The Platform: General Implementation Capacity
The concept of “Implementation capacity” as a workflow — design, development, production, deployment, and adaptation — provides a framework for understanding transformative AI.
Advances in AI will expand general implementation capacity — the ability of humans to achieve broad goals, not by waving a magic AI-wand, but by helping us to design, develop, deploy, apply, and adapt complex sociotechnical systems at scale. The conceptual implementation workflow from design to application and adaptation provides a framework for thinking about tasks, tools, costs, and bottlenecks, and how AI capabilities could be applied to accelerate progress and scale results.
As a bonus, dissecting implementation into concrete stages — design, development, production, deployment, and adaptation1 — helps us consider prospects for strongly transformative AI without debating the possibility and potential misbehaviors of willful, omni-capable superintelligent agents.
Stages overlap and interact, but can be divided like this:
Design: Generating and evaluating alternatives2
Development: Iterative testing, refinement, and modification
Production: Making physical products
Deployment: Applying systems in the real world
Adaptation: Modifying systems based on experience
AI can accelerate every stage:
Generative AI can accelerate design and development.
AI-based automation can accelerate development and deployment.
Interactive AI systems can accelerate application and adaptation.
As always in this series, “AI can…” is shorthand for “As AI advances in capabilities and generality, it can eventually…”3
The implementation workflow framework
The workflow framework outlined above differentiates and organizes the various aspects of implementation capacity. I will be saying more about these in later posts, but here’s a taste with some simplifications:
Design
Design begins by articulating goals and desired functionality, often serving broader objectives. In systematic design, informal descriptions are refined into specifications that establish performance criteria and metrics. Guided by these, the design process explores and evaluates alternative approaches, incrementally refining and selecting preferred architectures.
AI can accelerate and improve design. Systems with deep knowledge of technological possibilities and human preferences can converse4 to clarify fuzzy aims. Generative design models5 can propose creative alternatives, drawing on both fundamental principles and past designs. Conversation and generation can be iterated, refining designs and perhaps changing goals. From vague aspirations to detailed plans, each step can be accelerated and improved.
Development
Development interleaves testing and design refinement, producing prototypes that mature into fully functional implementations. AI can accelerate development by building and testing prototypes more rapidly, gathering and evaluating performance data to correct and improve designs. By accelerating design iteration, AI-enabled automation can both accelerate development and improve products.
Production
In transitioning from prototypes, production typically specializes and automates manufacturing systems engineered during development. AI can cut cost and time not only through robotic automation, but also by designing specialized high-throughput machines that are faster and cheaper than robots.6 Production systems themselves are products of an implementation workflow.
Deployment
Deployment at scale requires not only production, but distribution, integration, and use. Humans can provide direction, while AI can do most of the work, handling tasks like transportation, assembly, and setup. