DSTL - Audio in. Alpha out.

dstl

TERMINAL

LIBRARY

Brett Adcock on Why Humanoid Robots and AI Will Deliver an Age of Abundance

Join

Brett Adcock on Why Humanoid Robots and AI Will Deliver an Age of Abundance

Apr 1, 2026

Shawn Ryan Show

2:57:06

362K Views

THESIS

The founder of Figure AI argues humanoid robots running entirely on neural networks are now proven in real factory shifts — and the scaling bottleneck has shifted from physics to data.

ASSET CLASS

SECULAR

CONVICTION

HIGH

TIME HORIZON

3 to 5 years

PREMISE

The world is built for the human form factor, and coded robotics cannot navigate its complexity

Human civilization has constructed its entire physical environment — doors, stairs, tools, appliances, vehicles — around the human body's specific dimensions and capabilities. Any general-purpose machine that aims to perform useful labor in this environment must replicate the humanoid form. However, a humanoid robot with approximately 40 degrees of freedom produces more possible body states than there are atoms in the universe (360^40), making traditional hand-coded software completely intractable. Previous generations of humanoid robots were hydraulic, heavy, expensive, leaky, and could only run for about 20 minutes on a charge. There was no demonstrated precedent for neural networks successfully controlling a humanoid robot in real-world tasks. The structural imbalance is that the demand for labor automation — across manufacturing, logistics, healthcare, construction, and domestic settings — is enormous and growing, while the supply of capable humanoid systems was effectively zero until very recently.

MECHANISM

End-to-end neural networks replacing coded controllers unlock scalable, autonomous humanoid deployment

The critical forcing function was the successful demonstration in 2023 that a humanoid robot could execute dexterous tasks (like making coffee from a K-cup) using only neural networks — taking in raw camera pixels and outputting motor trajectories with no hand-coded logic. This proved the viability of an AI-first architecture for humanoid robotics. Figure AI then deployed second-generation robots at BMW's manufacturing facility, where they operated 10-hour shifts daily for six consecutive months on the body shop line, with the same physical robot completing both the first and last day. The key learning was that the coded portions of the software stack were the primary failure mode — they could not generalize to novel conditions. This led to a complete architectural refactoring into Helix 2, which removed over 100,000 lines of traditional code and replaced nearly the entire stack with neural networks. The result: robots now run 24/7 shifts autonomously for days without faults, self-coordinate charging swaps in under 10 seconds, and operate across multiple use cases including logistics, manufacturing, and office navigation. The mechanism converting this from a research project into an economic force is that once the hardware is standardized (like a phone that accepts new apps without hardware changes), scaling becomes a data problem — train neural networks on more tasks, deploy to more robots, without redesigning the machine.

OUTCOME

Billions of humanoid robots driving unprecedented productivity gains and a collapse in goods and services costs

Adcock projects that humanoid robots will eventually reach a scale of millions to billions of units, deployed first in commercial workforce settings (manufacturing, healthcare, construction) and ultimately in every home — comparable in ubiquity to phones and cars. Combined with digital AI agents that can operate computers autonomously, this creates what he describes as synthetic human intelligence at massive scale. The economic consequence is a radical increase in effective GDP per capita, as the ratio of productive labor units to humans explodes. Goods and service prices collapse to unprecedented levels, producing what Adcock calls a true age of abundance. In the near term, Figure AI is preparing to redeploy robots to customer sites running entirely on neural networks, with third-generation hardware (Figure 3) designed as the best humanoid hardware in the world. The commercial model begins in industrial settings and extends to consumer home robots within several years.

NECESSARY CONDITION

Regulatory frameworks must remain permissive to innovation (avoiding the 'European' model) and open source development must remain unencumbered by downstream liability.

I think you'll see the some of the most transformative events in technology happen over the next like 36 months we've ever seen in in our like um ever.

