CEO Dinner Insights: October 2025

Editor’s Note:

This month’s gathering featured three Jeffersonian questions: Are we in a bubble? How are we combining different kinds of intelligence (poly-intelligence) to drive discoveries? And what life hacks are we using to leverage technology in meaningful ways?

The bubble discussion revealed a fascinating split: while most agreed we’re in an investment bubble, leaders disagreed on whether this matters. Some see troubling financial mechanics—long-term debt financing short-term assets—while others pointed to massive unmet demand for compute that could enable a soft landing. The consensus: like the dot-com era, there will be spectacular failures and extraordinary winners, with the forest fire clearing brush for the strong trees to thrive.

The poly-intelligence conversation surfaced the most exciting theme: we’re entering an era where knowledge scarcity is being replaced by knowledge abundance. Leaders shared examples of breaking down disciplinary silos in biotech, using AI to accelerate cross-training among deep specialists, and enabling people to participate in fields that previously demanded monastic dedication. The recurring insight: the real competitive advantage lies not in hiring the smartest specialists, but in creating systems that allow diverse intelligences—human, machine, and natural—to collaborate at unprecedented speed.

Life hacks ranged from practical (Claude for task management, NotebookLM for company knowledge) to profound (Sunday family meetings, earning the right for adult children to want to hang out with you). The through-line: technology works best when it augments human connection rather than replacing it.

Twelve technology leaders gathered to discuss artificial intelligence, knowledge integration, and the future of human capability, revealing five critical insights:

The Compute Arms Race Is Creating Strange Market Dynamics

Training compute operates as a pure game theory problem with no principled pricing mechanism. Labs buy compute based on competitor spending rather than ROI calculations, creating an arms race mediated by a single supplier who benefits from the escalation. Meanwhile, inference compute faces massive unmet demand—Google is 3x oversubscribed for basic enterprise AI applications, not experimental use cases. The financial mechanics are concerning: companies finance capital expenditures with 40-year debt while depreciating assets over 2.5 years, essentially funding OpEx with long-term debt. Yet this may enable a soft landing rather than a crash, as real productivity gains absorb the capacity.

“There’s so much demand and they cannot offer enough compute to satisfy the amount of demand that companies want. 3x oversubscribed. And these are not extra experiments. It’s really, really basic stuff where it’s like we’re trying to analyze data.”

Knowledge Abundance Rewrites Competitive Strategy

The transition from knowledge scarcity to knowledge abundance fundamentally changes what creates competitive advantage. Hiring the smartest specialists no longer makes sense when AI provides expert-level knowledge across disciplines. The new strategic imperative: build systems that enable diverse intelligences to collaborate. One biotech CEO described accelerating cross-training among scientists who spent 20 years studying single cell types, using AI tools to enable hard conversations across disciplines at unprecedented pace. The shift requires moving from forecasting (predicting the future with scarce knowledge) to falsification (testing whether bad outcomes resulted from missing knowledge rather than bad luck).

“We have these debates coming: is AI leveling the playing field so mediocre software people are as good as great? I’ve always thought, no. The rich get richer. Really great people can use the tools and do amazing things.”

Disciplinary Boundaries Are Artificial Constructs Ready to Collapse

Human knowledge disciplines exist as simplified taxonomies for limited human cognition, not as natural divisions. As AI augmentation expands cognitive capacity, these boundaries become unnecessary constraints. Fields previously demanding personality self-selection—programming’s monastic dedication, science’s narrow specialization—now open to diverse participants. This enables horizontal thinking similar to pre-1900s scholars who were simultaneously scientists, philosophers, and mathematicians. The practical implication: specialization-based hiring becomes obsolete. One CEO questioned why they need specialist salespeople when communicators can use AI tools for technical depth.

“I think we are now coming to a world in which some hard things are just not hard. You’re going to find really interesting people emerge over the next five to ten years that are going to look very different from our prototypical ‘this is the software engineer, this is the sales guy.’”

