The AI Framework — Develop a Holistic Understanding for Developing, Using and Learning AI

Howdy 🤖,

In this article I’d like to introduce the AI Framework — my current mental model of the entire Artificial Intelligence playing field.

The framework helps me keeping an overview of what’s happening across the many areas of AI — cutting-edge frontier models, new exciting use cases, new tools by AI startups, breakthrough hardware developments and many more. Besides the main aspects of AI development and usage, the AI framework also covers foundational aspects like ethics, governance and the return on investment (ROI) of AI.

Let’s dive in and I tell you more about it…

The AI space was multi-layered with many bits and pieces in place, which enable the development of an AI solution, for a long time. This hasn’t really changed. But it has become even more complex since the steep rise of Generative AI from 2022 onwards.

We see an accelerated mainstream interest and mass adoption of solutions like ChatGPT. This leads to lots of additional development efforts across the levels of the AI stack and a general push to advance AI as quickly as possible to reap the benefits, especially by tech companies.

Every day there seems to be a new groundbreaking announcement within the AI ecosystem. A new foundational model released. A new feature available. You can also read a lot about advances in database technology and ethical & regulatory questions being addressed.

The 7 Layers of the AI Framework

To organize everything that’s going on into a mental model, I have structured it into the following 7 layers, which make up the AI framework in its initial version:

This framework helps to think about AI in a structured way and sort topics that pop up from news or releases into the right section to keep a good overview.

In the following, I’d like to dive a bit deeper and explain, which topics I see as the building blocks of every layer. I also give examples along the way.

Thought Leadership & Idea Layer

• Strategy: Strategies for AI to create new solutions, advance/augment existing capabilities with AI and to embrace AI for internal and external optimization. Examples: Race for the best frontier models, partnerships among tech firms (e.g. OpenAI & Apple partnership), competitive advantages by bundling AI capabilities with existing products (e.g. Canva AI Tools)

• Narratives: Narratives that describe and drive larger changes/developments within the AI industry. Examples: Open-source vs. closed-source models, AI as the new operating system, Generative AI’s potential replacement of Search

• Big Ideas and Features: Aspirational big ideas and features that can be the main drivers of AI benefits and have the potential to describe a whole category of AI solutions. Examples: artificial general intelligence (AGI), agents, multimodality, context windows vs. RAG, memory

• Important Questions: Questions that can make or break AI developments and can guide the advancement of AI. Examples: How do good actors in AI stay ahead of malicious actors? How does open-source help to stay ahead, drive innovation and use the advantages of a large AI community? Will AI be deflationary to the economy and cause mass unemployment?

• Problems: Important challenges that should be addressed and solved to make AI successful. Example: Overallocation of compute capacity, missing or insufficient ROI, unclear path to safe AGI, fair regulation

Usage Layer

• Business/Individual Users: AI application development and usage by both organizations and individual users. Example: Corporate business model (e.g. subscription, pay per token) or individual usage model (e.g. life augmentation, convenience usage, freedom of work) for AI

• Use Cases: Practical usages of AI to create value across industries (e.g. automotive, health care, high tech), functions (e.g. marketing, sales, support, R&D), and types of AI (e.g. document search, text generation, data analysis). Examples: Programming assistants, marketing and sales content generators, anomaly detection solutions in health data and treatment recommendations

• Usage Patterns: Approaches of adopting and integrating AI. Examples: Complete human labor replacement, work/life augmentation

Solution Layer

• Services, Products & Applications: Developing marketable AI-driven services and products that offer tangible benefits and solve real-world problems. They lead to revenue/benefit growth for organizations and individuals. Also, specific non-commercial AI programs (e.g. open-source) that address particular challenges.

• Examples: ChatGPT, Perplexity, Microsoft CoPilot

• Additional important aspects: a) Types of AI services, products, and applications (e.g. translation, document search, data analysis); b) Solution stack and architecture for each type of AI solution

• Models: AI models of various types (e.g. large language models, vision models) that can power applications and lead to improvements across various domains, i.e., drive intelligent decision-making.

