Act 1 came from the technology-out — foundation models as a new hammer generating a wave of novelty apps. Act 2 must come from the customer-back, solving real human problems end-to-end. The biggest challenge is not finding use cases but proving lasting value — retention, not novelty. Invoked Amara's Law: we overestimate technology in the short run and underestimate it in the long run. As foundation models commoditize, the real value shifts to the application layer (product, UX, workflows). Published alongside the V3 generative AI market map.
Introduced the generative AI market map and thesis. Argued that a new class of AI had emerged that creates rather than merely analyzes — shifting the marginal cost of creation toward zero. Every industry requiring original human work (coding, design, law, marketing, gaming) is up for reinvention. Outlined four waves of development from transformer breakthroughs to killer app emergence. Became the foundational reference document for the entire generative AI ecosystem.
AI is progressing from 'thinking fast' (rapid pre-trained pattern matching, System 1) to 'thinking slow' (deliberate reasoning at inference time, System 2). The reasoning layer — inspired by AlphaGo-style approaches — endows AI with problem-solving that goes beyond pattern recognition. This enables 'service-as-a-software,' where the addressable market expands from the $1T software market to the $10T+ services market. Foundation layer has stabilized to five scaled players. Reasoning models are strong on logic-proximate domains (coding, math, science) but still developing on open-ended domains (writing, strategy).
Identified a massive gap between the revenue expectations implied by the AI infrastructure build-out (projected from NVIDIA's data center revenue run rate) and actual revenue growth in the AI ecosystem. The AI industry needs to generate $600 billion annually in revenue to justify current hardware spending levels — raising the question of whether this is sustainable or a bubble.
Declares that AGI is functionally here. Long-horizon agents — AI systems that can sustain multi-step work, correct their own errors, and persist toward goals autonomously — are the realization of AGI. Coding agents are the first proof point. One litmus test: can you hire an agent? In 2023-2024 AI apps were chatbots; in 2026-2027 they will be doers that feel like colleagues. Usage shifts from a few queries per day to all-day, every-day autonomous work. Human roles shift from executor to manager of AI teams.
If 2024 was the 'primordial soup' year for AI, the building blocks are now firmly in place. Data centers are the new rails of the digital economy and will be securely built by end of 2025. Five finalists emerged from the big model race (Microsoft/OpenAI, Amazon/Anthropic, Google, Meta, xAI). AI search will proliferate (Perplexity hit 10M MAU). The key question shifts from 'can we build it?' to 'what freight will ride on those rails?'
AI has reached its 'synesthesia moment' — models that natively understand and generate across modalities (text, image, code, video, audio, voice) in a unified latent space. AI synesthesia converts strengths in one cognitive domain into capabilities in another: if you write well but cannot code, AI bridges the gap through semantic representations; if you design beautifully but struggle to pitch verbally, AI transforms sketches into narratives. Creativity becomes translation, expression becomes multidimensional, and intelligence becomes fluid.