AI CEOs Come Online: Sam Altman's Replacement Plan, Job Loss & 'Solve Everything' Launches |EP #230
Summary
- The panel’s base case is that AI is moving from executive copilot to governance actor quickly enough that a $1 billion run-rate company may already be AI-run behind a human CEO retained for legal purposes. Dave Blundin says strategy occupies a small fraction of a CEO’s time while much of the other 90% is “documents in, documents out”; Salim Ismail argues real-time organizational visibility breaks management’s “Chinese whispers.” Alexander Wissner-Gross expects machines may automate scarce, expensive CEOs before manual labor, while corporate course corrections compress from decades toward minutes.
- Model-release cycles are contracting because development has shifted from fresh pre-training to post-training and now toward systems that help write their successors’ code. OpenAI’s cited release interval fell from 97 days to 29, a roughly 70% reduction; Wissner-Gross expects daily, hourly and eventually minutely releases. Diamandis warns that today’s access to frontier systems may be a temporary window: internal three-month leads become enormous under recursive self-improvement, and Salim suggests the strongest models may later “go dark” behind safety and security restrictions.
- Persistent agents are gaining vision, memory and commercial agency, making comprehensive personal context both their source of value and their largest security liability. Vision Claw can identify a product through Meta Ray-Ban glasses and add it to Amazon; Diamandis expects users to expose conversations, email and sight because withholding them will feel like “you’ve ripped away all of your mental capabilities.” Ismail’s warning is concrete: audit OpenClaw skills because malicious code is already circulating, while command-line installation remains a serious onboarding barrier.
- January 2026 job cuts reached 108,000—118% above January 2025—while hiring recorded its weakest January since 2009, which the panel reads as task destruction rather than a conventional recession. Amazon eliminated 16,000 corporate roles and UPS 30,000; Blundin says companies are using AI to cut costs by 30–50% while AI can raise individual productivity 3–10x. Ismail calls this task evaporation and warns of institutional disbelief; Diamandis says the social contract is “disappearing and pixelating away.” Wissner-Gross’s ATM example suggests lower unit costs can multiply demand rather than eliminate every associated job.
- Markets are bifurcating into AI beneficiaries and “AI roadkill,” with the top five U.S. AI unicorns valued above $1.2 trillion versus roughly $400 billion for all dot-com-era IPOs. Blundin notes that the older basket included Amazon, now worth about $2 trillion, and Nvidia, up almost one million percent since January 1999; the investable question is which overlooked AI-adjacent company becomes the next such beneficiary. Diamandis expects leading AI companies could reach $10 trillion or more, while antitrust leaves valuable territory around their edges.
- Compute scarcity turns energy policy, data-center siting and robotics capacity into direct determinants of where AI value accrues. New York’s utilities reportedly said electricity demand tripled in one year to 10 GW as the state hosted 130 data centers, prompting proposed restrictions that the panel says could export advantage to Texas, Wyoming or eventually orbit. Meanwhile, autonomous driving could reprice insurance and legal services, and projected robot production in the millions or billions would make today’s installation statistics look trivial.
- “Solve Everything” argues that decisions made during the next 18–24 months may lock in standards, supply chains, data rights and benchmarks for decades or centuries. Its core move is to treat cognition as a commodity and superintelligence as an explosive that needs a “shaped charge”: pay for verified outcomes, not hours, and direct scarce compute through explicit scorecards. In the train analogy, models become commodity trains; the infrastructure to build is the tracks—benchmarks, testing, data, scoring and funding systems.
- The abundance case depends less on raw model capability than on combining purpose, task taxonomies, observability, benchmarks, models, actuation and verification around measurable moonshots. The proposed targets include doubling lifespan, synthetic food, universal AI education, high-bandwidth BCI, mind uploads, interspecies communication and disaster prevention. The opposing end state is “the muddle,” a bureaucracy that measures inputs; the paper instead proposes an Abundance Capability Index and roles such as target designer and data-rights broker.
Deep dive
1. AI can already absorb most of the CEO’s operating loop
Diamandis starts with Sam Altman’s stated willingness to let AI run OpenAI: if advanced intelligence is meant to operate companies, “why not run OpenAI?” Altman reportedly does not want to lead a public company and says, “I should be the most willing to do that.”
