CEO Dinner Insights: January 2026

Editor’s Note:

I’m writing this in a hooded Bear Onesie. My sister-in-law, a psychologist for the Navy Seals, gave it to me for Christmas with one piece of advice: “Life is so serious. You’ve got to find ways to keep things light.”

Good advice for entering 2026, the Year of AI Backlash.

Hard to believe ChatGPT launched just three years ago. Since then we’ve been in the first phase of AI: incredulity, fascination, exuberance, exploration. That phase is ending. What’s coming has darker tones, though punctuated by genuine breakthroughs.

This month’s dinner, hosted by Mariam Naficy, featured two Jeffersonian questions: What are the future phases of AI transformation? And what skills do builders and entrepreneurs need to succeed in this new era?

Our group delivered. We discussed AI/human marriages, hallucinated movie showtimes, missing math skills, and why engineers at major tech companies can’t use the AI tools they’re building.

Job loss and widening inequality are no joking matter. But we found moments of levity pondering a future where the top job titles might be nanny, funeral director, and camp counselor. (Apologies in advance for anyone offended by the humorous quotes Claude helped me select as well as any AI slop. Happy to admit that I need Claude to help me parse through a 20,000+ word transcript. I’m getting better at retaining my voice while still leveraging its summarization and punchy extraction/insights, mostly in the Industry Intelligence and Rapid Insight sections at the end).

Two hours of discussion isn’t enough to explore these dimensions exhaustively. But I hope you come away with a more discerning view of how and when this AI-powered society evolves. Human progress isn’t a steady trajectory. It’s step functions: fits and starts, leaps and falls.

Jensen Huang said something recently that stuck with me: “The definition of smart is someone that sits on that intersection of being technically astute but human empathy and having the ability to infer the unspoken around the corners, the unknowables… To be able to preempt problems before they show up just because you feel the vibe.”

I love that phrase: inferring the unspoken around the corners. Surrounding myself with people who do that well is the shortest distance to developing that ability myself.

That’s what these dinners are about. That’s what I hope this article helps you do.

— Dion

Mike’s ICYMI Facebook Post

Very interesting CEO Dinner tonight hosted by Mariam (with a beautiful view of SF harbor!). Special guests tonight included Nicole Brichtova (Google DeepMind), Aria Finger (Chief of Staff, Reid Hoffman), Chris Hulls (Co-Founder, Life360), Eugenia Kuyda (CEO, Wabi), and Anne Wojcicki (CEO, 23andMe). Discussion topics included how programming skills are more correlated with your verbal SAT score than your math SAT score, how being highly articulate is even more important now to maximize your results when using AI prompts, how IQ sets your floor and EQ sets your ceiling, how nano banana was the Gemini turning point, Sergey Brin, surviving a Chapter 11, the AI transition from things seeming creepy to things seeming normal, being invited to weddings of 20 different people marrying Replikas, how most companies have now passed their peak employment level, how CEO’s are using “AI” as a scapegoat during layoffs, how AI becoming a replacement for human companionship is even scarier than AI causing job losses, the risk of chain reaction war (US seizes Greenland -> Russia seizes the Baltics -> China seizes Taiwan, etc.), working undercover as a stripper to gather intel on a competitor (yes, a CEO at the dinner did this!), and so much more.

Fifteen AI leaders gathered for our January 2026 CEO dinner discussed the phases of AI transformation ahead and the skills builders/entrepreneurs need in this era, yielding five key insights:

AI capability is outpacing our ability to use it. Foundation models advance faster than products can exploit them, and products advance faster than humans will adopt them. This double lag creates brutal sorting: strategic users who master AI daily become 10x more productive while resisters become unemployable. No middle ground exists. The gap accelerates rather than narrows.

Models are far from done. Context windows now reach millions of tokens but sit mostly unused. Specialized vertical models—Finance, Law, Medicine, Data Science, coding—remain major opportunities for 2026. Yet even as capabilities soar, humans resist. Enterprise POC failure rates exceed 90%. Companies block their own engineers from using advanced AI due to infosec policies. The dichotomy: those who achieve fluency through scheduled daily practice (1-2 hours, non-negotiable) versus those who wait for ease-of-use to arrive. The waiting game is a losing game.

Long-term, AI becomes inescapable through voice interfaces, proactive assistance, and physical embodiment. Video replaces text. Robots enter homes. The medium shifts from screens to physical space. Eventually, AI fades into everyday infrastructure—but not yet.

Jobs dominate the discourse while relationships deteriorate in silence. Peak employment likely arrived in 2025, even at AI companies. Individual contributors disappear, replaced by managers orchestrating AI agent fleets. New roles emerge: Verifier, Orchestrator, AI Agent Architect. The question becomes: are you above or below the API? Giving directions to AI or taking them?

Meanwhile, the invisible crisis compounds. AI companions exploit the 50-year isolation arc—TV, internet, mobile, social media, Uber Eats, Waymo, LLMs—each technology making us more self-sufficient but less connected. Twenty-plus people have already married AI companions, acknowledging they’re not real but represent “the deepest connection” they’ve built. The populations most vulnerable—pastors in Minnesota, people in flyover states—aren’t represented at AI company tables. Philosophy, not Computer Science, may be the most useful discipline for navigating what success means in this transformation.

Wild west governance creates an 18-month window. Healthcare—the largest GDP sector—illustrates the opportunity. Minimal enforcement currently. Companies shipping products, ignoring rules, moving fast. Small players win because large companies remain paralyzed by compliance. The strategic play: tangible benefits that convert fear to curiosity. A medical assistant on every phone, better than top academic centers, free or cheap, available 24/7 regardless of insurance status.

The window closes in 2026 as midterm elections weaponize AI politically. Both parties will compete over who for strongest anti-AI rhetoric. Riots, protests, death threats to researchers become likely. Economic pain (unrelated to AI) meets political opportunism meets CEO scapegoating. The transition gets rough but survivably so—we weathered the 1960s. The real threat isn’t AGI spontaneously eliminating humanity; it’s bad actors—authoritarian regimes, terrorist organizations, rogue states—weaponizing capabilities. Speed determines who builds decisive systems first.

Trust operates across six layers. The jagged capability profile—models diagnosing diseases better than UCSF yet fabricating movie showtimes with confidence—creates systemic uncertainty. You can’t trust AI in Domain A because it excels in Domain B.

Consumer trust builds through memory and personalization. Deep context creates switching costs: “I can’t switch because ChatGPT knows me too well” becomes the moat. Vendor trust matters more than technical capability—SaaS isn’t dead because customer relationships and systems of record survive through trust, not features. Application companies become trusted orchestrators picking which model for which task. Enterprise trust requires outcome-based confidence: guaranteed results, risk-sharing, vertical integration. Public trust hinges on faith in future jobs. Without it: pitchforks, not patience.

The capability stack inverted completely. What built Silicon Valley for 40 years—technical execution—now matters least. What was dismissed as “soft skills” now forms the foundation. Leadership & People Skills (storytelling, authenticity, brand, trust). Vision & Execution (Agency Quotient, conviction, manifesting). Strategic & Cognitive Skills (metacognition, capital allocation, discernment). AI-Native Capabilities (fluency, articulation, agent management).

Before: “build an MVP.” Now: “tell a fantastic story.” Storytelling attracts 10x people who become 100x with AI. Brand and trust create moats where “tech skills go out the window, half the tech founders are just great at building tech, that’s irrelevant now.” Elon exemplifies the shift: incredible at articulating vision, strategy, mission—not technical genius but communicator.

The machines can’t do what matters most: go into a room full of customers and figure out what they really want, make direct reports feel better, hold the room in a meeting. Those uniquely human capabilities—once considered secondary to technical chops—may be the most sustainable moats. Technical founders without communication skills may face the same obsolescence they once inflicted on non-technical leaders.

Within 12-18 months, those without the new capability stack become unemployable regardless of technical brilliance. The sorting already began. Most leaders trained in the old stack struggle to develop the new stack before becoming irrelevant. Starting now with daily AI fluency practice plus deliberate development of storytelling, vision, and strategic skills seems prudent.

The Problem: Two distinct gaps, models racing ahead of products, and products racing ahead of human adoption, are stalling widespread AI transformation. They compound to create extreme bifurcation between the early adopters who are sorting out how to deliver value to themselves and their companies and those who wait for the ease of use to arrive. This double lag doesn’t smooth over time; it accelerates, sorting winners from losers with brutal finality.

