Summary

Dixon’s first filter for any technology investment is whether an exponential force—not a tactical product advantage—is carrying it forward. Moore’s law, open-source composability, and network effects compound until incumbents built for the previous curve are overwhelmed. His operating rule: “These forces are going to overwhelm you for better or worse.”
AI products can bootstrap with single-player utility, then add a network only when it becomes genuinely useful. Instagram attracted users with free filters and Twitter distribution before its own graph mattered; Dixon says Substack similarly began with email and Twitter. “Come for the tool, stay for the network” solves the cold start while addressing the weak defensibility of standalone tools.
AI defensibility may increasingly live outside the product—in brand, internet-wide distribution, capital, and ecosystem attention. Midjourney tutorials, Cursor’s reputation, search rankings, recommendations, and creator coverage can form an “externalized” network effect. Getting there early enough to “own the meme” matters, but maintaining it still requires costly product velocity; Dixon agrees that capital can eventually become a moat.
Consumer AI is showing rising willingness to pay rather than a zero-sum squeeze so far. Acharya cites Google’s top SKU at $250 a month and Grok at $300, then advances the extreme view that consumer disposable income could become “food, rent, software.” The market may support both ever-larger platforms and the “single-person $100 million run-rate company.”
Small, intense movements are useful leading indicators, but enthusiasm alone is not a market. Dixon looks for often roughly 20,000 “hyperenthusiastic, sometimes cultish” participants with their own language and insider-outsider norms; that pattern informed Coinbase and Oculus. Acharya argues 3D printing lacked an exponential physical-world force, while Dixon counters that Function Health may reflect a slower health movement. The timing risk remains: a movement might unfold over “100 years or 100 days.”
AI founders must choose the right idea maze for a potential decade-long journey while treating today’s interfaces as possibly skeuomorphic. The industry-wide AI “meta-process” can keep compounding even if individual techniques hit diminishing returns, but general-purpose “God models” may absorb some use cases. Prompts resemble AI’s command-line era; the durable opportunities may be in deep domains and new media forms that cannot yet be predicted.
Open source remains the principal check against a handful of model providers collecting rent from every startup and consumer. Training capex makes its steady-state funding harder than Linux-era software, so Dixon’s acceptable equilibrium is that open models remain slightly behind the frontier but broadly sufficient. He is “cautiously optimistic,” while warning that four vastly superior closed systems would be a bad outcome.

Deep dive

1. Exponential forces determine which products become platforms

Dixon’s foundational question is why technology companies can emerge from nowhere, reach hundreds of millions or billions of users, and create outcomes rarely seen elsewhere. His answer is a set of “superlinear forces” that compound rather than merely reward competent execution.
Moore’s law is the canonical example: semiconductor performance roughly doubles every 18 months or two years, while storage and networking improve alongside it. Apple’s insight was not that the first iPhone was already unlimited—it was seeing the curve and building to ride it.
Composability supplied a second curve. Linux grew from a 1990s hobby project because open source harnesses internet-scale intelligence and turns software into reusable “Lego bricks”; as Dixon quotes the principle, “All bugs are shallow with enough eyeballs.”
Network effects are the third curve: Facebook began as a Harvard yearbook and moved by “lily pads” to other schools and eventually the world. In a separate example of technology improving along a curve, Dixon recalls that chatbots had another moment “around 2016 or something,” but were not good then; OpenAI and other pioneers made the bet before the technology improved faster than some optimists expected.

2. The strongest AI moats may emerge after the tool finds users

Dixon’s “come for the tool, stay for the network” pattern starts by making an otherwise empty network useful. Instagram offered filters that competitors charged for and distributed through Twitter; Dixon’s sense is that Substack similarly piggybacked on email and Twitter before its own app began gaining traction as a network.
The network layer varies in strength. Google Docs gains collaboration but remains replaceable; leaving Instagram can mean abandoning a following. Stripe’s Link and Shopify’s Shop similarly extend merchant tools into consumer-facing products; Dixon describes Shopify as now having “kind of a network,” while avoiding the dating-site problem where “no one wants to be on a dating site with like two people.”
Acharya’s pushback is that incumbents now recognize this playbook and deplatform potential threats, while AI tools can coexist through specialization and aesthetics—Midjourney need not look like another image generator. Founders therefore face a real uncertainty: predesign a network, or keep extending the tool until one becomes obvious?
Dixon’s alternative is that network effects may now be “externalized to the internet”: tutorials, influencers, search placement, model recommendations, and brand reinforce products such as Midjourney or Cursor. Timing can help a product “own the meme”; sustained quality, product velocity, and—especially in AI—capital then help maintain the effect.