09:42

RISK

Steel Man Counter-Thesis

The strongest counter-thesis is that general-purpose humanoid robots will follow the trajectory of autonomous vehicles rather than smartphones — a decades-long slog through edge cases, regulatory ambiguity, and unit economics that never reach consumer scale within the founder's projected timeline. The empirical precedent is damning: autonomous vehicles, which operate in a far more constrained domain (2D roads with lane markings, traffic signals, and known physics), have consumed over $100 billion in cumulative industry investment across Waymo, Cruise, Uber ATG, Argo AI, and others, yet after 15+ years remain geofenced to a handful of cities with human teleoperators standing by. Humanoid robots face a strictly harder problem — 3D manipulation in unstructured environments with orders of magnitude more variability than driving. The neural network approach, while necessary given the state space, introduces the same generalization failures that have plagued self-driving: long-tail edge cases where the model encounters situations outside its training distribution and produces unsafe outputs. In driving, this means a fender bender; in a home with children, the consequences could be far worse. Furthermore, the business model requires achieving automotive-scale manufacturing for a product category that has never existed at scale, with no established supply chain, no regulatory framework for home safety certification, and no demonstrated willingness by consumers to pay what the unit economics likely demand. The SPAC-era comparison is instructive: Archer itself, despite being public for over three years with a $6 billion market cap, has yet to carry a single paying passenger. The pattern of brilliant engineering demonstrations that cannot cross the commercialization chasm is the most common failure mode in deep tech, and nothing in this conversation provides evidence that Figure has solved the manufacturing, regulatory, or unit economics problems that determine whether this becomes a real business or remains a perpetual R&D project funded by successive capital raises.

RISK 01

Hardware Reliability and Safety at Consumer Scale Remains Unproven

THESIS

Adcock's thesis depends on humanoid robots operating autonomously in unstructured home environments alongside children and vulnerable populations. Yet by his own admission, after years of development he still does not feel comfortable leaving robots unsupervised with his own children. The BMW deployment — his most successful commercial reference — involved a single repetitive task in a highly controlled industrial environment with human oversight, running for only six months. The leap from a controlled factory floor performing one sheet-metal placement task to millions of robots autonomously navigating the chaos of human homes (candles, boiling water, children climbing on them, pets, shag carpet, uneven terrain, novel objects) represents an exponential increase in failure modes. Each of those 40 degrees of freedom represents a mechanical wear point, and running 24/7 shifts at the office already produces software faults within days. Scaling to millions of units in uncontrolled environments with no human monitoring creates a product liability exposure that could be catastrophic — a single serious injury to a child could trigger regulatory shutdown of the entire category.

DEFENSE

Adcock acknowledges this directly — he states he is not yet comfortable leaving robots unsupervised with his children and that there is still significant wood to chop on safety. He describes both intrinsic hardware safety measures and semantic safety layers being developed. He also describes the transition from coded controllers to neural networks (Helix 2) as improving robustness in novel environments. However, the defense is aspirational rather than demonstrated: he projects several years before home autonomy is trustworthy, and no specific safety certification framework analogous to the FAA's 10^-9 standard exists for home robotics.

RISK 02

Neural Network Opacity Creates Unpredictable Failure Modes at Scale

THESIS

The core technical thesis is that hand-coded software cannot handle the combinatorial complexity of humanoid robotics (360^40 possible states), therefore neural networks must replace code entirely. Adcock celebrates removing over 100,000 lines of code and running almost the entire stack on neural nets. However, this creates a fundamental interpretability problem: when a neural network fails, diagnosing why it failed is extraordinarily difficult. In safety-critical applications — a robot holding a knife, operating near a stove, carrying a child's toy near stairs — the inability to deterministically predict or explain behavior is a first-order risk. The original question from the Patreon member directly addresses this: AI systems hallucinate and invent answers. In the physical realm, a hallucinated motor trajectory doesn't produce a wrong text answer — it produces a physical action that could cause injury or property damage. The more the system relies on end-to-end neural networks with no code guardrails, the harder it becomes to guarantee behavioral bounds. This is the opposite trade-off from aviation, where the FAA demands deterministic proof of safety. No equivalent regulatory framework exists for home robots, meaning Figure is essentially self-certifying safety in a domain with no external validation.

DEFENSE

Adcock discusses the safety question only in general terms — intrinsic hardware safety and semantic safety layers. He never addresses the specific epistemological problem of neural network opacity: how do you certify that an end-to-end neural network will never produce a dangerous output in a novel situation it has never encountered in training? He frames the move away from code as purely positive (removing brittleness) without acknowledging that code at least has deterministic, auditable behavior. The aviation analogy he draws actually undermines his robotics thesis — the FAA requires 10^-9 reliability with deterministic proof, while his home robot strategy is moving toward less interpretable systems with no external certification body. This asymmetry between his aviation safety philosophy and his robotics safety philosophy is never reconciled.