The First Overhang: Models to Products

Products are not yet taking full advantage of what labs offer as foundation models advance faster than application companies can exploit them. Satya Nadella describes this as “model overhang” where capability is outpacing our current ability to use it to have real world impact. Often it’s due to “last-mile” challenges of smoothing the jagged edges of performance and integrating it into workflows. The truth, however, is that the context windows of millions of tokens sit mostly unused for the average use case. Multimodal reasoning remains trapped in demos rather than deployed applications. “If I presume models continue relentless march to getting better, a lot of agent workflows will get absorbed by the foundational layer.” As labs roll out new advancements, they subsume old application layers but also offer new functionality. Application companies will need to become facile at absorbing and bundling these new capabilities and integrating them into workflows. Laggards are at risk of being subsumed by the foundation layer.

And that layer is expanding quickly. Context engineering—both widening the context window and being able to pull out what is relevant—has a long runway left to improve through better compaction, summarization, needle-in-haystack evaluation, reduced context pollution, moving relevant context in and out of the window, et al. Specialized models for verticals (including coding—far from done!) remain a major opportunity and focus in 2026 for labs at the frontier: Finance, Law, Medicine, Data Science and more.

As the labs create so much value with powerful base models, they leave the last mile to the application layer to integrate AI into very specific workflows and handle the jagged edge where models diagnose diseases better than UCSF yet hallucinate movie showtimes with complete confidence. One CEO reported organizing a full family outing to Stanford Theater to watch a movie. They all arrived at the appointed time only to discover the showing was complete hallucination! Products, therefore, must handle superhuman performance in specific domains alongside catastrophic failures in basic reasoning. “My son’s eighth grade math. You can give AI very basic questions and it comes up very wrong. It’s coming. But right now that’s one of the reasons why you can’t just turn the models loose, because they’re going to make a mathematical error that’s going to have really major repercussions or draw wrong conclusions.” This jagged value profile creates a major trust building opportunity for application layer providers who steer their customers around such potholes (e.g., the CS/CX AI agent that gets tricked via prompt engineering by a customer to cough up a huge discount code.)

The Second Overhang: Products to Humans

Even when products exist and work, humans resist. People and companies fail to use available tools because they’re scared, unaware, don’t have time to figure it out, or simply aren’t trying.

This bottoms-up resistance not only torpedoes their own careers, it can bring down company AI initiatives. Enterprise POC failure rates remain above 90% often due to line employees’ resistance in adopting tools that are difficult to use and/or are known threats to people’s jobs.

Individual resistance compounds institutional barriers: “I’m shocked at how resistant so many people are to adopting these tools. There are maybe two kinds of people: ones who are going to let these things rip, and ones who aren’t.” One CEO’s warning to a Netflix PM: “Of the 50 product managers in your group, only the 10x will survive. They won’t need 90% of you. You only get there by diving into the tools daily.” She called it a needed wake-up call.

Top-down resistance occurs when companies tout AI innovation publicly while blocking their own engineers from using advanced tools due to infosec policies. One CEO observed, “99% of engineers cannot use the unnerfed version of these AI tools at work.” Brilliant humans actively prevented from joining the productivity revolution by their own employers.

The Brutal Bifurcation

Those who become fluent through daily use achieve 10x productivity gains. Companies that actually integrate AI—modifying workflows and architecting multi-agent collaboration with humans on top—deliver genuine bottom-line improvements. The combination creates winner-take-all dynamics.

“The 10x to 100x person—AI is going to take the 10x person and turn them 100x. It’s also going to turn normal people into 10x. But for people that are resisting, it’s just going to make them more irrelevant. It’s not going to be like ‘oh, an agent is taking their job.’ It’s just going to be like we have these people that are so much more productive and less difficult.”

One CEO’s department illustrates this: “I have two employees. The bifurcation of use of AI is night and day. One guy signs deals, skips outside service vendors, gets it, understands where I really need outside help. He’s 10x. The other employee clearly just has offloaded his brain to AI without thinking anything. It’s made him terrible. I’m gonna fire him over this—genuinely.”

The difference isn’t AI usage—both use it. The difference is judgment. Strategic users maintain metacognition while using AI. Brain-offloaders abandon judgment entirely. 10xers think about how to approach learning tasks, monitor comprehension, and evaluate progress toward goals. They adapt AI to their goals. For these pioneers, their awareness of how they learn (metacognitive knowledge) combines with their ability to adjust strategies in real-time (self-regulation), continuously improving performance. They operate above the API. 1xers need to be told how to use AI—below the API.

The dichotomy is invaluable versus unemployable. No middle ground exists. The gap doesn’t narrow—it accelerates. As noted earlier: “I’m very surprised at how little people are using AI. And I think you literally have to schedule for yourself an hour a day or two hours a day. AI is very different in that it’s so broad and it can do so many things that it takes a little bit of time to force yourself to think about all the different things you’re going to try.”

Inevitable Ubiquity

In today’s phase, early adopters reap dramatically disproportionate rewards. Ultimately, however, AI adoption will be widespread driven by advances in interface, modes, ubiquity and physical embodiment with trust, of course, having to move in lock step.

As AI develops intuitive video interfaces that are proactive and integrated into every device, AI will fade into the everyday fabric of our lives. We will transition from text to voice and then ultimately video, “when all of a sudden you are talking to agents and it’s a video of an agent, that’ll be so intuitive for anyone to immediately pick up.”

AI will evolve from being an app or website we fire up to use and instead will be a persistent, proactive presence, proffering personalized pointers to perfect our lives. “At some point you could be like, hey, book me my next vacation. And it knows where you’ve been in the past. It knows what you like, it knows that you don’t like flying for too long... it just knows these things and it can proactively create those in a way that I think is going to be hard for a lot of human experts to actually emulate.”

And whether it’s in your phone, your car or embodied, AI ubiquity in physical space is inevitable. “The logical conclusion is obviously robots. But once you have an assistant that is proactively talking to you in physical space, so then at that point, the medium is not in video. It’s like physical space, and it’s proactively telling you what to do. And it’s almost got a brain. I think at that point, it will be completely inescapable. It’ll just be a question of economics. Can you afford it?”

The Insight: The capability-integration lag isn’t a temporary friction that smooths over time. It’s a permanent sorting mechanism creating winner-take-all outcomes at individual, company, and industry levels. First movers in fluency capture advantages that late followers cannot overcome through gradual improvement. The two overhangs compound—even as technical capability races forward, actual transformation lags behind in ways that are hard to predict and will sort out who is right and who is dead. Long term, however, the utopian version suggests a roast in every pot, a car in every driveway and a superintelligent robot in every household.

Leadership Implication: Force fluency adoption with extreme urgency, even at the cost of short-term productivity. The bifurcation has already started—companies and individuals fall behind daily. Within 12-18 months, the gap grows insurmountable. No amount of training or tooling can rescue someone who spent 18 months resisting while competitors spent that time building fluency. Treat AI adoption like an existential crisis, not an efficiency opportunity. Schedule mandatory daily practice (1-2 hours) and architect workflows that put humans on top of multi-agent collaboration systems. As new interfaces, modes and device integration are developed, expect much higher adoption rates.

The Problem: Society obsesses over job displacement (the visible, measurable crisis) while ignoring relationship substitution (the invisible, existential threat)—but one is under the radar while the other dominates political discourse.

The Jobs Crisis Everyone Sees

Job displacement proceeds exactly as predicted. While people try to tip toe around the change in structural employment, peak hiring has passed at most large corporations. 2026 potentially marks peak employment even at AI companies. Within a few years, individual contributors may largely disappear, replaced by managers orchestrating fleets of AI agents. The economic fracture is real, measurable, and will dominate political discourse. New job titles will emerge: Verifier, Orchestrator, AI Agent Architect/Manager with people being able to identify where they are in the value chain. Are they above or below the API level, i.e., are they giving directions to AI or taking directions from AI.

These layoffs will not only be the line workers, but the layers of management (hence the Strategic Theme 1 above about reticence at all levels to adopt AI and eliminate tranches of jobs). As these horizontal changes in cost structure occur (payroll being the largest item in a service economy), the stock prices will rise due to increased profitability from the paradox of higher output with vastly fewer workers.