The Investment Bubble Mirrors Dot-Com But With Critical Differences

Leaders unanimously agreed we’re in an investment bubble, with valuations detached from fundamentals and massive capital deployed without clear returns. However, this bubble differs from dot-com in crucial ways. Web 1.0 burned money on advertising and hiring; AI investment builds infrastructure—data centers, GPUs, power generation. Like laying fiber optic cables that remained valuable after telecom bankruptcies, AI infrastructure will persist beyond company failures. The bubble also concentrates wealth differently: revenue concentration creates fragility (some “neo-clouds” derive 50-75% of revenue from single customers), while a handful of survivors will acquire failed competitors’ assets at discounts.

“It feels like we’re playing a venture capital game with a significant fraction of the investment economy. When Softbank tried that with Vision Fund, we’re now doing VC at 1000x, 10,000x scale.”

Nature Remains the Ultimate Compute Resource to Integrate

Multiple leaders emphasized that evolution represents billions of years of computation that cannot be ignored. Scientific breakthroughs increasingly occur at intersections of PhD-level fields, and AI enables accessing 100 fields simultaneously rather than the typical one or two. However, pure machine learning approaches miss crucial insights embedded in biological systems. The epigenome—the “software layer” controlling how identical DNA creates different cell types—remains dark matter that could unlock new therapeutic classes. Physical robot experiments in labs can test 3,000 chemical combinations continuously, but the real acceleration comes from combining AI simulation, human expertise, and biological principles.

“Most advances in human discovery occur at the intersection of two PhD fields. There are very few humans that have PhDs in more than one field. Now you have an AI that has a PhD in 100 fields.”

The Problem: AI compute purchasing operates as a pure arms race with no rational pricing mechanism, creating potentially unsustainable financial structures.

One executive revealed the stark reality: “Every lab is looking at: if I’m spending this much in compute, but my competitor is going to spend twice as much, they’re going to have a model that’s X percent better. They’re going to pull forward the future by one year.” No one has principled methods for determining compute budgets—decisions are made by attending “the right parties in San Francisco” and learning competitor spending. An arms dealer in the middle benefits from escalation.

The financial mechanics compound the risk. One leader noted companies spending $40 billion financed with 40-year debt while depreciating assets over 2.5 years: “Are people actually funding OpEx with long-term debt? That is a big, big problem.” Yet countervailing evidence suggests genuine demand absorption. Google faces 3x oversubscription for basic enterprise compute needs—not experimental moonshots, but fundamental data analysis applications. Companies request $200 million in compute capacity and receive only $70 million.

The inference side presents different dynamics. As models improve, crappier models thinking longer approximate better future models, effectively pulling tomorrow’s capabilities into today. A glut of inference compute could accelerate AI adoption by making advanced capabilities available earlier.

The Insight: We’re witnessing simultaneous bubble dynamics (irrational training compute arms race) and genuine scarcity (unmet inference demand). The outcome depends on whether productivity gains absorb capacity fast enough to prevent financial unwinding.

Leadership Implication: Separate your AI investment strategy into training and inference. For training, recognize you’re playing game theory against competitors with limited visibility. For inference, identify high-value applications with measurable ROI where compute scarcity currently limits deployment. The companies that survive will be those treating this as infrastructure investment rather than capability speculation.

The Problem: Organizational structures, hiring practices, and strategic frameworks assume knowledge is scarce, creating systematic disadvantages as AI makes knowledge abundant.

One executive framed the fundamental shift: “None of us have really fully internalized or deeply taken advantage of this change around knowledge going from a scarce resource to an abundant one.” This transformation invalidates core assumptions. Hiring the smartest people made sense when knowledge was scarce; now AI provides expert-level knowledge across domains. Forecasting made sense when predicting the future required scarce expertise; now the question shifts to falsification—when bad things happen, was it bad luck or missing knowledge that AI could have provided?