RISK 03

Unit Economics and Manufacturing Scale Are Assumed Rather Than Demonstrated

THESIS

The thesis requires humanoid robots to become as ubiquitous as phones and cars — billions of units. Adcock states they currently have hundreds of robots and will reach thousands this year, but need millions to make an impact. The path from hundreds to millions requires solving manufacturing at automotive scale for one of the most mechanically complex consumer products ever conceived — 40 actuators, custom gearboxes, advanced sensors, dexterous hands, battery systems, and onboard compute, all at price points consumers or businesses will accept. He draws the analogy to phones but phones have approximately zero moving parts and are manufactured on established semiconductor fabrication lines. Humanoid robots are closer in complexity to automobiles but with tighter tolerances and more novel components. No cost target is stated. No manufacturing partner or strategy is described beyond the company's own campus. The history of hardware startups shows that the gap between prototype and profitable mass production is where most ventures fail — Fisker, Nikola, and many EV SPACs from the same era Adcock describes provide cautionary precedent. The $2 billion raised sounds large but is modest relative to automotive-scale manufacturing capex (Tesla spent over $6 billion on Gigafactory Nevada alone).

DEFENSE

Adcock does not discuss unit economics, manufacturing strategy, bill of materials cost, or the path to profitability at any point in the conversation. He mentions the robot needs to be cheap enough but provides no specifics. He acknowledges the company has hundreds of robots today and targets thousands, but the leap to millions — which he says is required for impact — is treated as an inevitability rather than an engineering and capital challenge that must be solved. Given his own experience with Archer, where FAA certification timelines remain uncertain years into the process and the company required a billion-dollar public listing to survive, the capital intensity risk is something he has lived through but does not address for Figure.

ASYMMETRIC SKEW

The upside case requires simultaneous resolution of at least four independent hard problems: neural network reliability in unstructured environments, mass manufacturing of novel electromechanical hardware at consumer price points, regulatory acceptance with no existing framework, and consumer adoption of an autonomous physical agent in intimate spaces. Failure on any single dimension prevents the thesis from materializing. The downside is that Figure consumes its $2B+ in capital over 5-7 years, produces impressive demos and limited commercial deployments, but never achieves the unit economics or safety certification required for mass adoption — following the pattern of numerous deep-tech SPAC-era companies. The skew is unfavorable on a risk-adjusted basis: the probability-weighted downside (capital destruction with no path to profitability) is substantial, while the upside (trillion-dollar humanoid economy) requires a conjunction of low-probability breakthroughs. The asymmetry resembles a deep out-of-the-money call option — enormous potential payoff but with a high probability of expiring worthless.

ALPHA

NOISE

The Consensus

The market consensus holds that AI may be in an economic bubble, with speculative overinvestment in AI and robotics companies that have not yet demonstrated scalable, profitable commercial deployments. Poly Market prices an 18% chance the AI bubble bursts by end of 2026. The broader consensus is that general-purpose humanoid robots are a distant-future technology (10-20+ year horizon), that the path from lab demos to commercially viable products at scale is extremely long and capital-intensive, and that top-tier venture capital has largely avoided deep-tech hardware plays like humanoid robotics due to unfavorable risk-return profiles. The market also holds that autonomous systems — whether vehicles, aircraft, or robots — face enormous regulatory, safety, and consumer trust barriers that will slow adoption significantly.

The market's logic rests on several pillars: (1) Prior humanoid robotics attempts (Boston Dynamics, etc.) failed to achieve commercial viability despite decades of investment; (2) Hardware-heavy deep tech ventures have historically poor venture returns and long development cycles; (3) Neural networks hallucinate and make errors, which in physical systems creates exponentially worse downside risk; (4) Regulatory bodies like the FAA move slowly and certification timelines are inherently unpredictable; (5) The installed base problem — replacing billions of existing cars, infrastructure, and labor systems — creates multi-decade adoption curves that don't match venture capital return timelines; (6) Top VCs have systematically avoided humanoid robotics investments, signaling informed skepticism about near-term viability.

SIGNAL

The Variant

Adcock believes we are emphatically not in an AI bubble and that the most transformative technology events in human history will occur within the next 36 months. His variant view is that the humanoid robotics timeline has been compressed by roughly a decade — that electric humanoid robots running neural networks are viable now, not in 2035-2040. He believes every human will eventually own a humanoid robot, comparable to owning a phone or car, and that synthetic human intelligence (both digital and physical) will scale to billions or trillions of units, driving an unprecedented productivity explosion that collapses goods and services prices into a 'true age of abundance.' He sees the current moment not as peak hype but as the absolute starting line, with the hard technical proofs already in hand and only scaling and regulatory timelines remaining.

