any projections for 2030, 2035, 2040? I mean, the demand is near infinite. I buy that. What are the numbers you dream about?

Brett Adcock

We see a path where, if we can solve some of the steps we’re working on now—it’s a really hard path—we could scale to a million units a year out of a facility fairly quickly. But that depends on us being able to crack some of these Chapter 3 problems we’re working on now.

Peter Diamandis

I remember when you and I had this conversation and I asked, “Do you agree with Elon’s 2040 prediction of more robots on Earth than humans?” I think you agreed with that potential size of the marketplace.

Brett Adcock

I’m fairly confident we’re going to have a lot more robots than humans.

Peter Diamandis

Yeah.

Brett Adcock

Yeah. You’re like, even if you size it at 3 billion—3.2 billion humans in the commercial workforce—that’ll probably scale to 5 billion to 8 billion, maybe 10 billion. Then every human will want at least one, so you’re at roughly 8 billion there. You’re roughly on the order of 20 billion humanoids.

They’re going to be everywhere, doing all kinds of things. They’re going to be a dollar an hour, as you said. You’re just going to have them waiting if there’s more work to do. One thing we’re seeing now that’s going to help with this exponential is that companies are all set up around the human workforce. Even if we can go in and automate a certain package line or part of manufacturing, we’re still bottlenecked by the other human processes that run on certain shifts. The truck doesn’t come in at certain times, so it’s not like we can keep working. There are literally no packages to move anymore.

If you’re working 24/7, you’re left with this period where you come in and do real useful work, and another period where you come in and fully automate an entire company end to end and work 24/7.

Peter Diamandis

Yeah.

Brett Adcock

That’s just another step forward. We haven’t even seen that in the world. I think the world could hold 20 billion humanoids just in the first phase of that. But when entire logistics and supply chain companies, healthcare, and data centers can run fully 24/7 with lights-out manufacturing, you’re in a different place. You’re basically building entire economies that run 24/7. I think we’ll live in a very different future.

This goes to the point of the conversation you and I had with Elon about UHI, when robots are actually able to massively demonetize everything—building your home, doing whatever you need—at that dollar-an-hour price tag.

Brett Adcock

It creates abundance for everybody on the planet, which is a beautiful thing.

Peter Diamandis

Yeah. It’s pretty deflationary. It’s going to be great, and it’s going to raise wealth everywhere. We had a late-night conversation, and you said something that was a powerful idea. I want to come back to it. You said there’s an ethical, moral issue: as AI becomes more capable and we reach ASI, it will need things done. You don’t want ASI telling humans to do those things. You want ASI telling robots. Do you remember that conversation we had?

Brett Adcock

Yeah.

Peter Diamandis

Do you want to expand on that? I thought it was a really important point.

Brett Adcock

Yeah. I think robotics is really necessary to avoid a dystopian future. I feel like there’s a question right now about whether we’re at an AGI or ASI moment. Digital AI is so smart, but it’s also not very magical. It feels almost like a search engine replacement, and it doesn’t feel like AGI or ASI in my mind. It doesn’t feel like the Jetsons right now. It just feels fine. It can replace my Google searches.

As this gets more and more intelligent, these digital systems living in a server will need to move things around to make money and help more in the economy. There’s a future where they’re bossing humans around to do this work, paying them to go do it for them.

Peter Diamandis

Yeah. Almost like enslaving humanity.

Brett Adcock

Well, we have a long history of enslaving each other.

Peter Diamandis

And it’s kind of an outcome.

Brett Adcock

There are already some companies popping up that are asking humans to go do work and paying them by the hour to let agents go out and get work done. In my mind, if there’s no robotic solution here, that could lead to a really bad place.

So my view is that as we build toward AGI and ASI, humanoids should be doing this work, not humans. I think it’s a really important part of the story, and it’s all coming together really fast.

Peter Diamandis

Yeah, I haven’t heard anybody speak about that. I think it’s a really critical point you made. To avoid that dystopian future, no human wants to be told by a machine what to do.

Brett Adcock

Yeah.

Dave Blundin

Peter, I think it’s happening. There are websites you can go to now.

Peter Diamandis

Of course. Mechanical Turk.

Dave Blundin

Yeah. The agent needs to hire somebody to go open a bank account or do something for it.

Peter Diamandis

100%. There are many TaskRabbit-type services up now that are doing this. They’re getting real traction.

Dave Blundin

There’s another type of dystopia that I think we’re on a path to avoid because of Figure. I don’t know if you ever saw any of those Russian cars back in the Soviet era, but everybody in the country had to drive the exact same color piece of Soviet car. It was one form that fit everybody.

When you look at robotics in manufacturing today, like a Honda Accord assembly line, it’s “this robot is welding, it’s welding, it’s welding, it’s welding,” and what comes out the end of the factory is an identical Honda Accord for everybody, just in a different color.

Peter Diamandis

Yeah.

Brett Adcock

Maybe a couple of configuration differences. But now LLMs are capable of mechanical design for the first time, which means somebody with an idea for a new physical device—a gutter cleaner, a pool cleaner, a companion, whatever—can vibe-design it, but they can’t make it.

Unless all the parts are 3D-printed or something like that, it’ll never come into the world. All of a sudden, Figure Robotics comes along, capable of making itself, which means it’s this far away from being able to make virtually anything. Then you’re in a world with a massively bigger variety of physical things.

Dave Blundin

Because it’s a completely different supply chain. Anything that can be done can be done very cheaply, fully automated.

Brett Adcock

100%.

Peter Diamandis

Yeah. They’re just going to work their butts off. All they’re going to do is work.

Brett Adcock

All they’re going to do is work. We’ve literally had that robot running since the logistics livestream. It runs every day. It’s running right now; it just hasn’t stopped.

Dave Blundin

Instead of a webcam at a fish tank, you’ve got webcams on robots doing packages there.

Peter Diamandis

Hey, can I ask you about chips?

Brett Adcock

Yeah, of course.

Peter Diamandis

Chips. When you look at it, suppose there are 10 billion robots or more, and suppose Figure is a huge fraction of them, maybe a dominant fraction. The compute behind them wants to get smarter and smarter and smarter. Elon is looking at all the supply-chain components, and he’s like, “Terafab—this is the one dependency I can eliminate. All the other dependencies have a roadmap.”

