4 wild AI predictions from a $39B tech founder

Brett Adcock's Entrepreneurial Background

• Brett Adcock was raised in a rural area of Illinois and founded the company Vettery, which was sold for over $100 million. 10s
• Archer, a company focused on unmanned flying vehicles, was taken public by Brett Adcock. 10s
• Figure is a company currently valued in the range of $30 billion to $50 billion. 10s
• Brett Adcock is involved with Cover, an initiative aimed at preventing school shootings, and a new venture called HARK that has raised billions of dollars in funding. 10s

Financial Risks and Personal Commitment

• During the early stages of Figure, Brett Adcock invested nearly all of his personal capital—amounting to tens of millions of dollars—into the company, leaving him with a mortgage on his home and remaining funds tied up in Archer. 42s
• Brett Adcock expresses indifference toward reports regarding his personal net worth, including a $19 billion valuation cited by Fortune. 42s
• The current trajectory of Brett Adcock's companies is described as binary, meaning they will either scale successfully or fail, with no middle ground. 1m25s
• Brett Adcock focuses his energy on ensuring his companies reach their potential, noting that if successful, they could grow by 100,000 times their current size. 1m25s

The Future of AI and Digital Symbiosis

• The next five years are critical for Figure, as the company must transition from its current state to achieving large-scale production of robots. 1m55s
• Brett Adcock predicts that the impact of current AI developments will be 100 times larger than the impact of the internet. 1m55s
• Deep learning systems are currently advancing at a pace that exceeds historical trends observed in the software and internet industries over the past 15 years 0s.
• The AI lab Hark was established approximately one year ago with the goal of creating a digital symbiosis between humans and AI 25s.
• The vision for this symbiosis involves individuals having a personalized AI pairing that may include unique AI weights, personal memories, and dedicated hardware 35s.

Hark's Computer-Using AI Agents

• A fundamental requirement for AGI is the ability for AI to use computers for general-purpose tasks, such as managing financial models, booking flights, or ordering food 50s.
• Because only one in a thousand websites provides an API, AI must be capable of navigating the internet and using browsers like a human, including moving the mouse and using the keyboard 1m5s.
• Within two to three years, it is expected that users will be able to instruct an AI system to perform complex tasks online without relying on APIs or Model Context Protocol (MCP) 1m15s.
• Hark recently released a research preview of its first model, which is reportedly capable of performing nearly any task requested on the internet 1m35s.
• To enable computer use, Hark spins up a virtual computer for every agent, allowing for the creation of multiple sandboxed environments 2m15s.
• The approach taken by Hark differs from others by focusing on a specific reinforcement learning process during post-training that is believed to be unique 2m55s.

Figure's Humanoid Robotics Development

• Figure is developing humanoid robots designed to operate in the physical world by performing tasks such as laundry, dishes, cooking, supply chain management, and healthcare duties 0s.
• The concept behind Figure involves creating a single piece of hardware capable of performing various physical tasks, which is then powered by smart artificial intelligence 2m15s.

Historical Data and Future AI Assistants

• A database created at age 24 tracked the career trajectories of approximately 50 successful individuals, documenting their birth years, apprenticeship start dates, initial successes, and breakthrough years 1m15s.
• This historical data on successful figures was aggregated to provide inspiration and insight into what is possible for those seeking success 1m5s.
• A second, distinct path for artificial intelligence involves a close digital symbiosis between humans and AI, functioning as a constant, superhuman assistant 2m35s.
• This digital assistant is envisioned to function similarly to Jarvis from Iron Man, possessing access to a user's memories, accounts, and systems to perform tasks and provide support in real-time, such as rebooking flights during travel disruptions 2m45s.
• Current AI technology is limited to coding agents and chatbots, lacking the capability to act as a comprehensive, omnipresent personal assistant 3m15s.
• Advancing toward a Jarvis-like AI requires improvements in model capabilities, specifically the ability to interact with computers, maintain near-perfect memory, engage in natural human-like conversation, and utilize vision to understand the surrounding world 3m25s.