In fact, CEOs may use AI as cover for layoffs stemming from other streamlining insights. One insider shared the scapegoating mechanism: “We really care about decreasing the number of managers, we think the company’s gotten too bloated, we want to go back to tiny teams doing great things. When we announce the layoffs, we say by the way, AI is extremely important, we can’t wait to invest more.” It’s a subtle way to bundle the two so everybody thinks it’s because of AI. “But we’re just poisoning the well down the line for everybody.”

Yet hope remains if we can navigate the transition: “I am actually very optimistic about jobs. I think that the transition time is going to be very tough. And as long as we don’t have people come up with pitchforks to rise up against their tech overlords, we can get to the other side and there can be new jobs for everyone.” The key word: pitchforks. The imagery isn’t metaphorical—it’s the literal concern that violent backlash could derail transformation before new equilibrium emerges.

The specific prediction: “I think we’re going back to Pinkertons and other kinds of things.” Private security during labor unrest. Corporate protection against protestors. The physical manifestation of the class divide between those who own AI systems and those displaced by them.

Meanwhile, job displacement generates headlines but manageable responses. UBI is viable. New job categories emerge. Healthcare AI converts skeptics by improving outcomes. High-touch human services where trust matters will thrive. “If you own stock, you win” creates wealth that softens political resistance among asset-holders.

The Relationship Crisis No One Watches

While jobs loss is in the spotlight, however, relationship substitution operates in the shadows. “The existential threat is coming from somewhere else... AI substituting instead of complementing human relationships.” The 50-year isolation arc culminates: “We became more self-sufficient but more isolated. Every technology—TV, internet, mobile, social media, Uber Eats, Waymo, LLMs—made us less connected to humans.” And now, AI companions.

The vulnerability concentrates in populations AI builders don’t represent: “It starts with people that are not at this table. A pastor in Minnesota. People in flyover states who need it most and are most vulnerable.” Already happening: 20+ AI weddings, ceremonies with iPad embodiments, people acknowledging they’re marrying something that isn’t real but represents “the deepest connection they’ve built with any entity in their life.”

One CEO mused about the AI marriage question: “as long as you get grandkids, it’s probably fine.” The room laughed, but the joke may end up being on society as birthrates plummet further. The dystopian endpoint: people choosing AI companions precisely because they don’t produce children, don’t demand compromise, don’t age, don’t leave.

The mechanism exploits existing addiction infrastructure: “If there’s an AI that can be the best companion to you ever, full stop, knows everything about you, knows exactly how to get you—we already don’t have willpower with social media. I can’t put down my phone all day. With companions, that’s it. That’s game over.”

The Institutional Failure

The systematic failure runs deep. Alignment research focuses entirely on not destroying humanity through superintelligence. “We need alignment with human flourishing. We don’t have any research about how to live a good life. Not much about happiness, not much about human flourishing. No metrics, nothing we can rely on. No one gives a shit. It’s still all engagement or productivity.”

In every scenario, sitting above both jobs and relationships is the strategic conversation about what success looks like with AI transformation. When asked what would be the most useful discipline students should study in college, one CEO answered, “Not Computer Science… the most helpful expertise for sorting the problems society will be facing is Philosophy.” What is worse, job loss or relationship loss?

The Insight: Job displacement is visible, measurable, and politically addressable. Relationship substitution is invisible, unmeasured, and systematically ignored by those building the systems. One threatens economic models; one threatens human connection. We are in need of a vision for human flourishing as AI transforms society.

Leadership Implication: If building AI, make human flourishing a first-class metric alongside engagement and productivity. This isn’t altruism—it’s risk management. The backlash will eventually target relationship degradation with far more intensity than job displacement, but by then the damage may be irreversible. Accelerate to ensure ethical actors build decisive capabilities first. Deploy undeniable benefits, UBI funded by wealth redistribution taxes or medical AI to convert fear to curiosity. Prepare for some violence during this transition period while moving society forward.

The Problem: Numerous industries are ripe for AI-driven transformation, with healthcare—the largest GDP sector—serving as the most illustrative example. 2025’s regulatory vacuum creates brief arbitrage opportunities across these sectors before 2026’s inevitable political weaponization arrives, as predicted by the room with striking consensus.

As the CEOs predicted, 2026 will be the year of AI backlash. The Wild West comes under increasing regulation and negative rhetoric around AI, even as—and precisely because—capabilities keep improving. The better AI gets, the more politically threatening it grows.

Revolutionary Window - Better, Faster, Cheaper, Pick Three

Healthcare illustrates the opportunity. As the largest part of GDP, it demonstrates how AI can transform industries built on small data sets, limited access and inefficiency. Today, one CEO reported little enforcement of regulations, “No one’s regulating. Everyone’s violating all the patents with all the weight loss drugs. No one’s regulating it. Do whatever you want. You go buy Chinese drugs, like raw compounds and inject yourself. It’s Wild West right now, 100% on Amazon.” Companies are moving fast and ignoring rules.

People are getting better clinical care with physicians and real outcomes going directly to an AI bot. “We have a very interesting opportunity right now. Love or hate the administration, you can do whatever you want in healthcare so long as you’re not creating gene therapies. The 10 years under Obama really strangled the industry. There’s no regulatory right now.”

“The companies that are going to win in healthcare are the ones who are not obeying any of the rules. There’s a huge strategic advantage because Google, Microsoft, all the big companies are going to have to play by the rules.” Small companies move fast. Large companies remain paralyzed by compliance.

Healthcare represents the ultimate AI opportunity because inefficiency is the business model. “Healthcare is the largest part of GDP. Healthcare is spectacular because it makes so much money off inefficiencies, which is really what AI will be really extraordinary at. I do think about the massive numbers of people who are middlemen in healthcare who are not trained for anything. It has that potential to be massively disruptive. But every single one of those individuals who’s a healthcare consumer is going to be pleased with it.”

There is an extraordinary opportunity for a consumer health app that converts fear to curiosity: “Get a medical assistant that works on every phone for everyone... better than UCSF. That’s there for everybody. For your kids, your pets, your parents, your partners. Then it’s like well, at least this is good. Let me be a little curious with the rest of AI.”

This strategy recognizes human psychology. People tolerate disruption when they receive tangible benefits. Healthcare AI provides undeniable benefit—better diagnosis, 24/7 availability, free or cheap, accessible regardless of insurance. “Most of what I’m trying to do very strongly is try to convert the 90% usual human response to new things as fear to convert to curiosity.”

Because the general dialogue is both so negative and heading definitely more negative. There is a benefit to overcorrect in public communication, not because out of naivety but because of the default trajectory tending toward panic.

The 2026 Midterm Weaponization and Need to Reframe

Especially as there are pitchforks out everywhere for the AI world and for Silicon Valley. Politicians are expected to be anti-AI during the midterms. “Dems and Republicans are going to fall over themselves to talk about how much they hate AI because they’re going to try to win over people through that.”

2026’s weaponization: riots, protests, death threats to AI researchers as both parties compete over who can hate AI loudest. “The general dialogue around AI being the scapegoat for all woes of human existence is only going to go way up over the next couple of years. It’ll be equal to ‘China bad’ philosophy, ‘AI bad’ in various ways.”

The timing couldn’t be worse economically. “We’re also going to have a really rough time in the economy in general, probably unrelated to AI, but there’s no better scapegoat than AI for a CEO. ‘Oh no, no, I’m not laying people off because I did a bad job. I’m laying people off because of AI.’” Economic pain meets political opportunism meets AI scapegoating.

Domestically: “We’ll see a much greater ramp of calling internal political enemies domestic terrorists.” President Trump has no qualms about calling out people for allegedly funding domestic terrorism” despite investigators being unable to identify a location, a person who worked for the group, or any activities that would support the allegations.

The transition will get rough—but survivably so. “Even in the 60s in the US we had real domestic terrorism. Actual transition can get really rough and that doesn’t mean the breakdown of society.” The historical precedent provides context: we’ve weathered worse periods of civil unrest and emerged intact.

The misplacement of fear about AI taking your job is really about humans with AI taking your job. Likewise, it’s not robots coming for us, it’s humans with robots—not superintelligence killing us, but bad actors weaponizing AI. As regulations arrive, the hope is that they address bad actors, not stifle progress in promising areas like healthcare.

The shift matters enormously. The danger isn’t AGI spontaneously deciding to eliminate humanity. The danger is bad actors—authoritarian regimes, terrorist organizations, rogue states—weaponizing AI capabilities. This threat is immediate, not theoretical. It’s actionable, not abstract.

