Disclaimer: This represents my personal observations regarding AI progression and discourse, and does not reflect my employer’s views or opinions.

TL;DR

1. The modern AI stack is crystallizing into four layers: Infrastructure (compute, chips, and energy), Models (foundation systems), Tooling / Middleware (retrieval, orchestration, databases), and Applications (domain- and user-specific AI products).
2. The application layer will be the most dynamic frontier. Bret Taylor and Ben Horowitz argue a dynamic range of solutions will be born in this layer. Augmenting and applying foundation models will lead to a new wave of AI companies.
3. Differentiation will come from how systems model domains and users, construct context, and generate outputs that are expressive or exact depending on the need.
4. Context construction is the bridge between knowledge and action. Crude systems rely on static prompts; sophisticated ones assemble and adapt the necessary information to meet the user’s needs.
5. Two key axes define the application landscape: Orientation (domain-oriented ↔ individual-oriented) and Outputs (expressive ↔ exact).
6. From these axes emerge four archetypes: the Troubadour (creative worlds), the Scholar (precise reasoning), the Court (personalized expression), and the Chamberlain (trusted orchestration). Each archetype reflects human roles once reserved for elites - artists, advisors, attendants - now scaled through compute to millions of people.
7. Rather than a single leap to superintelligence, AI progress will unfold through thousands of bespoke systems built along these archetypes: tailored, contextual, and increasingly personal.

The Application Layer Before Superintelligence

The AI stack is taking form. At the base are Infrastructure and Model layers - capital-intensive, scale-driven businesses that will consolidate around a few players. Above them sit Tooling and Middleware platforms that handle retrieval, memory, and orchestration - the connective tissue linking models to data. And at the top lies the Application Layer, where these capabilities are translated into human and business outcomes.

As Bret Taylor and Ben Horowitz have both noted, the real frontier in AI isn’t building bigger models - it’s augmenting and applying them in novel ways. The next wave of innovation will come from companies that use these lower layers as raw materials to construct differentiated systems for specific domains and users.

The mechanics of that differentiation hinge on how effectively a system models the world it operates in, understands the individuals it serves, and assembles the right context at the right moment. This essay unpacks that chain in detail - showing how mastery of the world, understanding of the individual, and precision in context assembly will shape the archetypal solutions to emerge in the years ahead: Troubadours, Scholars, Courts, and Chamberlains.

First I’ll unpack some core concepts and then jump into the existing and future landscape of the Application Layer.

Modeling Domains and People for Targeted Use Cases

The defining strength of application layer AI lies in how it models domains and models users - constructing structured representations of both the world and the individual to produce targeted, contextually grounded outputs.

Modeling domains means encoding structured knowledge: laws, medical research, company documentation, creative canon, or financial data. The challenge is curating a clear, authoritative view of truth that can be reliably retrieved at the moment of use. These systems must know not just what’s true, but what’s relevant now. For example, regardless of whether you’re an intern or an executive, an AI assistant should provide roughly the same concise summary of “what topics were covered in the last company all-hands” by drawing from a shared organizational record.

Modeling users means constructing a dynamic model of an individual’s life in motion - their routines, goals, preferences, and constraints - by synthesizing signals across devices, services, and history. This data may live in timelines, interest graphs, or long-term model memory, constantly updated as behavior evolves. When prompted, the system must determine which signals matter now to deliver a contextual, personalized response. A query like “how can I get better sleep?” will differ meaningfully for each user depending on what the model knows about their habits, health data, and daily rhythms.

Together, these two modeling approaches define how AI will move beyond generic outputs toward systems that understand both the world and the person within it.

Crude vs Sophisticated Context

Once a domain or individual has been modeled, the next challenge is deciding what belongs in memory when reasoning over a task. This goes beyond just context window size, and instead focuses on the specific tokens chosen to go into the context window. Context construction is the bridge between knowing and doing - it determines which information is loaded, ignored, or synthesized before the solution acts.

Crude context construction is shallow: instructions, a prompt, and a hint of past interaction. Sophisticated context construction, by contrast, is selective and situational. It synthesizes signals across data sources - documents, history, sensors, and structured models - to assemble just enough context to reason intelligently. As Cursor CEO Michael Truell notes, high-performing systems already rely on “ensembles of models under the hood,” orchestrating retrieval, reasoning, and tool use rather than sending a single API call to a foundation model.

The difference becomes clear in a simple question: “What should I have for dinner?” A crude system might respond with follow-up questions or generic ideas, relying on short-term chat memory. A sophisticated one would draw on relevant, dynamic context - knowing that you’re traveling for work, landing at 7 p.m., staying at a hotel without a kitchen, tracking your dietary goals, and facing rain by 8 p.m. - to suggest a restaurant nearby that fits your needs.

The sophistication of a context window isn’t measured by its size but by its judgment: what to include, what to exclude, and when to adapt. That judgment will be a critical hallmark of tomorrow’s AI applications.

Mapping the AI Solution Landscape

Spectrums of Solutions

AI solutions in the application layer can be understood by what they know, who they know and the types of outputs they produce. These distinctions play out along two axes: Orientation and Outputs.

The horizontal axis of Outputs spans solutions whose outputs are expressive ↔ exact. Expressive systems generate prose, imagery, and ideas that aim for emotional coherence rather than technical correctness. Their strength is imagination and style - creating something plausible, novel, or resonant. Exact systems focus on verified truth. They retrieve facts, reconcile data, and deliver the “right” answer. Their success is measured in correctness, not creativity.

The vertical axis of Orientation spans solutions that orient towards domains ↔ individuals. Domain-oriented systems specialize in traversing structured fields of knowledge - finance, medicine, design, or even creative canon. Individual-oriented systems specialize in modeling the behavior of users - their routines, goals, preferences, and environment.

The incoming archetypes replicate human roles

From these axes emerge four archetypes, each reflecting a distinct path for how AI will build and apply context in the years ahead. It’s worth noting that for most of human history, these roles were specialist knowledge workers limited in number and only on-call to the most well-equipped and powerful. What’s striking about this moment in history is that compute will make these archetypes accessible to anyone. Tasks that once demanded entire households, libraries, or ensembles can now be simulated in silicon, replicated endlessly, and personalized for every individual. These roles, once reserved for monarchs, are becoming available to the masses.

1. Troubadours create expressive worlds through creative mastery.
2. Scholars deliver precision through disciplined reasoning.
3. Courts blend creativity with personal resonance, shaping experiences around individual taste and identity.
4. Chamberlains act with precision on the machinery of personal life, quietly managing day-to-day details

These archetypes form a map of the AI application layer - how systems model domains, model individuals, and wield context to generate expressive or exact outputs.

Expressive Master of Domain: The Troubadour

Troubadour solutions sit at the intersection of expressive outputs and domain orientation. Like traveling entertainers in ye olden times, they master a creative domain and can ad-lib and remix while still honoring the source canon. These solutions capture and reference visual identity, stylistic constraints, and lore so they can generate new novel experiences that feel authentic. The goal isn’t accuracy as much as a coherent “vibe.”

In practice, this looks like encoding a brand universe: businesses could taxonomize their voice, visuals, and messaging rules into a structured solution that generates endless work products or marketing assets (copy, visuals, ads). IP owners could intelligently reference character history, visual identity, and critical lore about their worlds for new creative ventures (think new offerings for Harry Potter or Star Wars or Nintendo). Much like trained Princesses and Mascots at Disney Parks not breaking character, outputs from these AI’s would aim to preserve the integrity of the creative output and stay true to the source material. 

We see hints of this world coming together. Brand solutions like Figma Buzz lay the groundwork for asset creation at scale by codifying guidelines and augmenting with AI. Inworld and Nvidia ACE are early examples of preserving convincing “characters” in simulated environments. Image and video solutions like runway and Pika along with audio solutions like udio and Eleven labs show the flexible creative stack that will power future agencies and in-house creatives. Over time, Troubadour systems will evolve from assisting creators to autonomously producing content across media (short-form video, full-length episodes, games) - each work generated from richly constructed context modeled on creative assets and canon.

Exact Master of Domain: The Scholar

Scholar solutions sit at the intersection of exact outputs and domain orientation. Precise and objective - they seek out and synthesize authoritative sources to serve up verifiable truth. Their purpose is reliability over originality - producing the correct answer quickly and consistently. Being “right” matters quite a lot, because erroneous information results in lost trust and fact-checking all future responses. 

These systems organize and retrieve structured knowledge in ways that accelerate expert work - whether that knowledge comes from curated industry datasets, scientific research, or internal company documents. Scholars model the logic, hierarchy, and interdependencies of a domain so the AI can reason within it, not merely recall it. We’re surrounded by early Scholars in the form of various “co-pilots:”

• Horizontal solutions cut across existing company knowledge bases: Google’s NotebookLM, Notion AI, and Microsoft 365 Copilot are practical examples of AI tapping internal documentation to formulate informed responses. 
• Vertical domain solutions model industry-specific knowledge like Harvey in law and Abridge in healthcare.
• All purpose solutions seek to turn unstructured materials into durable knowledge artifacts: Granola synthesizes distilled talking points from meetings, Delphi offers ways to transition their knowledge into a chat-ready knowledge base.