Next-Generation AI Hardware

• Current hardware systems, such as MacBooks and iPhones, are considered inadequate interfaces for AI because they were designed decades ago. 0s
• A transition is underway toward an upgrade cycle where existing computers and phones will be replaced by AI-native devices and systems. 15s
• Future AI systems are expected to operate in real-time, maintain constant awareness of their surroundings, and possess the ability to reference ongoing events. 25s
• The shift toward AI-centric computing may lead to the obsolescence of traditional app stores in favor of AI operating systems that abstract away the need for individual apps. 35s
• Users will likely possess their own personalized AI weights on portable devices that they own and carry at all times. 42s
• A team of approximately 80 to 90 people, including a lead hardware designer who previously worked on multiple generations of iPhone and MacBook products, is currently developing next-generation AI hardware. 50s

Advancements in AI Models and Hardware

• AI models are being developed to become more multimodal and expressive, as current text encoding is insufficient for achieving a true AGI experience. 1m12s
• A research preview of a computer-using agent was recently released, which has performed at the top level on leading browser computer-use benchmarks. 1m22s
• The computer-using agent is expected to improve in speed and intelligence on a monthly basis. 1m32s
• The ability to use HARK on traditional browsers, iPhones, and Android devices is scheduled to launch in approximately one month. 1m42s
• Hardware prototypes are currently being tested in a laboratory setting, with descriptions of the technology being compared to science fiction. 1m52s
• The objective for HARK is to create devices that can reach a scale of one billion units per year, a threshold currently only met by phones and computers. 2m25s
• Ancillary devices like watches and earbuds are viewed as ecosystem supports rather than primary platforms, as they do not reach the same volume of sales as phones or computers. 2m35s
• Replacing the current computer and phone market requires building new, superior systems that can replace existing hardware end-to-end. 2m45s

Hardware Prototyping and Wearable Devices

• Hardware development at HARK includes a family of devices that are not intended for mass-market volume but are considered important for the ecosystem 0s.
• The development of future AI hardware involves radically rethinking device form factors rather than simply creating an AI-native version of a smartphone 15s.
• Early-stage hardware currently in laboratory testing is described as being unlike anything previously seen 25s.
• Wearable devices such as glasses and pendants are considered peripheral to the main technological evolution 32s.
• Meta glasses are criticized for poor functionality, specifically noting issues with network dependency, the requirement to keep apps open on a phone, and difficult pairing processes 38s.
• There is skepticism regarding the long-term utility of wearing glasses for extended periods, and they may not rank among the top 10 most significant future devices 58s.
• The long-term trajectory for AI interaction is expected to move toward Brain-Computer Interfaces (BCI), with AI language devices serving as an interim step for the next decade 1m5s.

AI Capabilities and Design Philosophy

• AI is categorized as a new type of computer and automation that must offer a 10x improvement over existing phones or computers to be considered viable for widespread adoption 1m25s.
• Key capabilities of modern deep learning include the ability to think, use computer systems, engage in real-time speech-to-speech communication, and possess visual understanding 1m45s.
• AI systems are expected to achieve human-level performance in using computers and systems, eventually surpassing human speed and success rates 1m55s.
• AI systems can be equipped with near-perfect, long-term memory, effectively functioning as a personalized, intelligent assistant that accompanies a user 2m5s.
• The design philosophy for new hardware involves removing all constraints to imagine a "magical" experience, such as having an all-knowing AI companion that can see, hear, and provide advice 2m35s.

Proactive AI Assistants and Prototyping

• Artificial intelligence systems like Harkc offer competitive advantages by providing human-like capabilities and near-perfect memory that allows for referencing past information over time. 0s
• Current smartphone systems are described as inefficient because they require manual input for tasks, such as saving contact information or navigating apps like DoorDash, which can lead to user error or forgotten details. 0s
• Harkc functions by abstracting tasks, allowing users to complete actions like ordering coffee through voice commands without needing to manually interact with an interface. 15s
• Advanced AI agents are envisioned to act as proactive assistants that can anticipate user needs, such as suggesting a coffee order or rerouting a delivery based on the user's location or preferences. 28s
• The transition from current smartphone technology to next-generation AI is compared to moving from using a tool like a hammer, which requires manual operation, to having a handyman who performs the task upon request. 42s
• Rapid prototyping and a dedicated fabrication facility are used to design and test various devices, with the development process involving wearing or carrying prototypes to determine which designs are effective. 1m0s