A biotech CEO demonstrated practical application: scientists trained for 20 years on single cell types believed they “can’t possibly do anything” outside their narrow expertise. AI tools enabled them to “communicate for the first time and have really interesting hard conversations at a really accelerated pace.” Initial resistance (”what do you mean goals? I just do experiments”) transformed into thriving on aggressive OKRs within four years. The key: creating optionality and removing the feeling of risk when operating at higher complexity.

The shift enables horizontal knowledge integration resembling pre-1900s scholars who were simultaneously scientists, philosophers, and mathematicians. Fields requiring monastic dedication—programming’s narrow focus, science’s specialization—now open to diverse participants.

The Insight: Competitive advantage shifts from accessing scarce knowledge to orchestrating abundant knowledge. Organizations clinging to specialist-based hierarchies will be systematically outmaneuvered by those building systems enabling diverse intelligences to collaborate.

Leadership Implication: Audit every process assuming knowledge scarcity: specialist hiring, expert consultation, forecasting exercises, research timelines. Replace scarcity-based frameworks with abundance-based alternatives: generalist hiring with AI augmentation, real-time knowledge access, falsification over prediction, rapid iteration over careful planning. The winners will be those who redesign organizations around knowledge abundance before competitors recognize the shift.

The Problem: Human knowledge disciplines exist as simplified taxonomies for limited cognition, not natural divisions, creating artificial barriers to breakthrough innovation.

One leader observed: “All of these human disciplines we’ve created are not organic. It doesn’t have to be the case that these different segments of knowledge need to be separated from each other.” The boundaries exist because humans needed simple taxonomies to navigate complexity. As AI augmentation expands cognitive capacity, these constraints become unnecessary.

The practical implications are profound. Fields previously demanding specific personalities through self-selection—programming required monastic dedication, sales required specific communication styles—now open to diverse participants. One CEO questioned fundamental hiring categories: “I don’t really want specialist salespeople anymore. How do I find the person that’s good at both communication and can use these tools enough to get by with the technical stuff? Why do I need all of these specialties?”

Scientific discovery increasingly occurs at disciplinary intersections. One executive noted that OpenAI’s Kevin Weil identified “most advances in human discovery occur at the intersection of two PhD fields. Very few humans have PhDs in more than one field. Now you have an AI that has a PhD in 100 fields.”

The biotech sector demonstrates real-world application. After decades focusing on the 3 billion nucleotide genome, the field now races to understand the epigenome—the “software layer” of chemistry modifications that differentiate cell types. This “dark matter” requires integrating molecular biology, chemistry, data science, and medical records at unprecedented scale.

The Insight: The most significant innovations will come from collapsing disciplinary boundaries rather than advancing within them. Organizations that redesign around interdisciplinary collaboration rather than specialist depth will capture disproportionate value.

Leadership Implication: Eliminate specialist-based organizational structures. Instead of hiring the best programmer, the best salesperson, and the best analyst, hire people with strong communication skills and judgment who can use AI to access specialist knowledge across domains. Create evaluation frameworks testing interdisciplinary synthesis rather than domain depth. The temporary advantage belongs to those who act while competitors remain attached to specialization-based models.

The Problem: Current AI investment mirrors dot-com bubble dynamics but with fundamentally different capital allocation, creating uncertainty about crash severity and recovery speed.

Leaders unanimously agreed we’re in a bubble—valuations detached from fundamentals, massive capital deployment without clear returns, revenue concentration creating fragility. One executive noted: “It feels like we’re playing a venture capital game with a significant fraction of the investment economy.” The scale dwarfs previous bubbles: Softbank’s Vision Fund tried venture capital at 100x scale; AI represents 1000-10,000x.

However, crucial differences emerged. One leader contrasted: “In the dot-com bubble, people got venture money and spent it on advertising and hiring. A lot of this bubble is actually infrastructure bubble.” Spending on GPUs, data centers, and power generation resembles laying fiber optic cables—assets that retained value after telecom bankruptcies. Even if companies fail, the infrastructure persists for survivors to acquire at discounts.