Expressive Personalization: The Court

Court solutions sit at the intersection of expressive outputs and individual orientation. Like the royal entertainers of empires past, which adjusted their ballads and plays to the sovereign’s tastes, these systems tune creative expression to the preferences, emotions, and histories of each user. Their artistry lies in personal resonance rather than factual precision, crafting experiences that feel made for you.

Under the hood, they rely on sophisticated taste modeling - learning what delights, moves, or motivates each user. The interest graphs that power TikTok, Instagram, and YouTube hint at what’s coming next: fully generative content streams that create original videos, stories, or music aligned to each person’s aesthetic. We’re already seeing early prototypes in OpenAI’s Sora app and Meta’s Vibes, precursors to a “YouTok” era of personalized creation.

Advertising will evolve in parallel. Campaigns will be dynamically generated to show products in familiar, emotionally resonant contexts: a car cruising past your local landmarks, a coffee maker brewing your favorite roast, sports gear demonstrated by your favorite player. While still nascent, platforms like Adobe’s Gen Studio and Jellyfish’ Pencil hint at this vertical integration. 

In gaming and entertainment, adaptive worlds will learn from player behavior - rewriting quests, dialogue, and endings in real time. Physical toys like those pioneered by Curio & Bubble Pal may even retain a kind of “memory,” recalling past adventures with their owners - like if Buzz Lightyear of Woody could reminisce with Andy about their past adventures together. Each interaction deepens the feedback loop between system and self, transforming personalization from surface-level relevance into something closer to relationship.

Exact Personalization: The Chamberlain

Chamberlain solutions sit at the intersection of exact outputs and individual orientation. In royal courts, the Chamberlain was the trusted attendant who managed the sovereign’s private affairs - overseeing the household, schedule, finances, and correspondence. It was a position of deep intimacy and precision, quietly ensuring that the machinery of a monarch’s life ran smoothly. AI systems in this archetype will play a similar role for the digital era: ever-present, discreet orchestrators of modern existence.

These systems will construct and maintain a continuously updated model of your world - calendars, communications, finances, travel, and physical environment - to act on your behalf with minimal friction. Their intelligence lies not in conversation but coordination: knowing what matters, when to act, and how to weave information across surfaces and services into coherent action.

Early glimpses are already here. Google’s Gemini and Android ecosystem offer a clear path toward a unified life layer, integrating Gmail, Drive, Docs, Calendar, and Maps into a shared context. Meta’s Ray-Ban glasses extend this logic into the physical world - AI that can “see what you see and hear what you hear,” bridging perception and memory. OpenAI’s rumored device partnership Jony Ive hints at a similar “always-on” assistant, while Limitless offers the beginning offering for total recall - continuous, searchable memory from audio taken in daily interactions. Whoop’s new AI features convert their 24/7 tracking into a personalized health hub grounded in each user’s biometrics.

The challenge for Chamberlains won’t be capability; it will be trust. To inhabit a user’s private life demands discretion, reliability, and restraint. Apple and Google start with a structural advantage through device ubiquity and long-earned confidence. New entrants will need radical transparency - or devices so valuable they outweigh privacy hesitations. Over time, the Chamberlain may fade into the background entirely, becoming less of an assistant and more of an invisible operating system for personal life.

Closing thoughts

This framework isn’t a forecast of inevitabilities so much as a lens for exploration. I’ve found it useful for thinking about where new opportunities may lie - and how the next decade of AI products will differentiate themselves not by the size of their models, but by the sophistication of their context. The defining skill in the application layer will be engineering understanding - deciding what an AI should know, when it should know it, and how it should act on that knowledge.

Perhaps one day we reach artificial superintelligence, but for the foreseeable future, progress will unfold through thousands of bespoke systems that excel within these archetypes. Each will model domains and individuals in its own way - some mastering the logic of the world, others the rhythms of human life - and together they’ll populate the vast terrain of the application layer.

I expect this framework will evolve, just as the systems it describes will. For now, it helps me make sense of where we are: a moment when the specialized attendants of history - the troubadour, scholar, court, and chamberlain - are scaling up and out in digital form.

Thanks Bill Palombi, Danny Delaney, and Brice Morrison for helping me frame my thoughts for this post!

‍