Robotics Challenges and Manufacturing

• A major telecom executive, who previously worked with Steve Jobs on the original iPhone, reportedly expressed that the current AI developments have the potential to compete with major established technology companies. 1m35s
• The primary challenge in the field is not mass production, but rather the creation of a truly intelligent robot system. 1m52s
• Existing commercial robots are characterized as ineffective toys that often lack functional hands and require manual control, similar to the early experience of using a DJI drone that was difficult to set up and lacked practical utility. 1m58s
• Current robotics technology faces significant challenges, as evidenced by early attempts that failed to function effectively in real-world environments 0s.
• The primary objective for Figure is to achieve human-level intelligence in robots, enabling them to be deployed in any home to perform a wide variety of tasks 12s.
• Manufacturing consumer electronics is distinct from automotive manufacturing, as the former is highly scalable and capable of producing billions of units, whereas the latter is significantly more complex and difficult to execute 32s.
• Automotive manufacturing is uniquely challenging because parts are too large to be held by hand, requiring a complex, interconnected system of hundreds or thousands of robots to assemble a chassis; if any single robot in this chain fails, the entire production line stops 55s.
• Figure positions its manufacturing process closer to consumer electronics than automotive production, estimating its current capability at a 40% level of maturity relative to phone manufacturing 1m22s.
• The company recently produced its 1,000th EVT robot for the Figure 3 model 1m32s.

Robot Deployment and Logistics Applications

• Robots are currently distributed between two primary groups: an internal engineering and AI research fleet used for testing and development, and external customers 1m45s.
• Figure has begun shipping robots to a third external customer as of the current week 2m6s.
• Current robot applications for customers primarily involve logistics and package handling, such as sorting items off of conveyor belts 2m13s.
• The company is actively exploring the deployment of its robots into various other industries beyond logistics and manufacturing 2m25s.
• Businesses are currently seeking automation solutions due to significant labor challenges, including high turnover rates that can exceed 100% annually and a general shortfall of available talent 15s.
• Rising wages and the difficulty of finding and retaining staff are driving companies to seek external help to automate tasks that were previously difficult to address 15s.
• Robots provide a strong return on investment for customers because they can operate for multiple shifts per day, work seven days a week, and maintain high levels of uptime 15s.
• A logistics use case demonstrated that a robot could process packages at a rate of 2.9 seconds per package over a continuous 200-hour period, matching human speeds 15s.
• While current deployments are in the early stages of scaling, the long-term expectation is that these automation technologies will generate billions of dollars in value for client organizations 15s.

Identifying Progress in the Robotics Market

• Distinguishing between factual progress and marketing fiction in the robotics industry is difficult for the public because testing these products is expensive and inaccessible compared to other consumer goods 42s.
• The robotics market currently contains a high volume of noise, making it challenging to identify genuine signals of progress 1m15s.
• The most critical objective in the field is the ability to ship robots that can perform autonomous work at scale in practical environments, such as healthcare, construction, logistics, and household tasks like cooking or cleaning 1m15s.
• Achieving autonomous work requires onboard AI rather than traditional hard-coded instructions, as robots must be able to operate independently over long durations, navigate physical spaces, and manipulate objects 1m15s.
• The primary objective for robotics development is not to create machines capable of performing stunts like backflips, dancing, or running, but rather to develop functional robots that can perform practical tasks, such as working in a factory setting 0s.

Talent Acquisition and Hiring Standards

• Recruitment efforts focus on identifying individuals who are genuinely obsessed with the work, as the majority of time is spent searching for these specific types of people 1m5s.
• Despite the high concentration of engineering and AI talent in the Bay Area, it is estimated that 90% of professionals in the field are not highly proficient at their jobs 1m25s.
• Technical assessment is used to evaluate candidates, relying on the principle that individuals who have personally performed the work possess a deep, intuitive understanding that allows them to discuss details freely and reverse-engineer their projects 1m45s.
• Candidates who have not personally performed the work typically fail to provide depth, often struggling to explain details beyond the surface level 2m6s.
• The hiring process for mechanical engineers at Figure is rigorous, requiring candidates to demonstrate the ability to build complex, compact actuators from scratch, including components like bearings, motors, sensors, and gearboxes 2m35s.
• The hiring process is so selective that, despite conducting 10 case studies per week for six months, no mechanical engineers were hired during that period 2m55s.