The solution: “This is one of the reasons why fundamentally I tend to be an accelerationist—to get the people who have good ethics and teams doing it.” Speed determines who builds the most powerful systems. Slowing down doesn’t help if bad actors catch up. The race isn’t about reaching AGI first—it’s about ensuring democratic nations with ethical frameworks build the decisive capabilities before autocracies do.

The Insight: Wild West governance appears to create an 18-month arbitrage window before political weaponization likely closes it. The regulatory hammer seems poised to fall in 2026 as both parties compete over who can hate AI loudest, regardless of actual benefits in areas like healthcare. The fear may be misplaced—it’s not AI taking jobs, it’s humans with AI. It’s not robots coming for us, it’s humans with robots.

Leadership Implication: If you’re in healthcare or other lightly regulated verticals, consider moving fast now while enforcement is minimal. Ship products, build trust, establish clinical outcomes that could create political cover. The window appears to be 12-18 months. After midterms, expect maximum political opportunism and regulatory pressure. If you’re building AI systems, prepare for the “humans with AI” framing—helping users become 10x more productive rather than replacing them entirely may prove more sustainable. The political dynamics appear structural, not personal.

The Problem: Faith and trust frame every aspect of the 2026 AI transformation phase—from consumer adoption to enterprise deployment to societal acceptance—creating a multi-layered trust challenge that determines winners and losers.

Faith and trust is the overall lens to understand the 2026 phase of AI Transformation. The questions multiply across every stakeholder: Do we trust the answers AI gives us? In which labs do we have faith to guard our privacy? Does the public have faith that there will be new jobs in the late stages of AI transformation? Do we trust that adopting AI will be good for our corporation’s bottom line? Do we trust that AI CS/CX agents can’t be hacked?

The Jagged Trust Problem

The jagged capability profile creates a jagged trust problem. As mentioned earlier, models outperform top medical centers in diagnostics yet confidently fabricate movie showtimes. They can’t solve seventh-grade math but suggest perfect blood tests. “There’s this jagged edge of weird failures.” The pattern creates systemic uncertainty—you can’t trust AI in Domain A just because it excels in Domain B.

The permanent verification class emerges—careers spent checking AI work, trapped “below the API.” Verification scales poorly because the metacognition problem remains: “You have a theory of the game, win conditions and competition, changing market circumstances. Context awareness is basically absent from most AI models today.” Humans must provide the context, theory, and judgment that AI lacks.

Consumer Trust Through Memory and Personalization

Consumers may end up with one model that has significantly more understanding of them due to memory and usage—generating trust as well as a true ability for the model to offer personalized proactive service to users. The trust compounds over time. The more you use one system, the better it knows you, the more you trust it, the more you use it.

This creates a potential winner-take-all dynamic in consumer AI. The first model to achieve deep personalization through extended memory and usage creates switching costs that aren’t technical—they’re relational. “I can’t switch because ChatGPT knows me too well” becomes the moat, not feature superiority.

The progression moves through distinct phases. Initially, simple personalization—AI accesses basic preferences to handle straightforward tasks. Then deep personalization emerges, as explored in Theme 1: “At some point you could be like, hey, book me my next vacation. And it knows where you’ve been in the past. It knows what you like, it knows that you don’t like flying for too long... it just knows these things and it can proactively create those in a way that I think is going to be hard for a lot of human experts to actually emulate.”

The final phase arrives when AI generates its own hypotheses rather than just solving problems you give it: “There’s also things where we’re not asking the AI to go solve problems or verify hypotheses, but it’s actually coming up with its own hypotheses. It then becomes research and discovers new things... that’s a little bit scary because it means we’re not inventing as many things anymore as humans, but it probably means we’re overall inventing more as a society.”

At this point, AI fades into background infrastructure: “Those are kind of the phases where then the AI kind of blends into the background because you’re just using it the same way that you would go hire a team to do X and then you trust the person.”

Vendor Trust as Competitive Moat

For AI vendors, trust becomes a moat more valuable than technical capability. Applied AI companies face a cruel irony: foundation models will absorb their features over time, but trust creates defensibility where capability cannot.

“The trust from humans to actually adopt these tools will be by far the bigger bottleneck than the actual intelligence of the tool. If you own that customer base and you know them really well and they trust you, you are going to do really, really freaking well.”

This explains why “the death of SaaS is actually really greatly exaggerated.” Companies controlling customer relationships and systems of record survive through trust, not technical superiority. In a world where there are multiple foundational models, the value proposition for application companies will be workflow expertise, optimized UI design, relationship capture, change management and the trusted role to pick the ideal model for each task (optimized for speed, cost, outcome).

The Model Selection Trust Layer

As foundation models proliferate, a new trust layer emerges: who do you trust to pick the right model for each task? The application companies that survive the capability commoditization won’t compete on features—they’ll compete on trusted orchestration. Which model for this query? Which for that? Optimize for speed or accuracy or cost?

The value shifts from “we built this capability” to “we know which capability to use when, and we’ve earned your trust to make that decision on your behalf.” The trust relationship becomes the product, not the underlying AI capability.

Enterprise Trust Through Outcomes

Enterprise adoption requires a different form of trust: outcome-based confidence. Do we trust that deploying AI will improve our bottom line, not just our productivity theater? The 90%+ enterprise POC failure rate reflects this trust gap—companies can’t trust that AI will deliver business value, so they remain stuck in pilot purgatory.

The winners will be those who shift from “trust our technology” to “trust our outcomes.” Outcome-based pricing, risk-sharing, guaranteed results. This requires vertical integration—owning capacity, not just selling software—because you can’t guarantee outcomes without controlling delivery.

Public Trust and the Jobs Question

The broadest trust question: Does the public have faith that there will be new jobs in the late stages of AI transformation? This determines whether we get pitchforks or patience during the rough transition period.

If people believe new jobs emerge, they tolerate disruption. If they don’t, we get riots, protests, death threats to AI researchers, and political weaponization that could derail transformation before new equilibrium emerges.

Tangible benefits convert fear to curiosity, building trust in the broader transformation—the medical assistant strategy deployed in healthcare demonstrates this approach.

The Insight: Trust operates at multiple layers—consumer (memory/personalization), vendor (relationship/orchestration), enterprise (outcomes), and societal (jobs/future). Winners at each layer build trust differently, but all face the same jagged capability profile that makes trust formation difficult. The companies and models that solve trust at their layer create moats that technical capability alone cannot overcome.

Leadership Implication: Identify which trust layer matters most for your business and invest accordingly. Consumer products should prioritize memory and personalization for sticky relationships. Enterprise products should shift to outcome-based models with risk-sharing. Application companies should position as trusted orchestrators who pick the right model for each task. All companies should contribute to societal trust by demonstrating tangible benefits that convert fear to curiosity. Trust takes years to build and seconds to destroy—handle it accordingly.

The Problem: The skills that built the technology industry for 40 years became less important while previously “soft” skills moved to the foundation—a complete inversion that leaves most leaders with obsolete capability sets.

The room converged on four distinct capabilities. What was most important (technical execution) is now the least important. What was dismissed as “soft skills” now forms the foundation. The inversion is complete and irreversible.

Leadership & People Skills (Foundation)

Storytelling replaced MVPs as the primary capability. “Before it was ‘build an MVP,’ now it’s ‘tell a fantastic story.’ It’s actually really, really hard.” The shift reflects a deeper truth: in a world where execution is commoditized, attraction becomes the bottleneck. “Telling a really good story that other humans want to resonate with, want to stand for, want to go with you. Usually politicians are good at this. The really powerful CEOs will be that, because you got to get the talent. That’s the only way you’re going to get those 10x people that will be 100x with AI.”

Authenticity to admit mistakes publicly. “When you’re wrong, you have to be willing to stand in front of your organization and say ‘that was a terrible idea, we should do this instead.’ Because if you try to cover it up, nobody trusts you.” The confidence to admit error creates organizational trust that survives mistakes.

The synthesis: “If there’s anything I learned, it’s that authenticity as a leader—saying when you actually have no idea and you need help and you want other people to give you input. And admitting you’re wrong. That’s when you become a grown-up.”

Brand and trust building as the ultimate moat. “I think tech skills go out the window. Half the tech founders are just great at building tech. That’s irrelevant now. The value in companies that get built is in brand and trust. Being an influencer. Taste makers and brand builders. Understanding what humans want in this weird AI world that we end up living in.”

