AI eats the world (Spring 26) [pdf]

Platform Shifts and the “AI Era”

• Thread opens by comparing past eras (hardware, internet, mobile, cloud) and arguing each birthed new giants, implying AI will do the same.
• Some see this era-bucket list as arbitrary and not proof that incumbents fade; others say the point is about where new value is created, not old firms disappearing.

Models, Moats, and Commoditization

• Presentation deck series is read as moving from “maybe this is a platform shift” to “models likely become infrastructure; value moves to apps, workflows, data, GTM,” but framed as provisional.
• One side: open-source and multiple strong models (e.g., DeepSeek, others) at lower cost suggest rapid commoditization of the “model layer.”
• Other side: frontier models may form a duopoly/monopoly akin to advanced fabs; smarter models earn more, fund more compute, widen the gap. Counterpoint: current training costs still too low to force monopoly, and scaling dynamics remain uncertain.

Compute, Open Models, and Local Use

• Debate on whether compute is the real moat.
• Examples of running strong open models locally and via many third-party providers suggest some erosion of central control, but self-hosting remains expensive.
• Some expect specialized hardware to enable local frontier-ish models; others think datacenters will capture most such hardware.

Usage, Products, and UX

• Data from the decks suggests daily AI use is still low, even in tech; several see this as evidence we’re very early.
• Legal and other sectors are expected to change heavily but face institutional inertia and likely restrictions on “official” AI use.
• Disagreement on chat: some call it poor UX and “barely a product”; others see conversational interfaces as the best mix of power and simplicity.
• One view: chatbots are being used as a discovery surface for valuable use cases that will later be wrapped in agents and specialized apps.
• Another view: low daily usage shows limited use cases, not just capacity constraints; others argue compute shortages and weaker free tiers are artificially suppressing adoption.

Agents, Coding, and Small-Business Software

• Coding agents are widely seen as a strong current use case.
• Some argue better “harnesses” plus smaller models can deliver reliable, cheaper coding systems, enabling custom software for small and tiny businesses.
• One participant is building a multi-role, tool-rich coding agent system as an example of this direction.

Revenue Metrics and Hype

• Discussion of “annualized” revenue = last 4 weeks × 13; used by fast-growing startups where full-year data is uninformative.
• Acknowledged as effectively a prediction and potentially gamed, especially with volatile, usage-based revenue.

Technical Trajectories: Scale vs Structure

• Concern that trillion-parameter models resemble a “mainframe era,” potentially hiding large inefficiencies.
• Several discuss combining neural models with rule/heuristic or symbolic “world models” to get more compact, deterministic, domain-scoped systems.
• Debate over how narrow coding models should be: tightly scoped models that “don’t know about Ewoks” vs the idea that broad background knowledge actually improves coding performance.
• Mixture-of-experts is seen as an early step toward more structured, domain-partitioned intelligence; longer-term efficiency potential is viewed as a major unknown.

Societal Impact, Power, and Control

• Some worry AI data centers and solar infrastructure will “crowd out” humans and nature, citing cases where datacenter needs overruled local interests for land or electricity.
• Others insist technology should serve people, but note that, under current systems, it primarily serves shareholders and the wealthy.
• There’s skepticism toward relying on “experts,” given conflicts of interest and the field’s speed of change, alongside vague calls for collective political action.

Adoption, Bubbles, and Historical Analogies

• Many see strong parallels to prior tech waves: massive capex, fear of missing the platform shift, and bubble-like dynamics.
• Some emphasize that historical predictions around the internet and mobile mostly missed the actual winners and use cases, arguing for humility and acceptance that we’re likely “asking the wrong questions” about AI today.
• Others note that, unlike past eras, today’s incumbents are hyper-aware and aggressively investing, which may alter how displacement and consolidation play out.