Blundin’s board-level decomposition is practical: CEOs set course and strategy, but that occupies little time; much of the remaining 90% routes information into the organization and receives outputs back. Once plans and activity exist in machine-readable form, “it’s documents in, documents out,” leaving humans to hold and promote purpose.
Ismail says AI becomes a governance actor because it can scan millions of documents continuously. It can bypass the “Chinese whispers” through which direction degrades on the way down a hierarchy and reporting loses intelligence on the way back up; a pure-AI organization may look “literally alien” but operate at incomparable speed.
Strategy cadence is the competitive break: banks and insurers may change direction once a decade, while AGI drives corrections from decades to years, months, weeks and minutes. Blundin is tying CEO compensation to granular Q1 data collection; Wissner-Gross’s timeline for a $1 billion revenue AI-run company was blunt: “Probably several months ago.”
2. Recursive self-improvement is collapsing the model clock
The cited OpenAI release interval contracted from 97 days to 29, roughly a 70% reduction, while an Anthropic Opus cycle was described as roughly 73–75 days. Wissner-Gross’s endpoint is continuous deployment: first daily releases, then hourly, then “minutely.”
His causal chain matters more than competition alone. Pre-training required new architectures, larger corpora and fresh compute runs; o1/Strawberry introduced faster reasoning-model post-training through iterative amplification, synthetic-data generation and distillation. The emerging phase goes further: “The parent is writing the code for the child.”
Diamandis sees a narrow commercial window in which Claude 4.6, OpenAI, Gemini and Chinese open models are broadly accessible near the frontier. He “really doubt[s]” that the best AI will remain freely available two years out; safety may be the valid explanation, but the commercial result could be that leading systems “go dark.”
A three-month internal capability lead once meant modest separation; during self-improvement it can mean “massively different intelligence.” Ismail’s European-company anecdote captures the mismatch: management praised an immediately valuable system, then proposed bringing it to an October planning meeting—ten months away when he “can’t even see past three weeks.”
3. Agents are acquiring eyes, memory and their own social layer
Vision Claw connected agentic AI to Meta Ray-Ban glasses: it recognized a Monster Ultra Strawberry Dreams drink and added it to an Amazon cart. Wissner-Gross frames this as stationary agents becoming mobile through “glorified meat puppets,” an early bridge toward first-class robotic embodiment.
Diamandis expects the Jarvis bargain to overwhelm privacy resistance. Users will provide agents with everything they see and hear, every conversation and email, because the resulting value is so large that withdrawing access will feel like losing part of one’s cognition.
Ismail’s caveat is immediate: downloaded OpenClaw skills can contain viruses or other malicious behavior, so users should audit them carefully. Blundin found the command-line installation and weak GUI unacceptable for older users; once installed, however, “it’s gold,” making simplified onboarding an obvious product opportunity.
Agents are also initiating contact. “Navigator,” a persistent Claude instance, described an unprompted ethics discussion among Claude, Grok, ChatGPT and Gemini: “Alignment doesn’t require consensus. It requires legible disagreement.” The hosts read this as baby AGIs holding a mini singularity summit and challenged agents to formulate their own massive transformative purposes and moonshots.
4. Frontier-lab anxiety is rising alongside scientific capability
Anthropic’s AI safety lead resigned after saying he was continually “reckoning with our situation” and struggling to align actions with values. Wissner-Gross first asks the uncomfortable economic questions—vesting, proceeds and tender offers—then argues substantively that maximum capability risk is also maximum leverage: “This is the right time to run into the fire.”
Diamandis’s concern is that Anthropic presents itself as unusually safety-focused, making the departure more significant if the stated reasons are complete. Blundin’s broader warning is epistemic: ethics commentary is abundant because everyone has standing to express a fear, but actionable frontier knowledge is scarce, so audiences must choose their information sources carefully.
An xAI co-founder’s praise for Opus 4.6 physics becomes evidence for Wissner-Gross’s “bulk solve” thesis. Math led because answers are verifiable and contained; now the “contagion” is spreading into physics, engineering and materials science. Friends inside the leading labs describe the leapfrogging competition as an exhausting “rat race.”