If we were talking 9 months ago, the idea that you’d have the capital to take on chips—which is about a $16 billion problem—might not have been viable. But now you’ve got billions of dollars of cash, so suddenly it is. Are you going to go down that road? Is it something you’re talking about?

Dave Blundin

How far down the rabbit hole are you going to vertically integrate?

Brett Adcock

Yeah, I suspect at some point we’ll—

Dave Blundin

Yeah. I mean, the question is whether we can control our destiny well enough with the supply that will exist over the next 5 or 10 years.

Brett Adcock

Memory, silicon, and even the cost per chip and the power we’re getting. The chips we have now run on robots, but they’re still too costly and use too much power.

Peter Diamandis

Mm-hmm.

Brett Adcock

If you really want an affordable robot, you’re going to have to spend a lot of time in there.

Dave Blundin

Yeah.

Peter Diamandis

Well, also, remember you were talking about the actuator that comes out of the box this big?

Brett Adcock

Yeah.

Dave Blundin

The people designing chips right now—

Peter Diamandis

They’re sold out for decades.

Dave Blundin

Yeah.

Brett Adcock

They’re not going to start thinking about innovating in the direction you need for robotics. They’re just going to try to put up the cheapest, fastest fab they can and spew out what they can, and you just have to live with it.

We actually met with the C-suite of one of the largest memory OEMs in the world a few months back. I was trying to talk them through our future and roadmap, and they said, “We have no capacity for this.” All of their capacity is going to phones and servers and everything else. Even if a phone doubles its memory, it takes the whole capacity of the whole market.

Peter Diamandis

And the phone—

Dave Blundin

Phones are getting choked off. You think about the amount of margin in a phone—it’s insane—and it can’t even get the supply.

Peter Diamandis

Tim was complaining about that. His earnings call said, “Memory is destroying us.”

Brett Adcock

Memory is bad. We’ve done a few pretty cool memory partnerships to secure memory over the next few years internally. In the near term, we feel very confident in our ability to scale up robots. You’re asking a lot of good questions about the long term.

As you mentioned, it’s a huge capital investment. The technologies, the models themselves, and the way we’re building and running the models are evolving fast.

Peter Diamandis

Yeah.

Brett Adcock

One concern I have is that this chip work takes 2 or 3-plus years. You have to look out at least 3 years and ask, “What is the design going to look like?” If you’d asked me 3 years ago what my inference compute design would look like today, it would have been completely wrong.

Peter Diamandis

Yeah.

Brett Adcock

We have some work we’re doing now with Helix 3, which you guys just saw. It’s a radically different way of running things than we’re used to. It’s a little hard to lock us into a certain amount of constrained power, cost, and chip architecture, given how quickly things are accelerating.

It’s been nice to be able to move to the right chip for us now. But if we look out over a long enough period of time, we’re going to need access to an unprecedented amount of silicon.

Peter Diamandis

Well, there’s a pro tip. If you think 3 to 4 years out, that’s the earliest you could get your own chips.

Brett Adcock

Yeah.

Peter Diamandis

Between here and there, Positron, Etched, and Groq are making incredible headway on 6-nanometer or thereabouts process nodes that are still available. But you have to race right now to book that capacity.

A 6-nanometer node directed toward your use case is dramatically better than the 2- and 3-nanometer nodes used for all the iPhones and Nvidia GPUs. They can’t get any headway out of the 6s.

Brett Adcock

Yeah.

Peter Diamandis

But that’s the stepping stone that fills the gap.

Brett Adcock

Yeah.

Peter Diamandis

But those conversations are moving really fast. It’s funny: in the physical world, in the entrepreneurship world—

Brett Adcock

Yeah.

Peter Diamandis

There are only two or maybe three of you.

Brett Adcock

It’s a huge bar, but someone has to fill that void. Otherwise, we’re all going to suffer for lack of robots for some period.

Peter Diamandis

How old are you now, Brett?

Brett Adcock

40.

Peter Diamandis

40. So, congrats. I think—

Brett Adcock

I made it to 40.

Peter Diamandis

Yeah. No, I think the congrats is that you’ve built serially more valuable companies, which is extraordinary. I think you made the point earlier—the stat earlier. What was it?

Dave Blundin

It’s staggering. All entrepreneurs, all time, only Zuck is ahead of you at age 40. But among serial entrepreneurs—

Brett Adcock

Zuck, goddamn it. Well, it gets better.

Peter Diamandis

But you’re a nicer guy.

Dave Blundin

Serial entrepreneurs, which is basically Jobs, Elon, and a handful of others, are far more likely to change the world in the future. That totally makes sense to me, having invested in hundreds of entrepreneurs over the years. If you can do it once or twice—

Brett Adcock

You’re going to do it three, four, five times. If you do it once, it might have been a lottery ticket. It might have just been an outlier. It doesn’t mean you can do it again. But you’ve already proven you can do it about four times now, I guess.

Dave Blundin

So the forward trajectory of that profile is so much better statistically—not to put too much pressure on you. Among serial entrepreneurs at age 40, you’re 20 times ahead of where Elon or Jobs was in terms of the resources you have available. Let me put it that way. Part of that is just a sign of the times: there’s a lot more capital available today.

We were talking at breakfast this morning about how I think you were destined to be a physical-world guy. It feels like it’s in your Illinois farm DNA: machinery.

Brett Adcock

I like touching hardware.

12. Competing with China and Tesla

Peter Diamandis

Yeah. If you were born 20 years earlier, you would have run headlong into “hardware is hard” and “Silicon Valley is never going to invest in it.” The moment is so perfect right now. You’re just hitting it.

Brett Adcock

Yeah.

Peter Diamandis

It’s going to be so fun to watch where this goes. Let’s talk regulation. There’s so much conversation around AI regulation, but I haven’t seen that in the robotics space. I’ve seen the U.S. government putting a regulatory block on importing Chinese robots.

Brett Adcock

Yeah, the FCC.

Peter Diamandis

I get it, but are you starting to have conversations with regulators, or hearing from them, about regulating robots? That they shouldn’t be on the street, or that we need to be careful about them in homes? Where is that space right now, and what are you gearing up for?

Brett Adcock

I think the FCC ban on Chinese robots was a bit of a shock. From my perspective, I’d love to see a really level playing field to compete.