Adcock's causal logic rests on a specific technical inflection: the convergence of cheap electric actuators, neural network control systems, and modern AI training paradigms has created a qualitatively new design space that did not exist even four years ago. His core causal chain is: (1) Electric motors maintain ~90% efficiency at any scale, unlike combustion or hydraulic systems, enabling small, redundant, cheap actuator arrays; (2) The state space of a 40-DOF humanoid (360^40 positions) makes coded/heuristic approaches mathematically intractable, but neural networks bypass this entirely by learning from data rather than exhaustive programming; (3) He demonstrated neural-net-only control on humanoid hardware in 2023 — what he calls the most significant proof point in the company's history — proving the approach works; (4) BMW deployment validated 6-month continuous operation with hardware performing at A+ levels, with failures concentrated in the coded (non-neural-net) portions of the stack, which were subsequently replaced with Helix 2's all-neural-net architecture; (5) The remaining bottleneck is data collection and training scale, not fundamental feasibility — converting the problem from an engineering moonshot into a data scaling play analogous to LLM development. He further argues the human form factor is not arbitrary but necessary because the entire built environment is designed around human morphology, making humanoids the only general-purpose robot architecture that avoids the need to redesign every environment it operates in.

SOURCE OF THE EDGE

Adcock's claimed edge is genuine operating experience across multiple deep-tech hardware ventures — he has actually built and deployed these systems, not merely theorized about them. His specific informational advantage comes from: (1) Direct operational data from BMW's manufacturing line over six months, which revealed that hardware reliability was not the bottleneck but coded software was — an insight only available to someone running robots in production; (2) First-hand experience with the neural-net-on-humanoid breakthrough in 2023, giving him visibility into the actual capability curve before the market priced it in; (3) Cross-domain pattern recognition from Archer Aviation (electric propulsion, control systems, certification processes, sensors) that directly transfers to humanoid robotics; (4) Having robots in his own home for extended testing with his family, providing qualitative safety and interaction data unavailable to pure-lab operations. However, the edge has meaningful limits. His timeline claims — 24 months for comprehensive AI life management, every human owning a humanoid — are aspirational projections that extend well beyond what his operational data actually supports. The BMW deployment showed B/B+ software performance after years of work, and he himself acknowledges the robot is not yet safe enough to leave unsupervised with his children. The gap between 'neural nets work on humanoids' (proven) and 'billions of humanoids replacing human labor across all domains' (projected) is enormous, and his edge is strongest on the former and weakest on the latter. His credibility is high on technical feasibility and hardware timelines, moderate on commercial scaling timelines, and weakest on his most expansive societal transformation claims, where he is functioning more as a visionary evangelist than an operator reporting verified results.

CONVICTION DETECTED

• Absolutely not — in response to whether AI is in a bubble • Ever — repeated twice when describing the transformative events coming in the next 36 months • For sure work and it'll go really far in our lifetimes • The greatest increase in productivity we've ever seen in our lifetime • A true age of abundance • Every human to have a humanoid like almost like a phone and car • This is going to really work • Game on • It's the best humanoid hardware in the world by far • The most significant demonstration we've done in four years • It'll happen. It'll happen in our lifetime • In like 5 years everything will be fully autonomous and trusted and fine • We'll do that in like 24 months • Very confident in this even though it sounds ridiculous • There's no code left really on the robot

HEDGE DETECTED

• We're still not at that stage yet where I feel comfortable enough to let loose — regarding unsupervised robot-child interaction • We're not there yet. I think we will be in the next several years • It's just man it's just an incredibly complex problem • There's still a lot of wood chop of getting this thing to a point • It's early. It's earlier than AVs or EVs • We still see failures right now • It's not like these things have been around for decades and we understand that they're really mature. They're not. • It's not something you can like there's not a date on the calendar like you'll be certified here • We're just in this transitory state right now • I think from a software perspective we did like a B job, B+ • This is a very very tough problem • We need like millions of robots to make an impact. That's just going to take some time The ratio reveals a speaker with genuinely high conviction who hedges selectively and specifically — almost exclusively on timelines and current-state readiness rather than on directional trajectory. He never hedges on whether this will work, only on when and how polished it is today. This pattern is consistent with an operator who has real knowledge of the gap between current capability and full deployment, rather than someone performing certainty for fundraising purposes. The hedges are credibility-enhancing, not credibility-undermining — they suggest he distinguishes between what he has proven and what remains to be proven. However, the most expansive claims (24 months for full life automation, every human owning a humanoid, trillions of synthetic humans) carry zero hedging whatsoever, which suggests these function as vision-casting narrative rather than engineering projections. Weight the near-term technical claims heavily; discount the civilizational transformation timeline claims by a significant margin.