• Examples: OpenAI GPT-4o, Llama 3.1 405B, Claude 3.5 Sonnet

• Additional important aspects: a) Model benchmarking and tracking with specific performance indicators for different types of AI solutions; b) AI model types/families, reference architectures and components

Ecosystem and Platform Layer

• Ecosystems: Ecosystems and structures in AI that foster collaboration, innovation, and shared growth as well as interoperability among stakeholders. Examples: Relationship between hardware providers like NVIDIA, cloud providers like AWS and open-source LLM providers like Meta used to build a solution stack for an AI use case

• Platforms: Technical or commercial platforms for AI development, scaling and commercialization. Platforms connect various components and bring players together to, e.g., create services, products, applications and models. Platforms enable to scale and ensure reliability. Examples: AI development platforms like Hugging Face, cloud platforms like AWS, model API and development platform providers like Groq, data platform providers

Infrastructure and Hardware Layer

• Infrastructure: Essential technical systems and components as well as technical frameworks and standards in AI. Those build the basis for advancement of AI technologies and enable seamless operations and scaling. Example: NVIDIA Cuda toolkit that provides a development environment for creating GPU-accelerated applications.

• Hardware: Physical components that are key to AI performance, such as processors and storage devices.

• Examples: Compute (GPU, CPU, TPU), memory, storage, and network equipment

• Additional important aspects: a) Decision about building underlying hardware (make or buy decisions); b) Needs of specific models with regards to hardware stack

Resource and Raw Material Layer

• Resources: Key resources required for AI development. Examples: Data, talent, processing units/chips, energy, cooling, electric wires

• Raw Materials: Fundamental materials and components that are vital for creating and maintaining AI systems. Examples: Rare earths/silicon, metals, fuels and further energy sources

Foundational Layer

• Benefits/ROI: Achievement of returns with AI and definition of economics for financial success of organizations, governments and individuals

• Competition: Competitive dynamics in, e.g., research, development and commercialization of AI solutions

• Regulation: AI regulations of associations of states, nations and industry associations and their impacts on AI innovation and usage

• Compliance: AI compliance frameworks and acts that need to be addressed during AI development and usage

• Sustainability: Measures how AI can be used to reduce environmental damage and promote sustainability

• Governance: Frameworks and structures to govern AI research, development and usage in a fair, safe and effective way

• Philosophy: Fundamental questions about the nature and ethics of machine consciousness, AI decision-making, the relationship between humans and machines, and the societal impacts and moral implications of using AI

• Social Impact: Healthy and unhealthy interaction with AI, implications of AI on humanity, social deficits of AI, implications on human workforces and social welfare

• Ethics: Responsible design, development, and deployment of AI systems, focusing on fairness, transparency, accountability, privacy, and the societal impacts of AI technologies

• Culture/Region: Adaption of AI technologies to local contexts, respecting cultural norms, languages, and values, and addressing region-specific challenges and needs

• Diversity: Ensuring inclusive representation, equitable access, and unbiased algorithms, aiming to reflect and serve the diverse needs and perspectives of all societal groups

• Behavior: AI-supported advancement of behaviors like logical reasoning, creativity, first principles thinking, discipline, self-critics, out of the box thinking, healthy thinking. Also, clarification of what humans can but AI can’t do

• Investing: Investment decisions in AI considering, amongst others, technological innovation and AI market potential.

Summary

The space is so rapidly changing. The layers and underlying building blocks have very different levels of maturity and might develop completely independent from each other over time.

However, advances in one layer or building block can have crucial impact on the whole ecosystem. One breakthrough can be a catalyst for innovation or a big blocker for other parts of the AI framework.

What caught my attention is that most organizations that develop and market AI solutions are specialized players, who focus on one or very few building blocks of AI.

I am thinking of AI startups that create specific AI applications for mostly narrowly scoped use cases, or about hardware manufacturers that produce a special piece of network equipment for AI computing equipment.

Few players are active in many layers and building blocks. When they are, it’s mostly the large tech organizations (such as AWS, Google, Microsoft) or key suppliers (such as NVIDIA, AMD) of the industry.

I hope this framework can help you to better organize your thinking and work in AI.

Let me know what you think and where you see potential for advancement of the framework