3. Movements reveal the future, but enthusiasm alone is not a market

Dixon used to spend substantial time in subreddits and niche communities because interesting movements can be led by surprisingly few committed people. Wikipedia, Stack Overflow, and other community projects can be propelled by an often roughly 20,000-person hardcore group of smart, frequently technical enthusiasts with outsized building and distribution power.
The signature is a group that looks “hyperenthusiastic, sometimes cultish,” with distinct language, norms, and insider-outsider boundaries. Following such people took Dixon to Bitcoin. A related hobbyist thesis around 3D printing and VR informed investments in Oculus and Coinbase; he also discusses MakerBot, nootropics and Soylent, and drones.
The method is not foolproof. Acharya argues that 3D printing lacked a physical-world equivalent of Moore’s law; Dixon counters that Function Health may catalyze a large quantified-self movement, with nootropics as a predecessor. Their unresolved question is timing: apparent niches may remain hobbies or compound much later.

4. AI is reviving paid software while weakening the open web

Acharya says metrics suggest that more than 95% of internet traffic and revenue now sit with five to 10 companies. AI could deepen that consolidation by answering questions without a click-through; falling SEO traffic then pushes websites toward more pop-ups, worsening the experience and accelerating the “negative flywheel.”
Stack Overflow embodies the trade-off: Chris says some training data probably came from Stack Overflow, GitHub, and other places, yet Cursor-style tools reduce the need to visit such sites. He calls Cursor “an unbelievable tool” and “clearly good for the world”—good for users, even when destructive to the websites that supplied the knowledge.
Acharya sees “narrow startups” charging high prices for exceptional value and argues, controversially, that “there are no marketing problems, only product problems.” High AI costs can improve business models by forcing consumer founders to monetize early rather than subsidize weak engagement.
When Dixon asks whether paid niches will exhaust higher-paying users and turn to advertising, Acharya says addressable needs can keep subdividing: AI therapy, then ADHD therapy, then a specific life stage and interaction style. Dixon says the market has not been zero-sum so far: prices are rising and “everything feels like it’s working.”

5. The right idea maze matters more than the first implementation

Borrowing the “idea maze” concept from Balaji Srinivasan, Dixon rejects the choice between idea and execution. A founder must enter the right maze, then remain agile for a decade: Netflix kept its subscription-movie thesis while moving from mailed discs to streaming and then original content.
Individual AI processes, including LLM pre-training, might encounter diminishing returns—Dixon explicitly defers to experts—but the larger “meta-process” includes reinforcement learning and many other approaches. Like semiconductor fabrication, one technique can hit a wall while another sustains the industry’s smooth exponential curve.
That creates opportunity and a “brutal” competitive environment. Founders must ask whether general-purpose “God models” will subsume some of their use cases, then defend through domain depth, brand, users, or distribution. Dixon compares the possible churn to the PC hard-drive industry’s “fruit-fly Darwinian struggle.”

6. Prompting is probably AI’s command-line era, not its native form

Dixon defines skeuomorphism as importing an old medium’s grammar: early films resembled filmed plays, early websites resembled catalogs, and early YouTube contained viral clips before native creators emerged. Broadband, cultural shifts, network effects, and generational change all helped the internet’s distinctive forms become visible.
AI image and video generation may likewise automate existing media before creating a genuinely new one. Dixon’s analogy is photography threatening representational painting before cameras enabled film; he suspects AI’s native form will be surprising. Acharya suggests it may take another generation, or five to 10 years, to emerge.
Dixon calls today’s AI the “command-line era”: most people cannot describe desired music as an aesthetic at 110 beats per minute, so prompts are probably a poor long-term interface. Acharya notes that some people now call the work “context engineering”; Dixon points to Spotify history as better input than verbal description, while Acharya speculates that ambient devices could automate context capture.

7. Open source is the economic counterweight to model concentration

Dixon credits open source with making technology inexpensive enough for a $10 Android phone and allowing startups to launch competitive products with hundreds of thousands of dollars—or less. Without it, operating systems, back-end software, and the rest of the stack would each collect fees.
The immediate policy priority is preventing de facto bans: Dixon cites a California bill imposing unlimited downstream liability on developers. The harder structural issue is capital, because AI models require major training expenditure rather than simply “a bunch of coders sitting around.”
As Dixon recounts it, Satya from Microsoft argues that enterprises will fund at least one open alternative. Meta’s Llama, some startups, and China’s open-source push illustrate other possibilities; Dixon points to OpenAI releasing older models as an example of how open source might remain somewhat behind the frontier. A workable equilibrium may leave open models sufficient for most startups and inexpensive consumer health-care advice.
Acharya raises Android as the cautionary tale: nominally open code became operationally closed through proprietary services and permissions. Dixon nevertheless thinks conditions are better than three years earlier and remains “cautiously optimistic”—provided the market avoids four closed providers with vastly superior models charging rent across the ecosystem.

Karl Friston - Why Intelligence Can't Get Too Large (Goldilocks principle)Elon Musk on DOGE, Optimus, Starlink Smartphones, Evolving with AI, Why the West is Imploding