The forest fire metaphor captured the dynamic: “When forest fires sweep through, they’re incredibly healthy. The forest gets really overgrown, there’s all this brush, and the forest fire takes out all the weak weeds. But the strong big old trees get singed around the edges, their core remains strong. After the forest fire, they actually start to thrive even more once all the brush is cleaned up.”

Yet concerning dynamics persist. Revenue concentration creates fragility: “Neo-clouds should report revenue as how much money do I have without my number one customer versus how much money do I have. Sometimes the gap is 50 to 75%.” Data scaling limits compound the problem: “For every doubling in computation, you need 40% more data. We’re kind of out of general purpose data for the LLMs.”

The Insight: The bubble will produce spectacular failures and extraordinary winners, but infrastructure investment means survivors inherit valuable assets rather than worthless advertising spend. The critical question: which companies have genuine demand absorption versus financial engineering.

Leadership Implication: Position for the post-bubble environment rather than trying to avoid the bubble. If you’re building infrastructure, ensure you can survive long enough to acquire failed competitors’ assets. If you’re consuming infrastructure, prepare for consolidation and shifting power dynamics as suppliers collapse or merge. If you’re investing, distinguish between companies with real demand absorption and those dependent on continued capital infusion.

The Problem: Pure machine learning approaches ignore billions of years of evolutionary computation embedded in biological systems, missing crucial shortcuts to breakthrough discoveries.

Multiple leaders emphasized evolution as massive compute expenditure that cannot be ignored. One noted: “Nature is one of the biggest compute hogs out there, having done evolution for so long to get to a particular solution.” Another observed that beautiful art represents “the artist having spent a lot of their own compute to get to the simplicity and beauty.”

The biotech sector demonstrates the opportunity. After spending decades understanding the genome’s 3 billion nucleotides, the field now confronts the epigenome—chemistry modifications controlling how identical DNA creates different cell types. This “software layer” represents evolutionary solutions to complex problems that remain “dark matter, we don’t understand it.” A race to generate epigenomic data correlated with medical records could enable AI to “understand this freaking biology, develop insight, understand what’s the source of diseases and hopefully use it for new classes of therapeutics.”

One executive described pioneering platelet-rich plasma applications, creating an “anti-PubMed” to surface negative research results that never get published. The goal: “accelerate serendipity” by enabling “crazy possible correlations across different disciplines.” Physical robots in labs now test 3,000 chemical combinations continuously, but integration with biological principles remains crucial.

A biotech CEO emphasized balancing machine and human intelligence: “I don’t think the machines are going to be able to do it better than the humans. The machines will augment the humans. We can add nature in—there is so much insight there. Thinking we can do it without some of those principles, we’ll go much slower.”

The Insight: The fastest path to breakthrough discoveries combines AI computation, human expertise, and evolutionary principles embedded in biological systems. Pure machine learning approaches that ignore nature’s solutions will be systematically slower than integrated approaches.

Leadership Implication: For any complex problem domain, map the relevant natural systems that have evolved solutions over millions of years. Invest in extracting principles from biological, materials, or other natural systems rather than assuming AI can derive solutions from scratch. Create interdisciplinary teams that can translate between machine learning, domain expertise, and natural system principles.

Two autonomous vehicle leaders provided stark contrast to the 2018 bubble when 133 companies pursued self-driving technology in California. One executive recalled: “280 billion were dropped in that industry. General Motors lit 10 billion-plus on fire and walked away from it.” The current environment shows rationalization—Aurora operates as the only company driving trucks at 70 mph on freeways, while Zoox launched people-movers in Las Vegas transporting 1,000 riders daily.

The business model evolution reflects maturity. Aurora’s approach: trucking companies buy trucks, order them with Aurora driver systems, and sign subscription services—slotting directly into existing capital and operating expense structures. The unit economics target 250,000 miles annually (versus 150,000 for human drivers) at roughly $1 per mile in gross margin.