Competition for Top AI Talent

• Competition for top AI talent is intense, with larger, more liquid companies offering compensation packages that can reach tens of millions of dollars per year 3m10s.
• Aggressive hiring practices, particularly those driven by Meta, have persisted for over a year without showing signs of slowing down 3m25s.
• A senior AI infrastructure candidate was interviewed following the disbandment of the team at XDI, which occurred approximately six months prior 3m45s.
• A candidate was offered a stock package at HARK valued between $15 million and $20 million over a five-year vesting period, with the potential for the value to grow significantly if the company achieved its growth targets. 0s
• The same candidate received an offer from Meta consisting of $36 million in restricted stock units (RSUs) over four years, which was viewed as a guaranteed cash alternative compared to the speculative upside at HARK. 25s
• Meta has been aggressively recruiting talent for approximately one year, effectively buying its way into the AI race by offering high compensation packages. 42s
• There is a limited pool of experts who possess the specialized knowledge required for language pre-training, mid-training, post-training, and the associated supercomputing infrastructure; estimates suggest only 20 to 30 people in California are capable of building high-quality AI models. 1m15s
• Junior-level AI talent, including individuals in their 20s, are currently earning between $750,000 and $2 million annually, consisting of a base salary of $200,000 to $250,000 supplemented by RSUs. 1m45s
• Other AI labs have followed Meta’s lead by increasing their compensation structures to remain competitive in the talent market. 2m6s

Strategic Talent Acquisition and Team Size

• The strategy of buying talent is considered smart from a competitive standpoint, as it allows a company to secure a position in the AI race quickly. 2m15s
• A potential drawback of this strategy is the recruitment of "mercenaries" who are motivated primarily by money rather than a deep commitment to the company's mission. 2m35s
• Building effective AI models does not require a large workforce of hundreds of people, but rather a highly skilled team of 20 to 40 individuals who are deeply invested in the work. 2m50s
• Meta is recognized for its strong recruiting efforts and organizational structure, though there is uncertainty regarding the company's ability to ship successful products. 0s
• Meta is characterized as one of the most successful acquirers in history due to its purchases of Instagram and WhatsApp, but its recent work on products like Ray-Ban smart glasses is viewed as subpar. 15s

Robotics Training and Real-World Challenges

• The current AI landscape is described as being saturated with low-quality startups, with the exception of advancements in the coding sector. 1m15s
• A prediction made on January 1st regarding humanoid robots performing unsupervised, multi-day tasks in unfamiliar homes using neural networks is considered to be on track. 1m35s
• The process of moving from pixels to torques—translating camera feeds into motor commands—is already functional for current robotics technology. 1m55s
• The primary challenge for humanoid robots is operating in environments they have never encountered before, such as a new home with different lighting, furniture heights, and objects. 2m25s
• When a robot encounters a scenario outside of its training distribution, it lacks the necessary data to function, similar to how an LLM would struggle to discuss a topic if all relevant data were removed from its pre-training. 2m35s
• Because data for household tasks is not readily available on the internet, it must be actively collected from the real world to train models effectively. 2m50s
• While folding laundry is a common benchmark for robotics, commercial applications like warehouse work offer a more immediate path to scale and financial runway. 3m5s

Business Strategy and General Robotics

• The initial business strategy involved focusing on the commercial market to fund long-term development for home robotics, as commercial environments offer higher revenue potential, lower variability, and the ability to operate 24/7 0s.
• The current objective is to develop a robot capable of performing work in any environment using language, a goal that could potentially lead to a trillion-dollar market valuation with a team of 50 people and 100 robots 35s.
• The development of "general robotics" is currently underway, with the goal of being the first to achieve this capability 1m5s.
• Humans are currently capable of winning in a physical confrontation against existing robots 1m35s.

