Hard Fork Live, Part 2: Patrick Collison of Stripe + Kathryn Zealand of Skip + Listener Questions
Summary
Patrick Collison’s abundance call is blunt: the most urgent need is “building,” because physical construction faces the most unnecessary, self-imposed constraints. California once celebrated the Golden Gate Bridge by building Treasure Island; now its high-speed-rail project has removed a former 2033 completion estimate. The inversion is stark: creating superintelligence sounds plausible, while building a new California city sounds “just ridiculous.”
Collison’s institutional bet is that progress comes from freeing capable insiders, not parachuting in outsiders with “delusions of grandeur.” In a survey of top-tier scientists, 79% said they would change their research agenda a lot if existing funds came without today’s restrictions. He also said 48% would do something “very different,” a figure the transcript does not reconcile with the 79%. Arc Institute applies that premise to complex disease, combining single-cell sequencing, neural networks and CRISPR in pursuit of computational “virtual cells,” though Collison cautions that “time will tell” whether it works.
Stripe expects agents to offload checkout labor while stablecoins may extend the dollar’s reach rather than overthrow it. Collison imagines “critters” acting as “paperwork minions,” handling antiquated forms and transactions while users return to whatever else they were doing. Agents will still use locally denominated amounts, and because most cryptocurrency transactions are already in dollar-denominated stablecoins, crypto’s consequential outcome “may end up enabling” greater dollar adoption abroad.
Stripe’s operating model requires employees willing to underwrite decade-long infrastructure curves. After 15 years, Collison says the company still feels as though it is starting; even successful products can take a decade to play out because Stripe builds “roads,” not fast-growing consumer “cars.” On Meta’s reported $100 million recruiting offers, he declined to comment after only one board meeting, leaving Casey Newton to joke that the figure suggests, “You would have to pay me $100 million to work at Meta.”
Skip’s MO/GO exoskeleton is a $5,000 early-market mobility product, not yet a mass-market bionic upgrade. Its lightweight motors can provide up to roughly 40% of a healthy person’s muscle force, assisting with hills, stairs and standing for hikers, people with joint pain, injury recovery or reduced mobility. Preorders are intended to ship next year; lower pricing depends partly on improved international supply chains, while rentals target customers pursuing a specific “dream hike.”
The hosts reject the assumption that AI-created wealth or adequate safeguards will appear automatically. Kevin Roose invokes the 50-year “Engel’s pause,” when Industrial Revolution productivity and profits increased without corresponding wage gains: abundance is “a deliberate policy choice.” Casey argues industry cannot be trusted to lead regulation once commercial incentives change, while a proposed 10-year absence of state-level rules could last through the arrival of much more powerful intelligence.
The nearer-term AI risk is a widening gap between technical capability and social readiness, especially for children, schools and defense. Roose favors AI literacy and redesigned assignments over classroom bans; Newton insists students still need reading and writing to develop critical thinking, concluding, “Real friends over AI friends.” Newton’s practical AGI benchmark is software that can perform routine assistant work, while household robotics remains slower: he predicts people will “probably still be folding” their own laundry in 2028.
Deep dive
1. Building, not invention, is the self-imposed bottleneck
Collison’s one-word answer to the most urgent abundance need was “building.” Science and medicine also move too slowly—people still get common colds—but physical construction has constraints he sees as especially unnecessary, self-imposed and, “at least in principle,” addressable.
His California contrast carries the argument: California once built ships in a day, and San Francisco was so “drunk on the fervor” of completing the Golden Gate Bridge that it built Treasure Island to celebrate. That culture has yielded to a high-speed-rail project whose former 2033 completion forecast has disappeared from its website.
Roose sharpened the contradiction with an AI forecaster who could imagine superintelligence and robot factories within years but conceded they would not clear Berkeley zoning. Collison’s parallel: America has repeatedly built cities, yet the Solano County proposal produced outrage while plans to “populate the cosmos” with superintelligence sound reasonable.