A fundamental challenge persists: “The whole EV thing, we are totally not compact. U.S. transportation still doesn’t understand that it’s about a computer on wheels as opposed to a car with a computer.” Traditional automakers face structural disadvantages beyond technology—union heritage creates cost structures where they’re “basically a healthcare business that happens to make cars on the side.”

Key Insight: Autonomous transportation has moved from bubble to execution phase, with survivors demonstrating viable business models and actual deployment. Success requires treating vehicles as computers with wheels rather than cars with computers.

Leaders discussed enterprise AI adoption revealing the gap between proof-of-concept and production. Task management using Claude’s voice input with David Allen’s GTD framework shows practical wins, while ambitious leaders use AI to answer “what do I need to do today?” by analyzing Slack, email, and artifacts.

The data infrastructure opportunity remains massive. A top-10 bank maintains 1.5 million pages of standard operating procedures “because you have to design to the lowest common denominator when human beings are involved.” One executive explained: “Operating procedures are like code for people. Engineers write code that runs on AWS, humans write operating procedures that run on human labor.” AI conversion of human-readable procedures into executable code represents enormous automation potential.

However, organizational resistance creates friction. One leader noted: “Operationally, people are in the way of advancing AI because advances mean job loss. Most people are thinking how they can just hold on to their job.”

Key Insight: The largest enterprise software opportunity involves automating millions of standard operating procedures, but success requires addressing organizational resistance and job displacement concerns rather than purely technical challenges.

Leaders described revolutionary approaches to drug discovery combining AI with biological principles. One executive’s company works on therapy for their own child, leveraging AI to accelerate cross-training among narrow specialists: “Scientists trained for 20 years on one cell type think they can’t possibly do anything. But these tools enable them to communicate and have hard conversations at accelerated pace.”

The epigenome represents the frontier: the “software layer” of chemistry modifications controlling cell differentiation remains poorly understood despite complete genome mapping. Generating epigenomic data correlated with medical records could unlock “new classes of therapeutics. Everything we learned during the last 30 years, we’re going to outpace it within the next five years.”

Physical automation accelerates iteration: thinking machines use robots for lab experiments, testing 3,000 chemical combinations continuously. For fusion research, simulation enables varying magnetic fields, energy inputs, and electrostatic fields against clear objective functions.

Key Insight: Scientific discovery acceleration requires combining AI computation, human cross-disciplinary collaboration, and biological/physical principles rather than relying on pure machine learning approaches.

A CEO shared how Charles Schwab’s Dave Pottruck addressed power dynamic problems in executive teams. After becoming CEO, he observed people using his name to justify resource grabs: “Dave Pottruck said we need to do this” became a justification card. He clarified: “Unless I tell you this is a gun, you need to treat it as a light bulb. When I say an idea, it’s just a lightbulb—just a feature, like our website might look better with lighter blues. But if I tell you this is the gun, I will be very specific.”

One leader built on this: “You should never say ‘because Chris said so.’ If it says that, either I failed to explain why, or you didn’t understand the why. In either case, you should come back.”

Application: Create explicit language distinguishing directives from ideas. Power dynamics cause casual suggestions to become mandates, wasting resources and reducing agency. Clear signals preserve operational flexibility while maintaining accountability.

One executive described a frequent practice: “When I’m not being directive, I’ll say ‘I’m not being directive.’ Because by saying isn’t that direction, people could have great ideas. The power dynamic means when the CEO says something, everything flies. By thinking and seeing through that, it helps people get off autopilot.”

This connects to another leader’s practice of requiring “If you’re using my name, I need to see how you’re using it. You need that to be open to inspection.” The transparency curbs resource grabs and name-dropping while forcing clearer reasoning.

Application: Develop explicit signals distinguishing exploration from execution. Without clear markers, every CEO comment becomes a directive, crushing agency and forcing poor resource allocation. Pattern interrupts preserve collaborative problem-solving.