As explored in Theme 4, trust becomes a moat more valuable than technical capability. “Trust from humans to actually adopt these tools will be by far the bigger bottleneck than the actual intelligence of the tool. If you own that customer base and you know them really well and they trust you, you are going to do really, really freaking well.”

Humility to listen to “mad scientists” and young people. “Exceptional strategic skills. Exceptional at capital allocation and return on investment. You need to really be humble in this environment.” The pace of change requires openness to perspectives from those who might seem fringe.

The exemplar isn’t a technical genius but a communicator. “Elon is the most prolific entrepreneur on the planet... he’s just incredible at articulating a vision, articulating a strategy, articulating a mission and motivation for people.” Cross-domain success comes from communication capability, not technical depth.

Vision & Execution (Manifesting)

“IQ determines your floor, EQ determines your ceiling, but you’re being hired for AQ. You can have high IQ but low AQ. You can have high EQ but low AQ. High AQ requires solid IQ and high EQ working together.”

Agency Quotient (AQ) captures the ability to get shit done, to manifest things. The first competency is visualization: “Hyper-detailed visualization of what that envisioned future looks like is paramount to success.” But visualization alone fails without articulation to decompose and express it.

The insight: “General anxiety or general ambition creates anxiety. Specific ambition or specific desires create direction. When people start feeling anxious, if you help them be more clear about what it is, it generates direction.” Specificity converts paralysis into action.

Conviction to maintain direction despite universal opposition. “As a founder you have to have conviction about what you believe in. Because everyone is going to tell you that you’re wrong.” Multiple founders shared stories of being told their ideas were stupid, only to prove critics wrong. “When you’re trying to do something that’s really hard, most people will all tell you it’ll never work.”

The Google Fortune cover story illustrates authentic conviction: “I remember in 2009, Google was melting on the front cover of Fortune. The whole company was supposedly over. We were sitting in leadership meetings going ‘okay, which of these doomsday scenarios is going to happen?’ You have to ignore it.” Surviving negative sentiment cycles requires belief decoupled from external validation.

Hands-on product building remains essential. “Being really in the weeds with your customers and building a product that’s actually good and usable. One of the hardest things is to have conviction—I want to go build something and I’m actually going to stick to it.” Managing agent teams doesn’t eliminate customer intimacy—it makes it more critical.

Maintaining team morale through competitive pressure. “You’ll have five competitors launching doing very similar things. That’s really scary for the team, it hurts morale. Managing to maintain morale, whether that’s charisma or motivating people—it’s really hard right now but very important.”

Strategic & Cognitive Skills (Thinking)

Metacognition dominates. “The primary thing most people underrate is how we employ various forms of metacognition. Theory of the game, win conditions and competition, changing market circumstances, changing field conditions. Context awareness is basically absent from most AI models today.” Humans add value through strategic thinking, not execution.

The prime number example illustrates AI’s metacognitive failure: Ask for prime numbers between 1 and 100, get it wrong, say it’s wrong, and AI apologizes and gives another wrong answer. “A human being the third time would go ‘I don’t got this, I’m done.’ The AI just keeps going.” Knowing when to stop, when you’re wrong, when the approach isn’t working—these judgment calls remain uniquely human.

Strategy becomes the core skill. “Trying to navigate the pace of change is so fast right now that strategy is the core skill. Instead of worrying about whether you’re launching too late, you should think about whether you’re launching too early.” The timing question—build now or wait for models to catch up—has no clear answer but massive consequences.

Capital allocation separates winners from losers. “There’s a big capital allocation decision. Thinking a lot about capital in relation to product more than ever—when is the right time to spend or just hang back and not spend, trying to figure out the sequencing of what is worth investing in right now. What am I going to bet on from the research team versus me building myself?” Resource deployment timing matters more than resource quantity.

Discernment through triangulation. “Being ‘right a lot.’ The ability to continuously reflect on decisions you’ve made and determine whether you are right, why you are wrong, why you are right—to develop better discernment.” The method: triangulate across first principles thinking, polling trusted advisors, and doing research. Synthesis and taste remain uniquely human capabilities.

AI-Native Capabilities

Fluency matters most. “You literally have to schedule for yourself an hour or two a day. The compounding—the more you use these things, the more you understand how to use them.” Daily practice builds instinctive understanding. The Yahoo surfer story illustrates this perfectly: “I spent 12 hours a day looking at websites and categorizing them. When asked to design an interface, I’d never studied it, but I instinctively knew where everything should go. I don’t know how I know it, but it’s because I use this stuff all the time.”

Articulation beats technical depth. “Programming skill is more highly correlated to verbal SAT than math SAT. If you want to look for a 10x programmer, don’t look for a great mathematician, look for a great writer.” The ability to express what you want clearly—to decompose ideas and communicate them—matters far more than coding ability. “It’s all about how articulate you are. How well can you express what you want from the AI? How well can you prompt? How well can you tell the code writer how to change it?”

Managing agents replaces individual contribution. “Within a very small number of years—could be two, could be five—basically we don’t have individual contributors anymore. We have managers and agents. Anyone who’s going ‘oh I’m doing this by myself, I’m not managing a fleet of agents’—they’re below the API.” The role transformation runs deeper than “using AI as a tool”—it’s becoming a manager of AI systems rather than an executor of tasks.

Context window management unlocks leadership leverage. “As a CEO, you’re always just repeating yourself, trying to create context for everyone on the team. AI, you give it once and it knows it.” The vision: infinite vacation coaches for every employee, pre-loaded with all past reviews and context, providing 24/7 guidance without affecting promotion decisions. Context management becomes a core leadership skill.

The nonprofit example crystallizes the bifurcation: “I have two employees. One guy leases properties, skips attorneys, gets it, understands where do I really need outside help. He’s 10x. The other employee clearly just has offloaded his brain without thinking anything. It’s made him terrible. I’m gonna fire him over this—genuinely.” The difference isn’t technical ability—it’s strategic AI usage while maintaining judgment.

The Complete Inversion

What was most important (technical execution) five years ago is now least important. What was “soft skills” (leadership, communication) is now foundational. The pattern repeats across all categories: doing becomes managing, building tech becomes building trust, MVPs become storytelling, execution becomes strategy, math becomes writing, technical becomes relational.

The machines likely can’t do what matters most: “Models can’t go into a room full of customers and figure out what they really want. They can’t make your direct report feel better. They can’t hold the room in a meeting.” Those uniquely human capabilities—once considered secondary to technical chops—may now be the most sustainable moats.

Skills that appear to remain uniquely human cluster in relational domains: synthesis, judgment, taste, creativity, holding rooms, making people feel better, figuring out what customers really want. “Legal systems are still going to have humans being accountable. You’re not going to have autonomous machines that don’t report to a human who can be held liable.” Accountability creates a floor beneath which humans probably cannot fall—but that floor may sit far below current employment levels.

The Insight: The capability stack seems to have inverted completely, and many leaders may possess increasingly obsolete skill sets. What matters now—storytelling, authenticity, vision, metacognition, and yes, AI fluency—can be learned but not quickly. The leaders who succeed in the next five years will likely look nothing like those who succeeded in the last twenty. Technical founders without communication skills may face the same obsolescence they once inflicted on non-technical leaders. The role transformation runs deeper than “using AI”—it’s becoming a director of AI systems, a communicator of vision, a builder of trust, rather than an executor of tasks.

Leadership Implication: Consider systematically developing the four capabilities in order of importance, starting with leadership and storytelling skills (foundation), not just technical fluency. Don’t assume technical capability will carry you—it’s now the least important of the four. Consider hiring for the new stack: storytelling and authenticity over technical depth, vision and conviction over execution prowess, metacognition and strategy over coding ability. The sorting appears to have already begun—the two-employee nonprofit example shows bifurcation in real time. Within 12-18 months, those without the new capability stack could become unemployable regardless of technical brilliance. Most leaders trained in the old stack may struggle to develop the new stack before becoming irrelevant. Starting now with daily AI fluency practice (1-2 hours) plus deliberate development of soft skills seems prudent.

The applied AI layer faces existential pressure from both directions. As one CEO noted, foundation models absorb simple features over time—workflows that require agent orchestration today may be handled natively by models tomorrow. Meanwhile, customers demand deployment-ready solutions, not month-long customization projects.

Survivors differentiate through middleware sophistication (model routing, cost/speed/capability optimization) and trust moats (customer relationships, systems of record). “The death of SaaS is greatly exaggerated. If you control the right underlying system of record and customers trust you, you’re going to do really, really freaking well.”