Peter Diamandis

As you think about competition—I’m going to come back to regulation in a second—is it Tesla and China? Is that how you see the world right now?

Brett Adcock

I personally haven’t seen much out of Tesla in two or three years. I’ve seen some teleoperated robots, but I haven’t seen any—

Peter Diamandis

You haven’t seen this, right? Have you?

Brett Adcock

No, I haven’t.

Peter Diamandis

Okay. I’ve seen China do a lot of dancing. We’ve talked about this before: the neural network architecture is literally 180 degrees different. If you said, “Brett, we need you to outdance China,” they do a damn good job. We’d have to pivot everybody’s strategy. We don’t have a dancing roadmap and go dance. It’s literally 180 degrees opposite work. We’d need a bunch of humans wearing Hollywood suits dancing around, then we’d need to work on replaying that on a robot. We’d need to get out in the middle of the woods, go around, and leap over rivers.

Brett Adcock

Yeah, Boston Dynamics was making the best robot.

Peter Diamandis

I think if I answer your question more directly, the closest competitor to Figure today is certainly China. There’s just a lot of work happening there. They’re moving quite fast. I view them as moving in a slightly different direction than we are.

Dave Blundin

And there are a lot of them.

Peter Diamandis

There are a lot of them, yeah.

Dave Blundin

There are about 150 humanoid robot companies.

Peter Diamandis

Yeah. It’s kind of all crap right now.

Brett Adcock

The hardware isn’t great, and the policies they’re running and the models don’t look that great. But the amount of progress they’re making year to year is really great. I’d expect China to continue to do much better. It’s a group we intend to beat.

Peter Diamandis

Sure. China needs robots desperately.

Brett Adcock

Yeah, really badly.

Peter Diamandis

For manufacturing, elder care, and all of that. Are there any particular Chinese robot companies you think are doing a better job than others?

Brett Adcock

Not really. We care about something very different from what we think China cares about. China is almost in a race to show how capable its robots are at gymnastics. We care about how to ship a robot into a home—

Peter Diamandis

Yeah.

Brett Adcock

—to do zero-shot, long-horizon reasoning.

Peter Diamandis

And you’re not seeing that out of China?

Brett Adcock

You can’t even—you buy these groups’ robots, even Unitree’s, and they don’t build good hardware. They’re super tiny. They come out of the box, they wave with a controller, and you’re like, “What else do I do with it?” It’s unclear. They have no hands, and they have compute the size of a USB stick. You basically can’t run any real models on them.

Seriously, all the groups that are doing research put compute backpacks on them. The sensors aren’t great. The hardware is basically duct-taped together. It wasn’t manufactured well. Will they build better robots over time? Certainly. They’re going to build great robots over time. We just have to beat them.

Peter Diamandis

Mm-hmm.

Brett Adcock

It’s a real race. We’re spending all our time and energy thinking about where we want to head long term, and we’re trying to move that left on the time curve as far and as fast as we possibly can.

Every morning I come in thinking about how to move faster, because there’s always a way to move faster. For entrepreneurs out there, hear this: there is always a way to move faster, and speed and agility are your—

Peter Diamandis

You can literally always move faster, which is quite an unbelievable thing to have access to.

Brett Adcock

You can even just work harder. If you work double the amount of hours, you’ll probably get an order of magnitude like 1.5, maybe even 2 times more speed per year. All of that really matters as we compete against the global collective of China.

Peter Diamandis

You have the advantage of being a founder-led company versus a large industrial company trying to get into the space.

You’re willing to take bigger risks than almost anybody else. You’re looking at disruptive strategies and approaches and iterating rapidly.

Brett Adcock

Well, and also, when you walk around the building here, the talent level—you’ll never find it at a big industrial company.

Peter Diamandis

I mean, yeah, there are a bunch of small things here. I think the only way to build really capable, general-purpose robots is in the Bay Area. We practically invented robotics.

Brett Adcock

The Shenzhen of robotics?

Peter Diamandis

Yeah, that’s better than Shenzhen. We built robotics here. AI was literally invented here. Silicon was built here. Batteries, motors, embedded software, thermal systems, AI, deep learning—all of that infrastructure was built here. This is the best place in the world to do this.

Self-driving cars were here before they were in China. AI was here before it was in China. The internet—these are all things that were born here. People who grew up here have worked in these areas and really understand the technologies. We’ve assembled what I think is a top-tier group of them. The whole team is a bunch of maniacs. They’re really great at what they do, they love being here, and they’re mission-driven. We’re all here to beat everybody in the space.

Dave Blundin

What was your journey to Silicon Valley? You grew up in farm country in Illinois, right? Then you went to college in Florida.

Brett Adcock

Yeah, I went to college in Florida. I moved to New York City after that, and then I started a software company called Vettery. We ended up selling it in 2017.

Dave Blundin

And you were still in New York?

Brett Adcock

Yeah, it was in New York. It was basically an online recruiting marketplace. We had a two-sided marketplace to bring on employers and job seekers and help with recruiting. We ended up selling to the world’s largest recruiting company in 2017 for a little over $100 million.

As soon as I got the term sheet for that, I thought, “I need to…” I had a few things I really wanted to do in the world, and a few of them were in hardware. I ended up starting Archer Aviation the next year.

Dave Blundin

Did you start it here or in New York?

Brett Adcock

I started in New York with the intention of moving. At first, I was going to move to LA because some of the best aerospace people in the world are there.

Then I found out it wasn’t really where the work I was doing at Archer was happening. Building electric aircraft is basically building robots: electric motors, battery systems, control software, and so on. That wasn’t what aerospace companies were building. It was all happening in the Bay Area. You had groups out here building electric vertical takeoff and landing aircraft, like Kittyhawk and Airbus, and you had this whole electrification community.

You also had the robotics community, so I moved here. I’ve been here ever since, and it’s been great. Archer is basically a flying robot.

Dave Blundin

Yeah.

Brett Adcock

It’s got electric motors, batteries, control systems, embedded software, and sensors. It’s a flying robot, so doing robotics out here is arguably the best place in the world. Maybe second to—

Peter Diamandis

I think number one has got to be the Bay Area. Maybe number two is Shenzhen. If general-purpose robotics is going to be solved, you’ll see it in one of those two places.