One CEO shared their highest-impact productivity hack: “For each of our companies, all of our companies in our portfolio, every company has a NotebookLM workbook. I dump everything we know about it in the workbook. It saves me hours per week.” The system creates accessible institutional knowledge without requiring manual organization or synthesis.

Application: Create company-specific NotebookLM repositories rather than relying on scattered documents and tribal knowledge. The compound effect of instantly accessible, synthesized information dramatically reduces context-switching overhead and improves decision quality.

One leader emphasized: “The Sunday family meeting—really finding a time to sit down with your family, your spouse and kids, and really take the time to discuss how everybody’s doing and what everybody has coming up and who needs help, support or encouragement.” This creates structured space for coordination that prevents scheduling conflicts and ensures resource allocation.

Application: Establish recurring family governance rhythms similar to business practices. Weekly coordination meetings prevent emergencies and create space for proactive support rather than reactive crisis management.

Despite massive infrastructure investment, genuine compute scarcity persists for enterprise applications. One executive revealed Google faces 3x oversubscription for basic AI services—not moonshot experiments, but fundamental data analysis. Companies request $200 million in compute capacity and receive $70 million allocations.

This scarcity creates unusual market dynamics. On the training side, labs make purchasing decisions based on competitor spending rather than ROI calculations, with no one having “any real principled way of deciding how much” to invest. One leader noted: “You can go to the right parties in San Francisco and have a sense of what someone else is buying.”

Market Implication: Genuine demand absorption could enable soft landing despite concerning financial mechanics. Companies focusing on high-value inference applications with clear ROI will capture disproportionate value as compute supply expands.

Multiple leaders noted extreme customer concentration in “neo-cloud” providers: “Sometimes the gap is 50 to 75%” between revenue with and without the largest customer. This mirrors earlier patterns where companies would report “Google revenue with Groupon versus Google revenue without Groupon.”

One executive observed similar dynamics in AI companies: “Revenue without Cursor versus revenue with Cursor” represents meaningful differences. This concentration creates fragility as single customer decisions can eliminate majority revenue streams.

Market Implication: Evaluate AI infrastructure providers based on customer diversification rather than absolute revenue scale. Concentrated revenue structures will create consolidation opportunities as single customer losses trigger distress sales.

One leader identified a fundamental constraint: “For every doubling in computation, you need 40% more data. We’re kind of out of general purpose data for the LLMs.” This creates divergence between general-purpose models facing data limitations and domain-specific applications where proprietary data enables continued scaling.

The implication extends to market structure: “Smaller networks doing things in really interesting science domains where the data is not publicly available” will see continued advancement while general models plateau.

Market Implication: Domain-specific AI applications with proprietary data moats will capture disproportionate value as general-purpose model improvements slow. Focus on verticals with rich, unexploited data rather than horizontal AI platforms.

A biotech CEO described the fundamental shift from genome to epigenome focus: “We spent decades understanding 3 billion nucleotides that each of us has. All those nucleotides are shared across all cells in our body. What makes your eye cell act different than your heart cell and liver cell is not that genome. It’s the software layer—the epigenome, the chemistry modifications.”

Despite being “completely dark matter, we don’t understand it,” the race to generate epigenomic data correlated with medical records could enable AI to unlock “new classes of therapeutics. Everything we learned during the last 30 years, we’re going to outpace it within the next five years.”

Leadership Lesson: The most valuable pivots often involve recognizing that the problem you’ve been solving is secondary to a deeper layer you’ve been ignoring. After decades of genome focus, the field realizes the control mechanisms matter more than the base code. Leaders who identify these layer shifts early capture disproportionate value.

One executive described attempting to build an “anti-PubMed” for medical research: “Most research in medicine never sees the light of day. People go to conferences, information gets presented and dies.” The goal: “Accelerate serendipity by doing crazy possible correlations across different disciplines.”

The project ran out of compute a decade ago, but the vision remains relevant: combining negative results (failed experiments never published) with cross-disciplinary correlation could dramatically accelerate scientific discovery. Now compute abundance makes the vision achievable.