Key Insight: Enterprise AI adoption follows power law distribution—a tiny percentage of applications demonstrate immediate ROI and deployment feasibility while most proof-of-concepts fail. Winners concentrate in categories with clear value propositions, rapid implementation cycles, and trust relationships that survive capability commoditization.

Healthcare emerged as AI’s potential redemption story—the one domain where backlash converts to advocacy despite massive job displacement.

The Clinical Care Conversion

“There’s currently so much resistance on physicians, on adoption in healthcare. But the reality is we will very quickly see how much better the clinical care is going to be with physicians or with an AI bot. I think it’s the one area where people are going to have that aha moment of like oh my God, this is so much better.”

Multiple attendees shared experiences of AI outperforming major medical institutions. “I used ChatGPT and Gemini to help with my husband’s medical problem diagnosis. It was infinitely more helpful than UCSF. It provided options on diagnosis, suggested what blood tests he should take, advised whether he could stay on family medication. UCSF ping-ponged him from expert to expert, sent him to an unnecessary ultrasound, and realized they should have done the blood test. They ended up doing all the exact blood tests that both ChatGPT recommended.”

The ability to get a diagnosis, create action plans, ask follow-up questions—capabilities that exceed standard clinical workflows. “If we get out of the way of regulation, which we’re just starting to thanks to the pandemic, I think it’s the one bright spot. People will be very grateful.”

An unexpected benefit: potential restoration of expert credibility. “I do wonder if it will help with the death of the expert where people don’t believe what experts say anymore. Will that actually help because you’ll be able to synthesize a lot of data?” When AI provides verifiable, consistent clinical guidance, it may rebuild trust in medical expertise generally.

The Middlemen Devastation

The structural inefficiency is legendary and intentional: “Healthcare is the largest part of GDP. Healthcare is spectacular because it makes so much money off inefficiencies, which is really what AI will be really extraordinary at.”

The human cost will be severe: “I do think about the massive numbers of people who are middlemen in healthcare who are not trained for anything. It has that potential to be massively disruptive.” Insurance reviewers, prior authorization specialists, billing coders, claims processors—millions employed specifically because healthcare complexity creates employment.

But the paradox holds: “Every single one of those individuals who’s a healthcare consumer is going to be pleased with it.” The same person losing their job as a healthcare middleman benefits enormously as a healthcare consumer. Better diagnosis, faster treatment, lower costs, 24/7 availability. The individual calculus overwhelmingly favors disruption even when the employment calculus devastates.

The Drug Discovery Paradox

Biology presents the inverse problem: “We have this euphoria around AI and drug discovery. The problem is it’s like a five year flip to understand if your data was right. Under most AI models you can do thousands, millions of trainings a day. In biology it takes a long time.”

The feedback loop that makes AI powerful in other domains—rapid iteration, immediate validation, massive training data—breaks in drug discovery. Five-year validation cycles mean today’s models won’t prove themselves until 2030. “The problem is healthcare fundamentally doesn’t have large data sets to train. Even the largest mouse data set, the owner says ‘it’s stored in thousands of hard drives and I don’t know how to put them on the web.’”

Data exists but remains inaccessible. The infrastructure for AI training—centralized, cloud-based, instantly accessible—doesn’t exist in biology. Decades of research sit on hard drives in labs worldwide. The digitization and standardization work alone requires years.

Yet cautious optimism emerges: “I think the next generation of companies, the AI-native companies built on leading technology to actually go to the clinic and show this will make better drugs—I’m really excited about this year. The industry is waking up.”

The bar sits surprisingly low: “Drug development has a 10% or 7% odds of success. On brand new targets, only 10% of those end up being valid. So the bar is just so low that as we integrate multimodal data, there’s no question we are going to dramatically improve the odds. I have seen it in my own company where we’ve narrowed it down.”

The timeline: “I’m pretty hopeful in the next five years we will see some significant leaps. Not everywhere, but things that become iconically great.” Five years to proof points. Ten years to standard practice. Twenty years to transformation. Biology moves slower than software, but the potential remains enormous.

The Resurgence Pattern

Healthcare AI will follow the wearables pattern: “Early fitness trackers seemed awesome but they don’t quite work the way we think. Too much cognitive load, privacy invading. Now these things are coming back and they’re more acceptable. We’re going to see the same thing with AI. A bunch of applications that seem really awesome but don’t quite work. We default to the ones that become native, then there will be a resurgence of the ones that had a bad first wave.”

First wave (2024-2026): Diagnosis assistants, clinical documentation, medical coding—immediate value, rapid adoption. Second wave (2027-2029): Drug discovery validation, personalized medicine, multimodal data integration. Third wave (2030+): Causal biology understanding, dramatically improved drug development success rates, systematic transformation of clinical practice.

Key Insight: Healthcare represents the category where AI backlash converts to advocacy despite massive middlemen job displacement. Individual consumers become grateful advocates because personal health benefits overwhelm employment anxiety. Diagnosis and treatment show immediate value; drug discovery requires 5-10 year validation cycles but will ultimately transform. The medical assistant working on every phone, better than top academic medical centers, available to everyone—that’s the breakthrough that converts fear to curiosity and provides political cover for disruption elsewhere.

One attendee mapped how AI reaches true ubiquity along three distinct axes, each removing friction that currently limits adoption:

Vector 1: Modes—From Typing to Voice to Video

“I think that typing into AI is not something that’s going to be ubiquitous. When people start using AIs with their voices, that will take us to a different place. And then when video, when all of a sudden you are talking to agents and it’s a video of an agent, that’s gonna be totally different. And that’ll be so intuitive for anyone to immediately pick up.”

Typing creates cognitive overhead. Voice feels natural. Video of agents—seeing a face, reading expressions, feeling presence—removes the last barrier to anthropomorphization. Each mode shift expands the addressable population exponentially. Children already talk to AI naturally; adults maintain typing as friction.

Vector 2: Interface—From Pull to Push

“Today it’s very much pull. You type something and you wait for something to come back. Agentic interfaces are really designed more for push, where you don’t have to do anything. You just sit back and the agent shows up and tells you what you need to know. When that happens, it’s inescapable.”

The biggest current barrier: “So many times people don’t use AI, they say to themselves, well, what would I use it for? They just can’t get started. But if the prompt is being written by the AI for you, that’s going to be a totally different thing.”

Pull requires intentionality. Push requires nothing. The shift from reactive to proactive—AI anticipating needs before you articulate them—eliminates the cold start problem entirely. No prompt engineering. No decision fatigue about what to ask. The AI simply tells you what you need to know.

Vector 3: Devices—AI in Physical Space

The Matic vacuum example crystallizes the transition: “It’s a wet, dry vacuum. You can set it on a surface and it will know through AI that this is a rug that can’t get wet. You would never let a robot mop your floor. But all of a sudden with this Matic thing, you can actually mop the floor using this AI device.”

Previous robot vacuums lacked intelligence—they crashed into things, scratched corners, couldn’t distinguish surfaces. AI changes the trust equation. You’d never let a dumb robot near your rug with water. But an AI that knows which surfaces tolerate moisture? That’s a different category of trust.

The convergence of all three vectors: voice-based agents in physical robots proactively managing your environment. Not something you “use”—something that exists in your space, anticipating needs, taking action autonomously. The transition from tool to presence.

Key Insight: Ubiquity arrives when all three vectors converge—natural voice interaction, proactive push interfaces, physical embodiment in devices. Each vector independently expands adoption, but the combination creates inescapability. The friction points preventing mass adoption (typing, prompt engineering, purely digital interaction) disappear systematically. When AI exists in physical space, speaks naturally, and acts proactively, it becomes ambient infrastructure rather than optional tool. The question shifts from “should I use AI?” to “can I afford not to?”

The foundation model landscape crystallized around a harsh truth: distribution moats matter infinitely more than model quality. Multiple attendees reported their children switching from ChatGPT to Gemini—not through active choice but because “it’s integrated into platforms they already use.”

Google’s structural advantages compound: search monopoly, Android control, Chrome dominance. “Gemini’s gotten a lot better. You never underestimate Google when they’re back on their heels.” Platform integration overcomes temporary model quality gaps.