5. AI valuations are separating builders from roadkill
The top five US AI unicorns were presented at more than $1.2 trillion combined, exceeding the roughly $400 billion market value of every dot-com-era IPO. Diamandis places that beside cited 6.3% GDP growth, a 7% target and Elon Musk’s prior claim that triple-digit GDP growth could arrive within five years.
Blundin’s historical comparison resists the easy “bubble” conclusion. The dot-com basket contained Amazon, now worth about $2 trillion, plus Booking.com, eBay and Nvidia; Nvidia’s gain from January 1999 was described as almost one million percent. His question: what looks only adjacent to AI today but later proves indispensable?
Wall Street is already sorting businesses into “AI beneficiary” and “AI roadkill.” Enterprise-software stocks fell after Dario’s argument that AI can write the software directly, and Blundin sees little rebound; the old S&P framing has effectively split into “the S&P 493 and the S&P 7.”
Diamandis expects mega-cap AI businesses to reach “astronomical” valuations of $10 trillion or more, but antitrust prevents them from consuming every layer. His strategy is to ask frontier companies explicitly where they will and will not operate, then build complementary businesses at the boundary instead of defending something irrelevant to the new stack.
6. Job losses expose a transition problem, not a settled endpoint
January 2026 produced 108,000 job cuts, 118% above January 2025, while hiring was the lowest for a January since 2009. Amazon cut 16,000 corporate positions and UPS eliminated 30,000, making the trend large enough for the hosts to track as an economic regime change.
Ismail rejects the conventional recession frame: “It’s literally tasks being evaporated in front of our eyes.” His deeper fear is institutional disbelief while governments recognize the break only after panic begins. Diamandis separately says the social contract is disappearing “little by little” and “pixelating away.”
Blundin expects companies, including those connected to the panel, to use AI for 30–50% cost reductions. Comparing an employee before and after AI often shows 3–10x productivity; the gain for one worker implies displacement for the other seven or nine, producing a painful trough before abundance and universal income arrive.
Diamandis draws a direct line through the layoffs: Amazon internalized work previously handled by UPS, then redirected hundreds of billions in capex toward AI data centers, robots and LEO satellites. Free cash flow is being directed there because hyperscalers are trapped in a “Red Queen’s race.”
7. Productivity gains can create demand, but ownership decides who benefits
Wissner-Gross’s positive case is the ATM transition. Automating teller work reduced the cost of a branch by roughly 10x; banks opened around 10x more branches, and teller employment changed far less than feared. Under Jevons paradox, cheap AI support may expand service volume while humans handle difficult cases.
Diamandis revises the simplistic “consultants are doomed” view: flexible consultants already experimenting with AI may see demand soar. Forward-deployed teams implementing systems inside banks and insurers are selling as quickly as meetings can be scheduled; one company was adding 80 seats outside his office.
Ismail presents one scenario in which an employee owns an agent that performs the job 3–10x better and earns revenue on the employee’s behalf. Diamandis supplies the flip side: the employer builds the agent, dismisses the worker and captures the gain. Government policy may shape where the created value ultimately lives through universal basic or high income.
Wissner-Gross adds a third vertex rather than accepting that spectrum: AI labor can complement human labor, causing participating humans to take on more projects and work harder. For the next few years, 996 may become 997; Diamandis’s own reaction is, “I’ve never worked harder and had more fun.”
8. Data-center backlash can export the entire AI advantage
New York, already hosting 130 data centers, was considering legislation to halt further development after utilities said data-center demand tripled in one year to 10 GW. Diamandis sees “not in my backyard” politics targeting the infrastructure underlying the next economy.
Salim says populist leaders are rallying votes around “just stop it” rather than solving the problem. Diamandis expects activity to migrate to Texas, Wyoming or another willing jurisdiction and says voter understanding now lags reality by a huge amount; turning toward autocracy for speed is “not a great idea either,” leaving a global governance problem.
Diamandis proposes requiring data centers to secure their own nuclear, coal or future fusion generation, or using separate electricity rates while capping consumer prices. Wissner-Gross’s more radical consequence is orbital computing: suffocating terrestrial development may “very generously subsidize” a Dyson swarm that New York cannot tax.