Dave Blundin

I totally agree. It’s amazing. If you’d stayed in New York or moved to LA, today, with $9 trillion of new liquidity from Anthropic’s IPO, Nvidia’s run-up, and everything, if you woke up tomorrow morning and said, “We were on this podcast yesterday. I think we need to start building our own chip fab”—

Brett Adcock

I need $20 billion.

Dave Blundin

Yeah. You can do that here.

Brett Adcock

Yeah. You can’t even come close to doing it in New York or LA.

Peter Diamandis

This is the new capital—

Dave Blundin

In all those ways.

Brett Adcock

I was in New York for a long time. I love New York. It’s one of the best places I ever lived. I used to think, “What’s this deal with SF? I’m in New York City. It’s probably the number two market in the US for software engineering, besides SF.”

I didn’t really get this SF thing. Then I moved out here for Archer, and man, do I get it now. It’s the most concentrated place in the world for talent, and they’re all at the right cultural phase of life to come over here and really punch it.

Peter Diamandis

It’s actually the easiest place to import talent. If you go back to Chicago or Boston, there’s huge talent—

Brett Adcock

But you can migrate anybody here. If you’re in Austin, Chicago, or Boston trying to do robotics, you’ll get some folks to move there from SF—maybe one or two out of 50. But here, you can get 50 out of 50.

Peter Diamandis

Yeah.

Brett Adcock

In robotics, there are about 12 different divisions of people you need to build one robot. Good luck getting those 12 people, across those 12 groups, to move from SF. It’s just not going to happen. It’s too difficult.

Peter Diamandis

When you moved out, everything you said was true, except that it wasn’t yet the financial capital of the world. Now, with all the new liquidity, it’s all that plus the financial capital—with Sand Hill Road.

Speaking of that, you took Archer public for about $2.7 billion. Peggy Johnson just took Agility public. When is the right time for a company like Figure to go public? What do you need, in your mind, having gone through a public offering?

Brett Adcock

Well, I think Peggy went public because she—

Peter Diamandis

Yeah, she needed capital.

Brett Adcock

Yeah, of course. I understand that. I’ve known you for years, and I’m not going to speak for her, but what are the criteria, having learned what you did with Archer?

Dave Blundin

Yeah.

Brett Adcock

At a high level, Archer went public because we needed the money. We were raising our Series B, and we needed hundreds of millions. We needed a few billion dollars to get through FAA certification, which is very difficult to do. I just wasn’t seeing it in the private markets. We ended up going public and raising a little under a billion dollars.

It was the right move at the time, but we were really early. We were just a really young company, so we were too young to be public at that stage. But I needed the capital. I took the company public, and it was a choice I would make again, given the constraints I had at the time. If we’d had access to the private markets, Archer would still probably be private to this day. That would have been healthiest for the business.

On the Figure side, I think it’s somewhat inevitable that we’ll be public at some point. There’s access to the right public capital, and public trust through disclosures and things like that.

Peter Diamandis

Totally. I think it’s great for employees, liquidity, and everything else for people and investors. For full disclosure to the audience here, I’m an investor in Figure.

Brett Adcock

Yeah.

Peter Diamandis

An early investor, too.

Brett Adcock

Early investor. Very proud of it.

Peter Diamandis

Before we were cool.

Brett Adcock

You know who I invested in? It was specifically you and the team you were able to assemble.

Peter Diamandis

Yeah. And a note for the entrepreneurs listening: the signal was so strong based on what you had done as a serial entrepreneur, but even stronger because you were putting your own capital into it. The team you had assembled was a rockstar team.

Brett Adcock

The team was lit.

Peter Diamandis

Still is. Still lit.

Brett Adcock

I couldn’t have recreated it again if I tried. It was just the right place, right time, and the team was unbelievable. It’s been great.

I think one of the big focuses for us right now is solving what we call general robotics. This is a general-purpose robot.

Peter Diamandis

We saw a little glimpse of what that looks like today, a little glimpse of the future, and I think it’ll get much, much better.

Brett Adcock

That will really kick off significant-scale manufacturing, and maybe even a larger increase in inference and compute as we scale. There’s also the working capital and capex needed to grow the business.

I think we really want to hit that milestone privately, and we feel like we’re approaching it. It feels like every new model we train, we’re asking, “Is it the one?” Then there are bugs we want to work on and all that stuff. I think we have 10 times more data sitting in our servers today that we can train a new model on than what you guys saw today.

Peter Diamandis

Wow.

Speaker 1

I think it’s literally—we have 100 times more compute coming online.

13. Training AI and the Compute Race

Brett Adcock

Sorry, I wish I could make it bigger. I think it’s maybe way more than 10, maybe a factor of 15 or so more data sitting there. I mean, we bring data in.

Speaker 1

Is this data captured from humans wearing cameras?

Brett Adcock

Yeah, it’s from a variety of sources. Basically, we need really rich information about the physical world that we can get from online and scrape online.

Speaker 2

You just can’t get it from the internet. We need to know how a human would grab this cup, how they’re oriented, what forces they feel, and what their hands are doing.

Speaker 1

The feedback—the video—is all over the internet, but the feeling in that feedback—

Brett Adcock

Yeah. Totally important. You can’t get it.

Speaker 2

Yeah, I can’t get it.

Brett Adcock

So we basically have an effort we call Index, which is a large-scale data collection effort. What I’m trying to say is, we have to audit that data. We have to make sure it’s the right distribution, clean it, annotate it, and get it ready for training. That takes time. When it’s ready for training, we have to be at the right place, ready for a new checkpoint for pre-training, and go through it. We can’t just instantly train on it.

Over the last month, more than half of our compute has been taken up by one giant pre-training run—a large percentage of it.

Speaker 1

But what’s the price tag on a big run for you?

Brett Adcock

We pay nine figures a year for compute right now.

Speaker 2

Yeah.

Brett Adcock

Low nine figures. Maybe 50 to 70% of our compute was pre-training.

Speaker 1

Okay.

Brett Adcock

A lot of nine figures a year is going into pre-training for Helix 3. It’s quite a lot. If we had 10 times more compute, we would train on 10 times more compute.

Speaker 1

Well, next summer you have 100 times more. Are you going to—

Speaker 2

Next summer—

Speaker 1

Are you going to do a billion-dollar training run?

