Interface is Everything: The Evolution of Digital Interaction and the Role of AI

In the rapidly evolving landscape of technology, interfaces are paramount, shaping how users interact with digital products and, increasingly, with artificial intelligence. This exploration, featuring insights from Wesley Yu, Head of Engineering at Metalab, delves into the evolution of interfaces, the technical challenges they pose for backend systems, and how AI, despite its sophistication, often functions as a highly advanced form of a CRUD (Create, Read, Update, Delete) application.

Metalab, a firm specializing in designing and building digital products for ambitious companies, focuses on identifying the product DNA of early-stage ventures that aim to become household names, such as Slack, Coinbase, and Uber. They also collaborate with established leaders like Google, Amazon, and Meta to evolve existing products.

The Journey into Software Development

The path to software engineering is often non-linear. Wesley Yu's journey began not with traditional programming training but with photography and media studies, focusing on visual storytelling and communication. This foundation in understanding how to present information and engage an audience proved invaluable, later translating into content marketing for a tech startup. Exposure to the vibrant energy of Silicon Valley in 2013, despite initial failures in learning to program, eventually led to a coding bootcamp and a successful career in engineering leadership at Metalab. This background underscores a deep appreciation for the art of presentation and its technical underpinnings.

The Transformative Power of Interfaces

A well-designed, simple interface is often the key to a product's success, making complex technology understandable. The advent of the iPhone in 2007, for example, revolutionized user expectations, introducing adaptive screen sizes and gestural inputs like scrolling, pinching, and swiping, which replaced traditional point-and-click interactions. Interfaces transformed into direct manipulation tools, allowing users to drag, expand, and long-press elements. Furthermore, access to device capabilities like location, cameras, and sensors unlocked entirely new possibilities. The ability of mobile phones to deliver notifications also built "habit loops" into applications, drawing users back and demanding attention, sometimes in ways that feel more invasive than inviting.

With mobile devices, the interaction evolved beyond just users interfacing with the device; the device also began to interface with the user, particularly evident in social media platforms.

Crypto, Web3, and Information Density

While often contentious, the crypto and Web3 boom also established new user interface expectations, particularly around verifiability and trust. Unlike traditional systems where institutions provide trust and recourse (e.g., customer support), blockchain transactions are irreversible. This necessitated interfaces that clearly displayed information density, allowing users to verify on-screen actions while maintaining simplicity. This trend highlighted the importance of balancing comprehensive information with intuitive design.

Technical Paradigms for Evolving Interfaces

Modern interfaces, with their gestural interactions and sensor-based inputs, are largely extensions of existing paradigms rather than completely new technological shifts. A more significant impact on technical implementations comes from the rise of real-time collaborative productivity applications like Google Docs, Figma, Notion, and Miro. These tools introduced expectations of "presence" with multiple cursors, alongside features like history, versioning, and inline comments. This shift towards real-time collaboration places immense pressure on backend systems to ensure stability and consistent state across multiple users.

Managing state in such applications is a significant engineering challenge. Strategies involve tracking individual changes and robust merging capabilities to maintain a coherent system state for everyone. The rise of serverless and stateless systems often clashes with design demands, such as a simple "someone is online" indicator, requiring careful planning for potential future stateful requirements. The expectation now is that tasks are completed collaboratively, whether with other humans or, increasingly, with AI systems.

AI: Enabling Businesses and Accelerating Tasks

AI integration in products typically falls into two categories:

1. AI that enables businesses that couldn't exist without it: An example is an application designed to teach children to read. It automates and scales interactions between expert teachers and students by generating custom storybooks based on reading levels, using both text-based LLMs and generative image AI for illustrations. A speech-to-text model evaluates phonemic accuracy, breaking down sounds to a granular level. In such a product, AI is core to the experience, often invisible, and indispensable.
2. AI used as an accelerator or alternative modality: Consider a product for policymakers, lobbyists, and journalists that aggregates vast amounts of unstructured policy data (government documents, meeting minutes, articles, directories). While this data is valuable on its own, LLMs offer a powerful way to structure it for easier querying. Users can ask questions ranging from "What are today's top health policy headlines?" to complex queries like "What bills is Coca-Cola lobbying for in 2025, and how much have they spent?" or "Which lobbyists used to work for the Department of Energy, and what bills are they lobbying for now?" While users could manually sift through the information, AI significantly enhances efficiency, acting as an interface to this data.

Ultimately, many applications, including those leveraging AI, serve as interfaces into data, with developers essentially performing CRUD operations on databases.

Designing AI Interfaces: Shared Memory and Context Management

When designing interfaces for AI systems, it's beneficial to view them as a "pair-programmer" or a collaborative entity. Existing paradigms from productivity applications and coding IDEs offer valuable blueprints, including version control, diffs, commits, and inline comments, which make interactions with AI transparent and understandable for developers.

A critical concept in AI interaction is "transactive memory," the psychological idea of shared knowledge within close relationships or teams. As users spend more time interacting with AI, they develop a shared memory. This memory needs to be visually represented and stored within the interface, preventing its loss when an LLM session ends. This involves building context on the backend, often using knowledge bases or knowledge graphs like Neo4j, to enable systems to query this shared memory. Context compression and understanding the context window are also vital considerations, even if not explicitly thought about by the average consumer.

Managing context is a complex challenge, as larger context windows can lead to "context rot" or the "needle in the haystack" problem where LLMs struggle to retrieve information from the middle of a large input. Solutions include Retrieval Augmented Generation (RAG), which employs vector search for unstructured text, and more recently, graph RAG. This technique involves using LLMs to write queries (e.g., Sparkle queries for Neo4j) against knowledge graphs, which model entities and relationships. This approach builds "just in time" context that doesn't solely rely on the LLM's immediate context window, enhancing retrieval of up-to-date information.

Building a knowledge graph involves defining an ontology – the concepts (nodes) and relationships (edges) within the system – in a way that is understandable to both the LLM and the end-user. For a policy application, this might involve nodes for articles, people, votes, and bills, with specific edges defining their relationships.

Beyond Text: Voice Interfaces and the Friction of Learning

While chatbots provide a common LLM interface, voice interfaces are emerging as powerful tools, acting as a "peer programming partner" for quick lookups and documentation queries, especially when engaging with complex content. However, the manual process of summarizing and extracting information from these voice interactions remains a hurdle that needs to be addressed.

Regarding the criticism that LLMs reduce the beneficial friction in learning, it's argued that a reduction in friction can be positive if AI is designed to invite curiosity and accelerate learning. The ideal scenario envisions an LLM as a perfect tutor, removing all impediments to learning except self-motivation. This vision aligns with historical trends, such as the increased accessibility of marathon running, suggesting that technology can elevate broad capabilities rather than just elite performance.

However, in learning interfaces, careful consideration is given to avoiding systems that simply give answers, as this robs users of the learning experience. While many companies prioritize efficiency and automation, even at the cost of deep understanding, the design challenge remains to mandate user engagement in a way that facilitates true learning.

The Future of AI-Generated Interfaces

The prospect of AI spontaneously generating interfaces on the fly presents a complex future. While AI can verify code or features, verifying whether an application truly meets consumer needs – which often requires market validation – is significantly harder. Humans currently excel at understanding how other people solve problems, structuring tasks, externalizing memory, and implementing progressive disclosure to manage information density. While it's conceivable that LLMs could be taught these design principles in the future, the verification challenge suggests that human designers will remain crucial for ensuring applications genuinely meet user needs for the foreseeable future.