Leadership Lesson: Sometimes being early means being wrong, but the vision remains valid. Ideas that failed due to technical constraints deserve revisiting as enabling technologies emerge. The leaders who maintain conviction in sound concepts despite early failures can capitalize when constraints lift.

One leader shared their approach to maintaining relationships with adult children: “You make them during the year after Christmas discuss and decide where the family is going on vacation the following summer and you pay for it and make it really nice. You tell them partners are welcome. A free ticket to a really nice place with a guest if it’s not a significant other always works.”

The intent: “Getting your kids as they get older, doing everything I can right now to earn for my kids to want to hang out with me as they are adulting.” This represents intentional relationship investment rather than assuming family connections persist automatically.

Leadership Lesson: The most important relationships require intentional design and investment. Assuming adult children will naturally want to spend time with parents ignores the reality that relationships require value creation on both sides. Creating compelling shared experiences builds relationship capital that persists beyond obligation.

“Really great people can use the tools and do amazing things. I think it is going to be a synergy of super smart people plus the incredible power of AI.”

→ AI amplifies talent differences rather than compressing them—the rich get richer.

“We’re kind of out of general purpose data for the LLMs. Smaller networks in science domains where data is not publicly available will see interesting things.”

→ Domain-specific applications with proprietary data will outperform general-purpose models.

“Every lab looks at: if my competitor spends twice as much compute, they’ll pull forward the future by one year.”

→ AI development operates as pure game theory with no rational pricing mechanism.

“If it doesn’t benefit a lot of people, we will get pitchforks—whether they are virtual or physical.”

→ Concentration of AI benefits will trigger societal backlash through various mechanisms.

“Compute is becoming a new general fungible resource alongside money and time.”

→ Compute joins currency and labor-hours as fundamental economic building blocks.

“None of us have fully internalized this change around knowledge going from a scarce resource to an abundant one.”

→ Most competitive strategies still assume knowledge scarcity, creating systematic disadvantages.

“Most advances occur at the intersection of two PhD fields. Now you have an AI that has a PhD in 100 fields.”

→ Interdisciplinary synthesis becomes the primary value creation mechanism.

“We’re not going to find really interesting people that look very different from our prototypical ‘this is the software engineer, this is the sales guy.’”

→ Professional archetypes dissolve as AI removes barriers requiring specific personalities.

“Nature is one of the biggest compute hogs out there—evolution for so long to get to a particular solution.”

→ Biological systems embed billions of years of computation that pure ML approaches ignore.

“Unless I tell you it’s a gun, treat it as a light bulb.”

→ Clear signals distinguish directives from ideas, preserving agency despite power dynamics.

“You should never say ‘because Chris said so.’ Either I failed to explain why, or you didn’t understand.”

→ Authority-based justifications indicate communication failures requiring correction.

“If you’re using my name, I need to see how you’re using it.”

→ Transparency around leadership invocation prevents resource grabs and name-dropping.

“By giving them options and the permission that there’s enough optionality, you can throw things away.”

→ Abundant alternatives reduce risk perception, enabling higher-complexity work.

“It feels like we’re playing VC at 1000x, 10,000x scale with a significant fraction of the investment economy.”

→ AI investment represents unprecedented capital concentration in speculative technology.

“Neo-clouds: revenue without their number one customer is sometimes 50-75% lower.”

→ Extreme customer concentration creates fragility masked by absolute revenue scale.

“The forest fire takes out weak weeds, but strong trees get singed and then thrive even more.”

→ Bubbles serve as clearing mechanisms benefiting survivors rather than pure destruction.

“We laid fiber optic cables that remained valuable after bankruptcies.”

→ Infrastructure bubbles create persistent assets unlike advertising-based bubbles.

Started in 2008, CEO Dinner is a monthly gathering of leading entrepreneurs in Silicon Valley.

© 2025 Dion Lim