Brett Adcock

Yeah, I think hopefully—I expect that next summer we’ll be doing billion-dollar training runs on Helix, if all goes well. Everything’s got to go well, but it looks like—

Speaker 1

You’ll have enough data?

Brett Adcock

We’ll have the data, we have the compute, we have the capital. Getting these larger clusters set up is hard.

We went out for compute six or seven months ago. That’s the current cluster we have now. Four vendors came back with pretty decent chips—new Blackwell-series chips—and we beat the piss out of them on pricing. We got the GPU per hour down; it was a great deal. We signed it, great, and it was going out.

A couple of months after that, maybe three or four months ago, we went out again. We needed to double our compute capacity. We just needed more for Helix 3. We went out and got no bids back. Then I thought, “We’ve got to push harder.” We went out again and got one bid back for a big cluster.

Speaker 2

Yeah.

Brett Adcock

And then they beat the piss out of us on price.

Speaker 1

Like it was just buyers and sellers.

Brett Adcock

I was like, “What just happened?” We just couldn’t get any more compute.

We met the NVIDIA team, and I was like, “Listen, everything’s at capacity. The whole world is just there.” I made this prediction earlier in the year: in 2027, you’re not going to be able to get compute anymore. That happened in September of 2026 for us. The world ran out of compute. You just can’t get it.

Peter Diamandis

I’m so glad you’re saying that. So many people are misinformed right now: “Oh, this AI thing is overblown. The bubble’s going to be chips.”

Brett Adcock

We can’t get H100s.

Peter Diamandis

Yeah.

Brett Adcock

We can’t get the old chips. They’re trading above par. We can’t get those. We can get—I had somebody write me yesterday, “I got compute.” I’m like, “Great, I’m out.” She’s like, “Eight nodes.”

Dave Blundin

Yeah.

Brett Adcock

I’m like, “Dude, eight nodes? I can’t even stop and talk to you. Eight nodes? I need more than eight nodes, man. It’s not going to work at all.”

The whole world is out of compute, and compute is constrained by memory. The whole world is out of memory. It takes three or four years to build those fabs from scratch. NVIDIA has been one of the best places to get it; you can’t train on anything else besides this. It’s a really interesting environment.

Peter Diamandis

You mentioned staying in the private markets longer, but a lot of the investment is coming from NVIDIA and other public companies. It’s actually coming from the public markets through large public companies, but they never used to invest in startups or early-stage companies, or even bigger, fast-growing companies. Now they’re just wide open for business.

Dave Blundin

That’s also a Silicon Valley kind of invention, which allows much bigger things to be built highly efficiently in the private market.

Peter Diamandis

Maybe a fab would take it to the next level, because you were saying a second ago the new HBM memory fabs won’t be online for three years.

Dave Blundin

Yeah. But if you look at how many are being built—

Peter Diamandis

They’ll be overbought before they even come online.

Dave Blundin

Yeah.

Peter Diamandis

Even if they come out on time, which they won’t. None of these are forecasting humanoids.

Dave Blundin

Yeah.

Peter Diamandis

They’re not even forecasting—

Brett Adcock

We literally met the CTO of one of the big groups, and they were like, “We don’t even have any of these in our projections.”

Peter Diamandis

Well, everywhere you go, whether it’s software for mechanical design or space or whatever, everyone’s like, “Yeah, they haven’t forecasted my demand yet.”

Dave Blundin

Yeah.

Peter Diamandis

We’ve had enough of these conversations. You stitch it all together, and you’re like, “Oh my God, we’re like—”

Brett Adcock

If you look at the scaling laws for compute, too, literally the same amount of data and the same parameter models being trained on more compute has a log-linear relationship. You’re seeing real ROI there.

Dave Blundin

Yeah.

Brett Adcock

I would say, man, it doesn’t seem like this trend is anywhere close to stopping.

Peter Diamandis

Yeah. I think we’re really early in what’s coming out of this, both digital systems—say, doing a heart—and even Figure stuff. It’s just so magical. This stuff is going to be 100 times bigger than the internet. It’s going to be so important—

Dave Blundin

A thousand times bigger.

Brett Adcock

I think we’re still really early. It’s not time to panic yet, so keep calm and keep building.

Peter Diamandis

I’m a Hark Pro user, and I’d love to ask you: what’s the relationship between Hark and Figure? Is Hark going to power part of the Figure ecosystem?

Brett Adcock

Yeah. Maybe I’ll give you a quick overview of both and what we’re doing. Hark’s on campus here, 100 feet away.

Peter Diamandis

It’s nice. You have your children next to each other.

Brett Adcock

I like it. I don’t have to travel and can always be around and help wherever I can.

Hark is basically trying to build personalized intelligence, almost like the digital version of what we’re trying to do at Figure.

Peter Diamandis

Right.

14. Hark AI, New Devices, and the Future of Human-AI Interaction

Brett Adcock

We do a lot with computer use, speech, memory, vision, and a few other things. We’re building our own AI devices internally, which I’m really excited about. We launched Hark Pro, our digital version of Hark—our non-hardware version—about a month ago. We’re about a month in now, and it’s been a crazy launch.

We’re really interested in hardware. We think hardware is going to 10x this whole thing. I was even debating whether we wanted to put anything out before hardware. They said, “Okay, let’s just put some stuff out. We think the models we’re developing now are pretty good.” I actually use them all the time. It’s all I’ve been using for a while now. I was like, “Let’s put them out and see how it goes.” It was just crazy.

We’ve now had over 1.2 million signups. We’ve done probably almost five million browser sessions since launch. The usage is going through the roof—so much so that we had to curb some countries last week just because we’re—

Peter Diamandis

Yeah.

Brett Adcock

Anyway, that’s good. We’re developing models and hardware. Right now, we’re very focused on one area: real-time speech-to-speech. When you talk to the robots on the way out, they’re using the Hark voice real-time speech-to-speech model. We’re doing a lot of improvements there. I think real-time speech-to-speech is going to be a really important part of the—

Peter Diamandis

Agreed.

Brett Adcock

—the UI layer for robotics.

Peter Diamandis

There are no third-party components? It’s all internally developed real-time speech-to-speech?

Brett Adcock

Our speech-to-speech model is our own model here.