Voice-Based AI and Future Communication

• Daily AI usage is expected to shift from text-based interactions to highly multimodal experiences, with voice agents featuring persistent memory becoming common 1m50s.
• A version of a voice agent with persistent memory is currently in development and is scheduled for release within a month 2m6s.
• The rise of voice-based AI interaction is creating a need for changes in office design, as open-air office layouts are becoming less suitable for private, whispered conversations with AI systems 2m35s.
• Current speech-based AI technology is limited by session-based constraints, a lack of persistent memory, and difficulties with tool calling and computer use 3m15s.
• By 2027, AI systems are expected to pass a full human Turing test via speech, making it impossible to distinguish between a human and an AI during a phone call 3m15s.

Weapon Detection and Silicon Chip Development

• A full-scale scanning system designed to detect weapons from a 20-foot standoff is currently being built, with testing expected to begin before the end of the year 4m0s.
• A weapon detection project is currently underway, led by a chief engineer with a background at NASA and JPL 0s.
• The project is primarily an engineering challenge that combines hardware and AI development, with minimal business-related tasks beyond supply chain management 0s.
• A major technological pivot occurred approximately one year ago to transition to a system utilizing custom-designed silicon chips 0s.
• This pivot was intended to increase scalability, improve performance, and reduce costs by approximately 90% 0s.
• The custom chips required a specialized fabricator in Europe, resulting in a one-year design and production lead time 0s.
• Testing of the newly received chips has yielded positive results, though there is an additional six-month lead time required to manufacture more units 0s.

Personal Productivity and Health Management

• To manage professional and personal responsibilities, life is categorized into three distinct "pockets": work, family, and social obligations 1m15s.
• A decision was made to prioritize work and family at an "A+" level, which necessitated the complete elimination of the third social bucket, including activities like weddings, golf trips, and casual meetings with friends 1m15s.
• Every minute spent outside of the work or family categories is viewed as time taken away from those two primary commitments 1m15s.
• Daily workflow management involves handling a high volume of text messages and Slack notifications, which are not processed in real-time throughout the day 2m6s.
• The AI tool Hark is utilized for end-to-end management of professional tasks, including engineering projects, recruiting, email, Slack communications, and to-do lists 10s.
• Prior to using Hark, to-do lists were maintained in a Google document that was manually updated on a weekly basis 35s.
• Health management involves access to specialized medical services that provide quarterly blood tests, full-body MRIs, and heart CT scans 50s.
• These medical screenings are described as effective tools for early cancer detection, the prevention of heart attacks, and the identification of health anomalies 1m5s.
• While dietary habits have improved, time constraints currently limit the ability to maintain a consistent exercise routine 1m25s.
• A significant focus is placed on business development, which sometimes necessitates trade-offs regarding personal health and lifestyle habits 1m40s.

Startup Resilience and Financial Lessons

• Between 2012 and 2017, the experience of building startups involved recurring financial instability, including periods of personal debt and not receiving a salary 2m6s.
• During the development of the company Veter, a $500,000 convertible note was raised in 2015, followed by a successful marketplace launch six months later and a $110 million sale one year after that 2m15s.
• When facing extreme professional challenges or "rock bottom" scenarios, the recommended strategy is to focus on a daily "punch list" and prioritize surviving one day at a time 2m45s.
• A strategy for managing long-distance running involves breaking the total distance into small, manageable segments, such as half a mile or 200 yards, and telling oneself that quitting can be reconsidered only after reaching each specific milestone 0s.
• After selling the company Very and achieving a 12x return for investors, the transition to the company Archer was met with significant skepticism from potential funders who questioned the viability of the project 42s.
• Maintaining conviction in the face of widespread doubt requires having performed the necessary work and possessing a deep, ground-level understanding of the project 1m15s.
• The startup environment is characterized by high failure rates, with historical data from an NYU incubator showing that out of 50 companies, only two generated any revenue after five years, while 48 failed 1m35s.
• Funding the company Figure required extreme personal financial commitment, including investing all remaining capital from Archer, utilizing stock during a period of significant devaluation, and taking out a second mortgage on a house 2m15s.