2. Better institutions may mean freeing insiders, not importing saviors
A host’s pushback—worth keeping—was that technologists may face resistance because social media and other technologies left people feeling they received a bad bargain, while environmental and civic consequences weakened trust. Collison accepted the “Homer car problem”: nobody wants outsiders imposing a grand, dysfunctional vision.
His alternative is to unleash people already “plowing some lonely furrow.” Asked whether they would change their research if current funding could be spent freely—not increased—79% of surveyed top-tier scientists said their agendas would change “a lot.” Collison also said 48% would be doing something “very different”; the transcript does not explain how that figure relates to the 79%.
That result influenced Arc Institute’s design and Collison’s broader institutional view: urban planners, scientists and other insiders often possess better ideas than the system permits them to manifest. The constraint may be committee allocation, consensus and narrow field definitions rather than a shortage of ambition or expertise.
3. Agents offload checkout labor while stablecoins export the dollar
Stripe sits beneath the interface, enabling transactions and money movement regardless of whether commerce uses mobile, stablecoins or an AI agent. Collison welcomed changing modalities, while teasing today’s Operator experience: it can buy a burrito “if you have two hours” for completion.
Internet commerce supports niche businesses serving “preposterously large” global audiences, but checkout remains an antiquated interruption: a “30-field form,” perhaps to be faxed, with “parts of it in Latin.” Few customers regard that administrative step as the internet’s great contribution.
Collison’s preferred framing avoids the grandiosity of “agents.” These “critters” can become “paperwork minions”: dispatch one to handle the transaction, then resume whatever else you were doing. He also imagined agents paying scientific-literature paywalls automatically, fixing a practical obstacle that currently trips up AI research tools.
He expects prices to remain denominated in local currencies—dollars in the United States, euros in Ireland. Yet the majority of cryptocurrency transactions are already in dollar-denominated stablecoins, he said, and many currencies have lost 75% in five years; crypto therefore might expand access to dollars for people outside the U.S. rather than overthrow the dollar.
4. AI could turn biology’s months-long loop into computation
Collison’s foundational claim was categorical but carefully defined: “Humanity has never cured a complex disease.” He separated infections and monogenic disorders from conditions produced by intertwined genetic, environmental and time-dependent factors, including most cancers, cardiovascular, neurodegenerative and autoimmune diseases.
The bottleneck is experimental iteration. Biological researchers manipulate physical cells, purchase expensive reagents across supply chains spanning multiple countries and wait three or six months for results; software engineers can test an idea at the command line and receive an answer roughly ten milliseconds later.
A useful virtual cell would let scientists test hypotheses “in silico” quickly and cheaply before committing to physical experiments. Collison said three recent capabilities create “shimmering potential”: sequencing individual cells to read them, transformers to reason across complex data, and CRISPR or functional genomics to write precise changes. Arc had published its first virtual-cell model the previous day.
5. Stripe hires road builders for decade-long product curves
Asked which trait companies undervalue, Collison chose long-term thinking. Infrastructure “can’t happen overnight”; Stripe has worked on it for 15 years and still feels it is beginning, while even its highly successful products can require a decade to play out.
His hiring analogy divides technology into cars and roads. Cars are visible, fast and glamorous, but “someone’s got to build the roads,” so Stripe seeks people predisposed to the slower infrastructure layer rather than products with “TikTok trajectories.”
Collison called AI summaries useful but “powerfully inoculating”: reading a television season’s plot on Wikipedia can cure the temptation to keep watching, yet it provides neither deep pleasure nor betterment. His book recommendations were The Dream Machine, a history of internet thinking published by Stripe Press, and The Demon Under the Microscope, about the antibiotic that preceded penicillin.
6. MO/GO targets mobility before mass-market bionics
Skip CEO Kathryn Zealand positioned MO/GO as an “e-bike for walking,” initially aimed at hikers limited by knee pain or reduced mobility, especially on inclines, steep hills and stairs. Longer term, Skip wants mobility impairments to stop determining what people can do.