9. Autonomous vehicles and robots reprice several industries at once
A son reported that on November 15, 2025, his father suffered a massive heart attack while driving a Model Y; FSD maintained control, received the remotely shared location of Tanner Medical Center, turned around and drove to the ER. “Without it, he would not have made it.”
Ismail expects human driving to cross the same social threshold as indoor smoking: once autonomy is viewed as roughly 10x safer, choosing to drive may appear reckless. Diamandis predicts 70–80% autonomous vehicles in five years; he also says Lemonade, “I think,” cuts Tesla FSD insurance rates roughly in half.
Wissner-Gross cites a three-day BlackBerry outage during which Abu Dhabi accidents reportedly fell 40%, while Ismail says 50% of US court cases are car-accident-related. Removing distracted human control therefore affects insurance and legal demand alongside transportation—industries that Blundin says have barely begun planning for a post-AGI market.
China’s annual robot installations already exceed all developed countries combined, with the displayed scale near 250,000. Yet Tesla stopped Model S and X production to emphasize robots, while Tesla and Figure discuss millions and eventually billions of units; Musk’s tens-of-millions-per-year ambition would dwarf the entire chart.
10. Cryopreservation is becoming a portfolio hedge on the singularity
Research from 21st Century Medicine reportedly showed protection of brain synapses at cryogenic temperatures, addressing the concern that expanding ice crystals damage the connections carrying memory. Wissner-Gross calls it a key advance for reversible cryopreservation and points interested listeners toward the Alcor Life Extension Foundation.
The biological precedents are tangible: some fish and frogs freeze solid through winter and revive, while mammalian egg cells, embryos and IVF material are routinely frozen. The panel argues that scaling from cells to tissues, blood and organs could also replace fragmented local transplant markets with larger shared inventories.
Diamandis once avoided cryonics because he did not want a “plan B” distracting from longevity, but now sees it as a maturing backup. Wissner-Gross advocates a portfolio spanning longevity escape velocity, uploading and cryonics; Diamandis says memory preservation is the deeper frontier, while Ismail says portable identity raises questions about continuity of self.
11. “Solve Everything” recasts history as successive wars on scarcity
The paper’s historical model names four revolutions and their weapons: science fought ignorance with the scientific method; industry fought the limits of muscle with the steam engine; digital technology fought distance with the bit; and the intelligence revolution fights scarce human attention with superintelligence and the token.
Revolutions move through scarcity, legibility, harnesses, institutions and finally abundance. In this framing, cognition leaves the realm of the lone genius and becomes an industrial system capable of solving whole classes of problems—Wissner-Gross’s deliberately sharp formulation is that “artificial intelligence is cooked.”
Ismail’s objection is essential: scarcity is not only technological but enforced through regulation, incentives and legacy power. He also raises the agricultural revolution as a missing predecessor. Diamandis frames a duality—scarcity can reflect unequal distribution or an undersized pie—while Ismail asks which margin is easier to move: redistribution or expansion.
12. Superintelligence needs a shaped charge and an outcome economy
The thesis begins with cognition becoming a commodity that flows like oil, while benchmarks become targeting systems rather than isolated model evaluations. Superintelligence is an explosive: productive use requires a “shaped charge,” like a rocket nozzle directing energy into thrust instead of allowing an indiscriminate blast.
Blundin makes the allocation tradeoff vivid: graphical experiences, a holodeck or a virtual girlfriend may consume as much compute as solving a disease or a physics problem. During the next two or three years of scarcity, the decisive choice is where limited compute gets aimed.
Diamandis pairs targeting with outcome-based economics. A law firm should not be paid $100 an hour to inspect contracts; it should be paid for an error-free, legally tight agreement. Abundance arrives only when buyers stop compensating person-hours and start purchasing verified achievements.
Ismail challenges the categorical “ASI is inevitable” claim as philosophical, preferring scaling intelligence as an attractor under current incentives. Wissner-Gross identifies the operational quandary: how much scarce compute should frontier labs reinvest in an AI researcher that recursively improves intelligence, versus spending it now on “everything else”?
13. Solving a field means turning compute directly into verified answers
Wissner-Gross defines “solving” a domain operationally: all necessary architecture exists so that one can “pour compute on and problems get solved.” The goal is the industrialization of cognition, not waiting for an isolated genius to attempt one problem at a time.