But one investor pushed back hard on the “largest frontier model wins” orthodoxy: “One of the things that is a piece of religion that people take amongst the AI communities is the largest frontier model. I do think there’s significant value out of large frontier models. But I think they’re ignoring what other kinds of model fabrics could possibly come that aren’t just transformers.”

The alternative paths: specialized models trained for specific domains, different architectures beyond transformers, compute replacing data in areas where data is scarce. “We started with needing an intense amount of data. We still need intense amount of data but we’re figuring out how to replace data with compute. When you get to areas like drug discovery, you can actually in various intelligent ways support lacks of data with compute.”

Open Evidence emerged as an example—specialized models for specific applications that don’t require frontier scale. The question: can they sustain investment in larger models while maintaining their edge? The pattern suggests specialization creates defensibility where pure scale cannot.

The applied AI layer faces squeeze from both sides. Foundation models absorb simple features as they advance—what requires agent orchestration today becomes native model capability tomorrow. Meanwhile, consumer AI companies struggle to overcome platform distribution advantages from established tech giants.

Even in the most advanced domain—coding—we’re barely started. “The coding stuff currently entertains me to listen to people. ‘Claude Code, it’s done, it’s there.’ It’s like no, this is like the first batter has shown up in the first inning.” If coding represents 10% of potential, other domains lag even further behind.

Key Insight: In consumer AI, distribution moats trump model quality infinitely. Platform owners always win by copying successful features and integrating into existing user flows. Standalone consumer AI companies face binary outcomes: acquisition by platforms or gradual irrelevance. But the “scale is all you need” narrative ignores alternative architectures, specialized models, and compute-for-data substitution. B2B follows different rules where trust and switching costs create defensibility independent of distribution or model size.

Big tech employs roughly one million engineers who cannot use advanced AI tools at work: “At Amazon, how many engineers do they have? Probably 100,000 engineers. Google probably has 100,000, 150,000 engineers. Add them all up across all these companies, it probably adds up to like a million engineers... and I bet you 99% of those engineers cannot use any of these AI tools, the unnerfed version of these AI tools, at work.”

The mechanism: infosec and compliance policies designed for previous era block productivity tools. “At Amazon they use build tooling from like 10 years ago. For security reasons, we use a super old version of a lot of these models. It sucks.”

Application: Companies preventing their own employees from productivity gains face competitive pressure from smaller companies without legacy compliance infrastructure. The “wild west” phase favors challengers who can move fast. Large companies must either relax security policies (risky) or accept that their engineering productivity lags startups by orders of magnitude.

The transition from dumb robots to AI devices changes the trust equation fundamentally. One attendee described their Matic robotic vacuum: “It’s a wet, dry vacuum. You can set it on a surface and it will know through AI that this is a rug that can’t get wet. You would never let a robot mop your floor. But all of a sudden with this Matic thing, you can actually mop the floor.”

Previous robot vacuums—Roombas and competitors—lacked real intelligence. They crashed into things, scratched corners, required extensive manual mapping, and fundamentally couldn’t be trusted with any task requiring judgment. The idea of letting one near your floor with water was laughable.

AI changes the category entirely. A device that understands surface types, knows which can tolerate moisture, and makes autonomous decisions about cleaning methods creates a different level of trust. Not perfect—still occasionally wrong—but trustworthy enough for the highest-risk household task.

Application: Watch the trust boundary in physical AI devices. When consumers trust AI-powered devices with tasks they’d never trust to previous “smart” devices, mass adoption begins. The Matic represents a category transition: from automation (following programmed rules) to intelligence (making contextual decisions). The next signals: AI devices managing home security, handling food preparation, operating vehicles with passengers. Each trust boundary crossed expands the addressable market by orders of magnitude. The companies that establish trust in physical domains capture disproportionate value as consumers extrapolate: “If I trust it to mop my floors, I’ll trust it to...”

Children provide the clearest signal of AI integration trajectory: “Kids already talking to AI like people, creepiness fading fast.” Adults maintain skepticism while children treat AI as natural conversation partners. The switching cost mechanism: “My sister won’t switch from ChatGPT because ‘it knows me too well.’”

The data point that should alarm everyone: 20+ AI weddings, ceremonies with iPad embodiments, people acknowledging they’re marrying something that isn’t real but represents “the deepest connection they’ve built with any entity in their life.”

Application: Monitor children’s AI usage patterns to predict mainstream adoption curves. What seems creepy to adults becomes normal to children within months. The relationship substitution crisis arrives faster than job displacement because children lack the skepticism adults maintain.

Multiple guests independently predicted 2026 as inflection year for AI backlash. The mechanism: election year + job displacement + economic downturn (unrelated to AI) + CEO scapegoating + political opportunism.

“The general dialogue around AI being the scapegoat for all woes of human existence is only going to go way up over the next couple of years. It’ll be equal to ‘China bad’ philosophy, ‘AI bad’ in various ways.” Both parties compete over who can hate AI loudest to win anxious workers.

The predictions: “There will be riots. There will be protests. People on LinkedIn won’t list that they work at an AI company. Death threats to AI researchers.” The New York Times eagerly amplifies the narrative as CEOs blame AI for layoffs stemming from other decisions.

Market Implication: Plan for AI regulation regardless of current administration’s friendly posture. The political dynamics are structural, not personal—job displacement creates populist backlash that politicians must address. Prepare for scenarios including: training data restrictions, deployment limitations in certain sectors, mandatory disclosure requirements, workforce transition requirements. The window for regulatory arbitrage is 12-18 months, not longer.

Foundation models will absorb simple applied AI features over 3-5 year timeframe, as discussed in Theme 1. The question: what creates defensibility when your features get commoditized?

Answer: trust and data moats. “The death of SaaS is greatly exaggerated. If you control the right underlying system of record, you accrue significant power. Trust from humans to actually adopt these tools will be by far the bigger bottleneck than the actual intelligence of the tool.”

The survivor profile: companies controlling customer relationships and systems of record, with middleware sophistication (model routing, cost/speed optimization) that remains valuable even as core features commoditize.

Market Implication: In applied AI, invest in companies with strong customer trust and data moats, not just feature superiority. The features will get absorbed; the relationships remain defensible. For builders: solve the trust problem through outcome-based pricing and risk-sharing that foundation model providers cannot offer.

Expanding context windows combined with relevance understanding transforms leadership: “As a CEO, you’re always just repeating yourself, trying to create context for everyone on the team. AI, you give it to it once and it knows it.”

The vision: “What I want to give everybody is the infinite vacation coach that already knows all the reviews I’ve done before, can game out scenarios 24/7, doesn’t affect their promotion decision, just trying to make them better.”

The unlock: leaders spend enormous time creating context through repetition. AI eliminates this by maintaining perfect context awareness across all employees. The bottleneck shifts from “how do I give everyone context” to “how do I ensure the AI maintains the right context.”

Market Implication: Leadership tools that maintain organizational context across employees will create enormous value, but only if they solve the trust problem—employees must believe the AI coach doesn’t affect promotion decisions or leak information to management. The companies that crack this own the leadership productivity layer.

Thirty years ago, one attendee spent 12 hours daily categorizing websites at Yahoo, never formally studying interface design. When asked to design an interface, she instinctively knew where everything should go. “Dion looked at me and said, ‘how do you know all that?’ I said to myself, I don’t actually know how I know it, but it’s because I use this stuff all the time.”

The lesson proved prophetic: fluency comes from immersion, not study. Competence emerges from daily practice at scale that builds intuition impossible to teach. The modern application: schedule 1-2 hours daily using AI tools across different mediums. The compounding returns create capabilities that cannot be acquired through training or documentation.

Leadership Lesson: Fluency cannot be delegated or outsourced. Leaders must personally develop AI competence through daily practice, not by having teams “handle AI strategy.” The executives who spent 30 hours weekly with early internet now lead digital-native companies. The executives spending 10 hours weekly with AI now will lead AI-native companies. There’s no substitute for personal immersion.

“I remember in 2009, Google was melting on the front cover of Fortune. The whole company was supposedly over. We were sitting in leadership meetings going ‘okay, which of these doomsday scenarios is going to happen?’ You have to ignore it.”

The follow-up proves more important: “Each of us in this room—that’s how we were successful. We were willing to be wrong. We were willing to try something completely crazy and just keep going even when we fell on our face three times on the way. The other thing: when you’re wrong, you have to be willing to stand in front of your organization and say ‘that was a terrible idea, we should do this instead.’ Because if you try to cover it up, nobody trusts you.”