The Value of Difficult Projects

• Pursuing difficult projects is often more advantageous than pursuing easy ones because hard tasks typically face less competition and offer the potential for a larger Total Addressable Market (TAM) and a more significant exit 2m45s.
• Working on challenging problems attracts high-performing individuals who are motivated by the unique risk-reward trade-offs associated with difficult endeavors 3m5s.
• Investments in difficult projects often follow a binary payoff structure, which makes them attractive to those seeking high returns for a portfolio. 0s
• There is a nonlinear relationship between the difficulty of a project and its potential payoff, where tasks that are only three to five times harder can yield returns that are 100 times greater. 0s
• Building humanoid robots is estimated to be three to four times more difficult than building quadruped robots, yet the humanoid form offers a significantly higher return on investment, potentially reaching a million or billion times higher. 15s
• Quadruped robots are viewed as having limited market potential and lacking a clear, scalable economic value compared to humanoids. 35s

Long-Term Commitment and Startup Outcomes

• Current AI tools, specifically those referred to as "Open Claw," are characterized as low-quality "slop" that are unlikely to succeed in the long term, despite potential for minor consolidation or acqui-hires. 55s
• Founders should carefully evaluate their long-term commitments, as the choice of project will dictate the next 10 to 15 years of their career and influence the probability of various outcomes. 1m25s
• While some business models operate on a binary outcome of becoming a trillion-dollar company or going bankrupt, many other businesses consider a $10 million annual revenue to be a significant success. 1m45s
• There is a question of whether individuals who spend decades building smaller, successful businesses would regret not attempting a larger, more ambitious venture when looking back later in life. 2m5s
• Personal experience includes spending seven years at a company called Veterary, where an internal marketing automation tool was developed that later mirrored the functionality of Outreach.io, a billion-dollar company. 2m25s
• Internal development of technology can sometimes overlap with external innovations, making the strategic decision of where to allocate time and resources critical for startups 0s.
• The startup landscape is described as binary, with a high failure rate even for companies that reach $10 million in revenue, making the achievement of such milestones difficult, particularly for those with limited capital 15s.

Future Energy and AI Impact

• Humanity faces a significant challenge regarding energy generation, which is closely correlated with human standards of living 1m15s.
• Potential solutions for future energy needs include solar, wind, and nuclear power, with specific interest in the development of fission and fusion technologies 1m35s.
• Artificial intelligence is expected to be a dominant force over the next 10 to 20 years, with a projected impact 100 times greater than that of the internet 1m55s.
• Artificial intelligence is anticipated to integrate with and encompass the entire internet, influencing both physical and digital domains 2m10s.
• The current landscape for artificial intelligence products remains in a "foggy" state, with no clear breakout winners despite some early successes 2m35s.
• Existing artificial intelligence devices currently on the market have not yet provided a compelling or "crazy great" user experience 2m55s.
• Whisper Flow is noted as a standout product from the last six months that has meaningfully changed communication habits 3m10s.
• There is a high level of admiration for individuals who demonstrate full dedication to their specific craft, similar to the work ethic depicted in the Michael Jordan documentary 3m30s.

Mentorship and Scaling Breakthroughs

• Steve Jobs is characterized as an exceptional operator and a product-led founder 0s.
• Jeff Bezos has served as a sounding board for the company Figure and is noted for his involvement as an investor 0s.
• Jensen Huang is described as a highly hands-on operator who has managed Nvidia in a unique way for the past 30 years 0s.
• Jeff Bezos advised that the company is at a critical juncture where it must determine how to scale and achieve a major breakthrough within the next one to two years 35s.
• The current phase of development is described as a pivotal moment where the company has moved onto a larger field and must execute a successful push to ensure its future 35s.
• The company has successfully transitioned from early development stages, such as building a single robotic knee, to creating humanoid robots that operate autonomously using AI 1m5s.
• Current robotic technology utilizes onboard camera inference to control joint movements 1m25s.
• The primary objective for the next phase of development is to achieve higher operational scale for the existing robotic technology 1m25s.