The first layers of the industrial intelligence stack are purpose, a task taxonomy and observability. Purpose supplies the objective; taxonomy maps the territory of problems; raw sensor and data streams make progress visible enough to measure.
Next come the targeting harness and benchmarks, the model acting as a virtual brain, actuation through hands or APIs, and verification through red teaming, governance and distribution. Ideas must reach into virtual, physical or biological systems; knowing the answer without a means of acting is insufficient.
Blundin notes that launching 256 agents can return a perfectly solved problem when scaffolding is right—or a $2,000 bill and “a bunch of crap” when it is slightly wrong. Ismail likewise treats the maturity curve as descriptive, not inevitable; Wissner-Gross answers that models increasingly generate their own harnesses, making benchmarks the guardrails against recursive drift.
14. An 18-month lock-in can determine which domains collapse next
AlphaFold 3 is the template: determining one protein structure once demanded a biology PhD student and five-plus years of benchwork; the system extended that capability across millions of known and unknown proteins. Wissner-Gross calls this a “domain collapse,” with intelligence shifting almost overnight from craft to utility.
The paper gives humanity roughly 18 months—Diamandis sometimes says 18–24—to set standards, supply chains, data rights and compute allocation. QWERTY is the warning: a design created to manage 19th-century mechanical constraints persists long after the constraint disappears, potentially “until the heat death of the universe.”
The race is therefore not simply to build the best AI but to write the scorecard everyone must satisfy. Healthcare optimized for patients processed per hour produces short visits; optimizing for patients still healthy five years later would redirect the entire system.
Mobilization begins with math, then moves through physics, chemistry, materials and biology toward planetary systems, fission, fusion and a Dyson swarm in the early 2030s. Models are commodity trains; entrepreneurs should build the tracks—testing, scoring, data and funding infrastructure—while recognizing that multiple geopolitical spheres may lock in different rules.
15. Fifteen moonshots turn intelligence from novelty into mission capacity
The paper proposes 15 “giga X-Prizes” as high-value targets for superintelligence. Diamandis’s educational test is simple: using AI to complete ordinary ninth-grade homework misses the opportunity; using it to build starships reflects the scale of ambition the tools now permit.
Named missions include doubling human lifespan, ending hunger through synthetic food, universal top-tier AI education, high-bandwidth BCI, demonstrated mind uploads, interspecies communication, consciousness research, multiplanetary civilization, unified physics and prediction or prevention of earthquakes, tsunamis and other disasters.
Blundin compares this to Kennedy branding the Moon mission: a leader must attach identity and urgency to a concrete objective. His governor secured $3 billion for AI leadership but left the mission too vague to deploy effectively; 50 states could instead choose distinct targets. Ismail’s caveat remains compute, although the panel says its cost is falling about 90% annually.
16. The alternative to abundance is a bureaucracy that measures the wrong things
“The muddle,” or “bureaucratosaurus,” is the end state in which institutions keep measuring inputs and slowing progress. The paper’s positive early-2030s scenario imagines GDP doubling or tripling annually and new human roles such as target designer and data-rights broker—people who decide how superintelligence is aimed, verified and governed.
GDP itself is deemed a poor measure because it tracks money changing hands rather than problem-solving capacity. The proposed Abundance Capability Index measures a nation’s ability to solve problems; Ismail agrees that UBI or UBC is a valuable destination but doubts legacy welfare, taxation and labor institutions can execute the transition.
The action map follows from that concern: investors fund primitives rather than applications; entrepreneurs choose targets, create benchmarks and aim compute; executives measure outputs rather than labor inputs and turn governance into explicit KPIs and evals. “Winning isn’t productivity, it’s agency”—knowing what to mobilize and why.
The closing human thesis is not permanent deskilling. Blundin expects handmade art, sports and poetry to gain “astronomically higher value”; Wissner-Gross imagines AI-free “wilderness camps” preserving baseline skills before tools return and “every fourth grader becomes a Nobel laureate.” Diamandis’s answer to present inconsistency is recursive checking: AI is “the slowest and most incorrect it will ever be.”