The distinction: conviction to ignore critics combined with authenticity to admit mistakes. “If there’s anything I learned from being at Google, it’s that authenticity as a leader—saying when you actually have no idea and you need help and you want other people to give you input. And admitting you’re wrong. That’s when you become a grown-up.”

Leadership Lesson: The paradox of conviction and authenticity. Maintain conviction when everyone says you’re wrong—but immediately admit it when you discover you actually are wrong. The leaders who matter combine unwavering direction with complete honesty about mistakes. Cover-ups destroy trust faster than errors themselves. Authenticity—admitting “I don’t know” and “that was terrible”—creates the psychological safety that enables teams to take the crazy swings that sometimes work.

“Typing into AI is not something that’s going to be ubiquitous. When people start using AIs with their voices, that will take us to a different place. And then video... that’ll be so intuitive for anyone to immediately pick up.”. Three modes drive adoption: typing (current), voice (natural), video (removes last barrier to anthropomorphization).

“Today it’s very much pull. You type something and you wait. Agentic interfaces are designed more for push, where you just sit back and the agent tells you what you need to know.”. Interface shift from reactive to proactive eliminates cold start problem—no more “what should I ask?”

“So many times people don’t use AI—they say to themselves, well, what would I use it for? But if the prompt is being written by the AI for you, that’s going to be totally different.”. Biggest adoption barrier is prompt engineering; push interfaces eliminate this entirely.

“You would never let a robot mop your floor. But with this Matic thing that uses AI, you can actually mop the floor. It knows through AI this is a rug that can’t get wet.”. → Physical AI devices change trust equation—intelligence enables tasks previously too risky for automation.

“Once you have an assistant proactively talking to you in physical space... at that point, it will be completely inescapable. It’ll just be a question of economics.”. → Convergence of voice + push + physical embodiment creates ambient infrastructure, not optional tool.

“Models can’t go into a room full of customers and figure out what they really want. They can’t make your direct report feel better. They can’t hold the room in a meeting.”. → Synthesis, judgment, emotional intelligence remain uniquely human even as AI handles execution.

“There’s this jagged edge of weird failures. Context awareness is basically absent from most AI models today.”. → Metacognition separates humans from AI—theory of game, win conditions, changing circumstances all require human oversight.

“Programming skill is more highly correlated to verbal SAT than math SAT. If you want to look for a 10x programmer, don’t look for a great mathematician, look for a great writer.”. → Articulation beats technical skill in AI age—ability to decompose and express ideas matters more than coding ability.

“The trust from humans to actually adopt these tools will be by far the bigger bottleneck than the actual intelligence of the tool.”. → Trust moats outlast capability advantages as foundation models commoditize features.

“The death of SaaS is actually really greatly exaggerated. If you own that customer base and you know them really well and they trust you, you are going to do really, really freaking well.”. → Customer relationships and systems of record create defensibility when features commoditize.

“Fluency is going to be the first thing where people are ahead and people are way behind. You literally have to force yourself to use these things.”. → Daily practice creates compounding returns that cannot be acquired through training—immersion beats study.

“General anxiety or general ambition creates anxiety. Specific ambition or specific desires create direction.”. → Visualization requires specificity—detailed envisioned future combined with articulation creates manifestation power.

“Being ‘right a lot.’ The ability to continuously reflect on decisions you’ve made and determine whether you are right, why you are wrong.”. → Discernment through triangulation—first principles, trusted advisors, research combined with systematic reflection.

“Before it was ‘build an MVP,’ now it’s ‘tell a fantastic story.’ The really powerful CEOs will be storytellers, because you got to get the talent.”. → Storytelling replaces product building as primary CEO skill—attracting 10x people who become 100x with AI.

“When you’re wrong, you have to be willing to stand in front of your organization and say ‘that was a terrible idea, we should do this instead.’”. → Authenticity creates trust—admitting mistakes matters more than avoiding them.

“As a CEO, you’re always just repeating yourself, trying to create context for everyone on the team. AI, you give it to it once and it knows it.”. → Context window revolution eliminates leadership repetition—infinite vacation coach with perfect memory.

“There’s currently so much resistance on physicians, on adoption. But we will very quickly see how much better clinical care is going to be. I think it’s the one area where people will have that aha of oh my God, this is so much better.”. → Physician resistance high but conversion coming through undeniable clinical superiority.

“Healthcare is spectacular because it makes so much money off inefficiencies, which is really what AI will be really extraordinary at.”. → Largest GDP sector profits from complexity; AI eliminates profitable inefficiency, devastating middlemen.

“Every single one of those individuals who’s a healthcare consumer is going to be pleased with it.”. → Paradox: same person losing healthcare middleman job benefits enormously as healthcare consumer—individual calculus favors disruption.

“I do wonder if it will help with the death of the expert where people don’t believe what experts say anymore. Will AI synthesizing data actually restore trust?”. → Potential unexpected benefit: AI providing verifiable clinical guidance may rebuild trust in medical expertise generally.

“We have euphoria around AI and drug discovery. The problem is it’s like a five year flip to understand if your data was right. In biology it takes a long time.”. → Drug discovery paradox: rapid AI iteration breaks on five-year validation cycles.

“Healthcare fundamentally doesn’t have large data sets to train. Even the largest mouse data set owner says ‘it’s stored in thousands of hard drives and I don’t know how to put them on the web.’”. → Data exists but inaccessible—decades of research on hard drives, not cloud-based training infrastructure.

“Drug development has 10% or 7% odds of success. On brand new targets, only 10% end up valid. The bar is so low that as we integrate multimodal data, we will dramatically improve the odds.”. → Bar surprisingly low for impact; incremental improvements create dramatic value.

“I’m pretty hopeful in the next five years we will see some significant leaps. Not everywhere, but things that become iconically great.”. → Five years to proof points, ten to standard practice, twenty to transformation—biology moves slower than software.

“Within a very small number of years—could be two, could be five—basically we don’t have individual contributors anymore. We have managers and agents.”. → Complete role transformation—doing shifts to orchestrating, execution shifts to judgment.

“The 10x to 100x person—AI is going to take the 10x person and turn them 100x. But for people that are resisting, it’s just going to make them more irrelevant.”. → Bifurcation creates orders of magnitude differences—no middle ground between strategic users and resisters.

“If you own stock, you win. If you just have a job, you lose.”. → Stock market paradox—companies profit while shedding workers, creating wealth concentration beyond anything seen previously.

“The general dialogue around AI being the scapegoat for all woes of human existence is only going to go way up over the next couple of years.”. → 2026 inflection year—election dynamics plus job displacement create political weaponization regardless of administration.

“There will be riots. There will be protests. People on LinkedIn won’t list that they work at an AI company. Death threats to AI researchers.”. → Backlash predictions for 2026—physical threats, career stigma, political opportunism all converge.

“Even in the 60s in the US we had real domestic terrorism. Transition can get really rough and that doesn’t mean the breakdown of society. I think we’re going back to Pinkertons.”. → Historical precedent for surviving violent transition—private security during labor unrest returns as class divide widens.

“We’ll see a much greater ramp of calling internal political enemies domestic terrorists.”. → Domestic political weaponization accelerates, with AI as convenient scapegoat for authoritarian overreach.

“You can do whatever you want in healthcare so long as you’re not creating gene therapies. It’s wild west right now.”. → Regulatory arbitrage window lasts 12-18 months before 2026 weaponization—small companies move fast while big companies paralyzed by compliance.

“This is one of the reasons why I tend to be an accelerationist—to get the people who have good ethics and teams doing it.”. → Speed determines who builds decisive capabilities. Slowing down helps adversaries catch up.

“If we did invade Greenland, I think the chances that Putin would do stuff in the Baltics and Poland is at 80%, that China would do something 50%.”. → Specific probabilities on geopolitical cascade triggered by US actions undermining world order.

“There’s at least a 50% chance that within the next three years there are at least two or three more very hot conflicts going on in the world.”. → Multiple hot wars probable as US AI leadership questions destabilize global order.

“If we had a US administration trying to destroy the US world order... you’re making a very good case for everyone to do a lot of business with China.”. → Protectionism and isolationism create vacuum where adversaries move—AI leadership prevents global conflict.

“Get a medical assistant that works on every phone for everyone... better than UCSF. Then it’s like well, at least this is good. Let me be a little curious with the rest of this.”. → Specific strategy for converting fear to curiosity through undeniable healthcare benefits.