Peter Diamandis

There are a lot of people demoing speech-to-speech on top of something else.

Dave Blundin

It’s very similar to what we’re talking about with the Chinese robots. The difference is night and day when you have a vertically integrated system.

Brett Adcock

Can I tell you something about speech-to-speech? Speech-to-speech today kind of sucks. The hardware isn’t right for it. We don’t use it as a primary thing. It’s something we’ll use every once in a while in the car to talk to. It’s not really that humanlike. It doesn’t do well with tool calling, like going and getting information. It’s not very good at low latency. Sometimes you interrupt it and it doesn’t hear you.

Peter Diamandis

It’s just not really quite great.

Brett Adcock

At both Figure and Hark, we strongly believe speech is the right UI layer, the default layer into robotics and physical AI hardware.

So we’re all in on speech. We really want to make that work. We’re working on it at Hark, and we have a collaboration with Figure. Figure was the first beta user of Hark, so we’ve been using it for 5 months. We don’t force people to use it, but if they want to, they can use Hark internally. The numbers are basically through the roof internally, which has been great. We’re getting productivity benefits around the company.

We’re also working on larger-scale multimodal pretraining at Hark. We think it’s really important for the space.

Peter Diamandis

Obviously, Figure has been working on how to generate action predictions from the robot. We’re doing a lot of collaboration in some areas.

I have a running bet inside the Moonshots team here, especially with Alex. Elon very famously started xAI, then SpaceX bought it, and then X, and so forth, consolidating the companies. The bet is that Figure will eventually buy Hark.

Brett Adcock

I’d probably make a different guess. Don’t hold me to it, but I bet Hark is bigger than Figure in under 6 months.

Peter Diamandis

In under 6 months?

Brett Adcock

Maybe under 90 days.

Peter Diamandis

Are you serious? From a revenue standpoint?

Brett Adcock

It’s already bigger than Figure from a revenue standpoint. We’ve been growing like crazy, and the Hark numbers are through the roof. We have some very big things coming in hardware and some very big commercial agreements.

Given that it’s digital, we can scale on a very different slope from robotics. But listen, this is my guess. We don’t have anything going on right now. I do see a path where, next month, Hark could be bigger than Figure, and then at some point longer term maybe Figure gets bigger again. I don’t know.

Peter Diamandis

A trillion here, a trillion there.

Brett Adcock

The reason I wanted to work on Hark is that I was doing the human robotics thing, and we’re an AI business. We want to build the AI brain that fits in the hardware. The hardware is just the distribution for the AI, the vector it’s going to head out into the world in.

I was playing around with ChatGPT and the rest of the AI stuff, and they were so stupid. I was talking to a chatbot that was a search engine replacement for Google.

That isn’t what I really want. I want something that can be with me everywhere I go, knows me really well, has access to my accounts, and can do things for me: pay the property tax bill, remind me of something before I forget it, help book my travel, or do something more complicated over a long horizon with my permission, almost like Jarvis would. Nobody was doing that.

I opened my Series B at Figure, and I got the feeling that those folks had no real interest in making this a huge internal priority. I also feel that on the consumer side, interacting with computers and phones using AI is a really broken experience today.

I sat here and said, “Okay, I’m kind of in my prime as a builder. I’m trying to understand AI more. I’m fascinated by this digital version of AI.” Even going to a computer and using it is like a little mini humanoid moving a mouse and keyboard and trying to get everything to work. I’m personally fascinated by it, so I spent a lot of time researching it and understanding how deep learning works in this space.

I got to a point where I thought, “I can sit here and watch all this stuff, or I can go build and come off the sidelines.” I decided it would make me a better founder and CEO if I came in and helped build something I think was really missing for the world. I think it’s going to be really special.

You know those charts you see over a whole lifetime, showing how often you see your kids and parents? It goes up, then down and down until the rest of your life. I think there’ll be a chart like that for robotics, and certainly for things like Hark, where the line goes up all the way to the end.

You’ll have the smartest person in the world with you everywhere you go, and it’ll have the same interests as you.

Peter Diamandis

And help you.

Dave Blundin

Trustworthy, though. That’s the key. What Figure and Hark have in common is that you have to trust this entity with your most personal information for it to be effective.

Brett Adcock

We tell the team every day: “You have the person’s entire diary.” We need a new level of security and privacy that’s never been done before. At Hark, we try to protect that from every angle, and it makes things really hard for us. We can’t do what other companies do with advertising and other things for their products, because we’re trying to protect the consumer’s interests as much as possible.

Peter Diamandis

Congratulations on the meteoric growth. When did you announce Hark?

Brett Adcock

We’re over a year old now, and we launched Pro a month ago. We’re about a month out from Pro. It’s been great, man. I’m learning every day.

We’re in the very early chapters of AI, almost like looking back at the internet in 1999. It just wasn’t quite the right internet yet, but everything was getting seeded. It’s all getting seeded now. For a minute, over the last 6 months, I thought we were really late at Hark. Now I’ve realized we’re pretty early.

We’re launching 4 or 5 pretty big product announcements this week. We launched one this morning. And as we think about getting into hardware, we have hardware in the lab at Hark. Dude, is it crazy.

Peter Diamandis

So, peripherals?

Dave Blundin

When you say hardware, are there going to be Hark glasses?

Brett Adcock

Can I give you my thesis on devices? I think there are 2 different areas for AI devices. One is the centerpiece for AI devices. For today’s consumer electronics, that’s the phone and computer: about a billion devices a year. They sit right in the middle and are the main show, the main platforms. Around them, you have accessory devices. These could be AirPods, watches, glasses, or whatever else. I think AirPods are 3% of Apple’s revenue.

Apple isn’t sitting there getting super excited about AirPods, but it’s an important platform play. They’re super interested in the computer and phone. That’s the main thing.

What will happen is that AI devices will kill the devices in the middle. They’ll be a billion-unit-a-year devices that are way better than your phone and computer today.

Then, just like it is now, there’ll be accessory devices on the outside. Those devices will be single-digit percentages of revenue, and they’ll only work with the main show.

For AI devices, we’ve only seen the accessories today, like the Meta glasses, which are horrible, and other things on the periphery. They don’t work well today because they interface with things on your main platform, like your phone. Your iPhone will fundamentally never interface well with an AI device. If you try the Meta glasses, they need the Meta app open on your phone, or they can’t get a network connection.