Lightweight motors—made sufficiently efficient, affordable and compact only within roughly the past decade—can contribute up to about 40% of a healthy person’s muscle force. Cuffs integrated into the Arc’teryx pants transfer that force to the wearer and can assist standing, squatting and climbing.
Zealand expects broader use among people recovering from injuries or surgery, experiencing joint pain, or navigating pregnancy; Skip also researches Parkinson’s disease and more severe impairments. Truly able-bodied customers may find the product unnecessary, and anyone seeking resistance training can simply use ankle weights.
7. Wearable robotics must solve both fit and intent
“Hardware is hard,” but body-worn hardware raises the bar further: imperfect industrial robots can remain ugly, while exoskeletons must be comfortable, fitted and visually acceptable. Users are unusually sensitive to what touches and moves their bodies.
AI is the second constraint. Warehouse exoskeletons repeat predictable lifts, while people walking through a community fidget and move irregularly. In one prototype, Zealand tapped her foot during a nervous meeting; the system mistook it for an attempt to stand, threw her leg upward and spilled the table’s water.
During the live StairMaster demo, Newton described powered climbing as the physical equivalent of vibe coding: “I’m vibe walking.” Because the brain quickly normalizes assistance, Zealand switched the system off; the sudden difficulty made Newton immediately ask for it back. “This is genuinely how we often get customers,” she said.
MO/GO costs “$5,000-ish,” comparable to an e-bike, and can be preordered for shipping next year. Skip also offers rentals and experiences for people pursuing a dream hike; beyond friendly discount codes, Zealand said lower prices require improved international supply chains.
8. AI adoption is outrunning schools, families and household robots
Roose’s primary concern is the pace of change. San Francisco once felt about 18 months ahead of elsewhere; he now sees the gap widening while society lacks adequate safety nets, energy to absorb disruption or confidence that even well-intentioned systems will avoid social media’s unintended consequences.
Schools initially tried to ban and detect ChatGPT, which Roose now calls a mistake. He favors AI literacy, more classroom discussion and redesigned homework because graduates will use these tools; Newton’s condition is that AI not substitute so thoroughly for reading and writing that students never develop critical thinking.
Newton is more alarmed by AI companions than plagiarism: teenagers may decide ChatGPT is a better friend than anyone they know, with “nothing preventing us from walking down that path.” He also highlighted OpenAI signing a military contract and a Mattel partnership in the same week, despite unresolved questions about safe use by children. “Real friends over AI friends.”
Newton’s practical definition of AGI is software capable of routine assistant work—scheduling, advertising sales and subscriber support—rather than an abstract marketing threshold. Embodied automation is further away: despite billions flowing to companies such as Physical Intelligence, Newton expects people will “probably still” fold their own laundry in 2028.
9. Military use, regulation and distribution are policy choices
Newton described the defense “foot in the door” as an AI assistant inside the Pentagon answering questions such as which missile belongs on a jet. Facial-recognition technology is “basically perfect,” he said; combine it with AI and “things get dark in a hurry.” He sees sustained reporting on military deployments as a core journalistic responsibility.
Roose shared the concern, noting that early AI builders sought contractual restrictions on military use; DeepMind’s Google acquisition once included such a clause, though the position later changed. The countervailing force is the Defense Department’s huge budget and willingness to spend heavily to catch up.
Regulation cannot be delegated to industry, Newton argued: challengers promise safer systems and strong rules, then revise those views after becoming successful. Groups including the Future of Life Institute have mobilized around a potential 10-year period without state AI regulation, but Newton warned that ten years could carry the industry to “the end of the ballgame.”
Roose was equally blunt about distribution. Technological abundance does not automatically trickle down; it requires deliberate policy that many people oppose. During the 50-year “Engel’s pause,” Industrial Revolution productivity and profits increased while wages did not: “It sucked for a lot of people for a long time.”