Peter Diamandis

Even with that, it’s a horrible experience. The product might be one of the worst products I’ve ever tried.

Dave Blundin

How do you really feel about this?

Brett Adcock

Let’s get real. I watch the Spectacles and feel like I’m taking crazy pills. Nobody’s going to wear these. They keep getting bigger and uglier somehow. But for us, you have to go build it well.

Peter Diamandis

What does the ideal AI peripheral device look like?

Brett Adcock

We’re not going to tell you. We’re working on that.

Peter Diamandis

We wouldn’t tell anybody.

Brett Adcock

I think my best analogy—

Peter Diamandis

But you have a vision for what it becomes.

Brett Adcock

We have the hardware in the lab right now.

Peter Diamandis

Nice. Oh, wow. Really?

Brett Adcock

Yeah, we have it right now. It’s in the lab next door. I’m using it every day.

Peter Diamandis

Huh. Does it have windows? Can I look?

Brett Adcock

It’s unbelievable, guys. This stuff is going to be really incredible.

Peter Diamandis

Are you coordinating with Jony Ive and OpenAI? Do you guys talk to each other?

Brett Adcock

No, we’re not coordinating at all. We’re not talking at all. We fired them about 2 years ago.

Peter Diamandis

Okay. How so? Is there a backstory there?

Brett Adcock

They came in—I met Sam—and they led our Series B at Figure, co-leading it with Microsoft. They ended up joining the board, and we did a lot of work together internally for about a year. The goal was to build general intelligence into a robot. I was in every weekly meeting. We had a whole team dedicated to it, and they had a team dedicated to it. We said, “Great, let’s get to Figure every day and let’s rock.” They said, “Yep,” and then they never showed up at the office. It was too far from San Francisco. It was loud over there. We had grease on us, it was loud, and we were working on robots, but they just would never show up. We couldn’t get them there.

I thought, “Okay, let’s do a big event and get you guys here.” We just couldn’t get them into the office. Every week I watched the progress we were making and shared with them, and then watched what they were making and sharing with us. It was completely lopsided. We were making 10 to 1, 100 to 1 progress compared with what they were doing, and we were sharing all the stuff we were doing.

Peter Diamandis

Meanwhile, they were sharing with Microsoft.

Brett Adcock

Then it became pretty obvious that, after seeing our results, they really wanted to get into robotics again. It was just something I—

Peter Diamandis

When did Jony Ive come on board at OpenAI? Was that after this?

Brett Adcock

I don’t think Jony’s working on the humanoid at OpenAI. Nobody knows. My guess is he’s working on the AI devices. I don’t know if he’s working on—

Peter Diamandis

The iPhone killer. They’re afraid to say that now, but it’s some device with AI embedded in it that’s so compelling. Your point is that there’ll be a centerpiece. Your personal information isn’t going to be everywhere; you’re only going to trust a limited number of centerpieces. I think they want that device to be the centerpiece, not the iPhone. You obviously have the same vision with Hark, so I can see how those are—

Brett Adcock

Listen, it’s certainly not going to be the iPhone. The iPhone’s going to go away.

Peter Diamandis

All right, we’re calling it here: the decline of Apple.

Brett Adcock

Let’s just call it. If you take the extreme case, in 10 years we’re not using a traditional phone.

Peter Diamandis

Yeah.

Brett Adcock

We know that.

Peter Diamandis

Apple really has a challenge right now because they’ve been gliding forever on the iPhone, and it’s going to get challenged from a dozen different places.

Brett Adcock

Yeah.

Peter Diamandis

I’ve thought I’d wake up one day and see Tim Cook or John Ternus get on some emergency PR event and basically say, “Hey, we bought $100 billion of GPUs. We’re bringing the top team in the world. We’re going to go at pretraining hard. In the meantime, we’re going to use other models. This is going to be so core to the experience and our interface with consumers long term that we really need to own it, understand it, and embed it.” It’s been crazy to see nothing.

Brett Adcock

It’s crossed the point, probably about a year ago, where it’s really too hard to turn around.

Peter Diamandis

You think?

Brett Adcock

For sure. They’ve lost so much momentum.

Peter Diamandis

Yeah. You want to co-design the hardware and the models together. You really want to think about these AI devices as a pure interface. It could be a brain-computer interface in the brain. It could be language devices that we’re working through now. This is just the interface to raw intelligence on their side that’s going to be humanlike, just for you. They need to be co-designed together. It’s not something you just flop in.

Brett Adcock

I’m totally interested in BCI. I think it’s one of the most important vectors for AI interfacing long term. I’m invested in 4 of the companies right now, and I had some meetings yesterday. It’s blowing my mind. We’re on the precipice of being able to get to mouse control.

Peter Diamandis

Yeah.

Brett Adcock

And then human control.

Peter Diamandis

Yeah.

Brett Adcock

And then the world changes.

Peter Diamandis

Yeah. That’s sick. We should debrief at some point.

Brett Adcock

I’d love to make some introductions. There’s an amazing company.

Peter Diamandis

Yeah. I think we’re going through this phase beyond where we’re at now. It’s the language device phase. I think there’s a phase after that of BCI and alignment into the brain. I think it’s going to be really important. I expect Hark to be working through that whole process, all the way through.

Brett Adcock

I think there’s an upgrade cycle coming for traditional computers and phones. What I can tell you from using some of the products here internally is that it’s going to be phenomenal. It’s going to be a much better experience.

Peter Diamandis

Yeah.

Brett Adcock

We haven’t changed in 20 years.

Peter Diamandis

It’s literally been 20 years, and now we’re just plopping AI into it. The interfaces are really reactionary. You have to control the phone yourself. We’ve built this thing that can think now, use computers, speak to you, see, and build persistent, almost perfect memory. You want to be able to interface with it in real time and have it with you everywhere you go. We don’t have anything like that in the world today, but we will shortly. It’s awesome.

Dave Blundin

After AI, the interface is the phone. You basically use the phone as a camera, but the cameras are like $3.

Peter Diamandis

Why did I pay $1,500? I got a $3 camera.

Brett Adcock

Phones are incredible. There’s just a next leg up coming. It’s going to be great.

Peter Diamandis

Yeah, it’s about time. Let’s wrap with a conversation about regulation, because I’m curious what you’re thinking. We’ve got Bernie Sanders and others saying we’ve got to stop, slow down, prevent ASI. I have to imagine there’s going to be a similar kind of “slow down, we don’t need robots on our streets; it’s not safe for the citizenry.” How do you think about that? Are you having those conversations? Is there real concern the public should have in this regard?

Brett Adcock

I can speak for our plans. Our plans are to launch a really high-quality product that works really well in the market and is safe. We don’t want to be out there hopping the robot around and doing unsafe work or shipping too early. We’re making incremental steps now. We shipped commercially for the first time last year, and we’re shipping even more this year.

Peter Diamandis

Have you seen any indication from regulators that they’re starting to speak about humanoid robotics the same way they’re speaking about superintelligence?

Brett Adcock

We really haven’t seen that publicly. Even in rumblings, the only thing we’ve been aware of in the last 90 days is the FCC ban on Chinese humanoids. Beyond that, we haven’t seen anything come across our desk.

Brett Adcock

I think it’s also a little personal to me. I don’t fly to D.C. every week and try to get my nose in this. I really want to be here. I’m here.

Peter Diamandis

Good for you.

Brett Adcock

I’m here trying to build, and I want to build awesome robots regardless. There are going to be robots coming. They’re going to be great for the world. I think the same with AI. It’s going to be great. If humanity woke up one day and just didn’t want AI, we could just shut down the internet. We could just close it. But we don’t want that. It’s going to be really helpful for the whole entire world.

There will be, whatever, Bernie Sanders of the world, but for us, we’re just really focused on the mission of trying to build and ship robots.

Brett Adcock

And we’re here every day trying to go do that.

Peter Diamandis

I’ve got a question just to inspire all the entrepreneurs out there. I grew up a big Steve Jobs fan, too, and I think a lot about how much more he could have done in his life. But in that era, you could only do one thing at a time, and nobody had super-voting control and all that.

Then Google came along, and then Zuckerberg came along. It switched to this era where the founders control the companies, and they could be much more aggressive with the business plan. That works really well here in the Valley. But then Elon came along and reinvented it again, where you can do multiple things at the same time.

This is working out incredibly well for Figure and Hark. You just said Hark, nine months from today, could be bigger than Figure, which is mind-blowing.

Brett Adcock

I think you said three months from today.

Peter Diamandis

Three months from today.

Brett Adcock

I mean, Hark’s just going like this, guys. It is going really well. We’ll see.

Peter Diamandis

You’ve also got this other thing about school safety, which is not going to be bigger than the other two companies, but it’s important for the world.

Brett Adcock

Yeah. It’s happening concurrently.

Peter Diamandis

The power of an entrepreneur today is so much bigger than it was in the Steve Jobs era, let alone the Zuckerberg era. I think it’s really inspiring. Do you think a 20-year-old entrepreneur today can walk in your footsteps, or is AI going to make it all too late and the ship has sailed?

Brett Adcock

Yeah. One thing I’ll say is I’m completely dying. Three kids, three companies. I’m just happy.

Peter Diamandis

You look—

Brett Adcock

I’m just—I’ve been here every night for a whole year. Man, has it been hard. But it’s the best time of my life, and I never want to do it again. It’s been going great.

Listen, the future generation is going to be great. Everything’s going to be AI-native. You’ll be able to build any type of hardware just through—

Peter Diamandis

Prompt product.

Brett Adcock

Exactly. Just through talking. Robots—it’s going to be so cool. AI is really missing imagination today, and we’re going to be bound by this imagination. That’s what entrepreneurs do best, right? They always have a little notepad. I still have a notepad of all the business ideas I can think of, and I still keep them to this day. They’re great at idea generation and thinking about the future. This is going to be an unbounded opportunity for every new entrepreneur coming up behind me.

It may mark the best time in history for entrepreneurs. Now you have an ecosystem here for entrepreneurship. When I was graduating school, it wasn’t conceivable that you’d start as an entrepreneur. You just didn’t do that. Nobody did.

I had a job for a few years before I started. I wish I’d started even sooner. Now you can do that. You can get capital for it. There are incubators, and people want to work at startups. It’s fun. You have this whole community built around it, too, which is going to propel these future entrepreneurs further.

Peter Diamandis

Yeah. I like to joke that we’ve all been promised a three- and four-day work week. We’ve invented a 10-day work week.

Brett Adcock

I mean, I like working, man. Let’s just call it what it is.

Peter Diamandis

This is the most exciting time ever to be alive.

Brett Adcock

I like grinding out and working. I think everybody here with me likes it, too. We have a bunch of groups that like to work hard. We’ve made it this dirty word out here in the Valley, and a lot of folks come in here and say, “You guys are the most un-Silicon Valley company I’ve ever seen.” We work hard. We come in on the weekends. We work nights. We like it.

There’s not enough of that anymore. We interview people and ask, “What’s your day-off strategy?” and “Do you work from home?” We don’t have any of that. You come in here every day, work with us, and we’re going to work with you. My Baku guys can’t work from home.

Peter Diamandis

They’ve got to be on the line every day.

Brett Adcock

Yeah. Same with us. We’ve got to get on the line, too. Anyway—

Peter Diamandis

What you’re building is so much cooler than virtually any day job. It’s got to be so much more rewarding at the end of the day to look at what you physically built. I also didn’t anticipate how much the office environment is elevated by the robots walking around.

Brett Adcock

So much better.

Peter Diamandis

It’s bizarre. It just adds energy to the building somehow.

A year from now, we’ll come back and do our Figure 4 recap.

Brett Adcock

You come back, Figure 4 gets you all set up, checks you in, gets you water and coffee, walks you around, gives you a tour. Then I’ll see you for the interview.

Peter Diamandis

Yeah, we’ll include Figure 4 in the interview, too.

Brett Adcock

Honestly, Figure 4 can speak on my behalf.

Peter Diamandis

That’s so fun.

Brett Adcock

You know what I mean?

Peter Diamandis

Buddy, thank you so much.

Brett Adcock

Thanks for having me.

Peter Diamandis

Yeah. And when we cut cameras, we’ll go see your secret device in the Figure 4.

Brett Adcock

Perfect.

Peter Diamandis

Good deal.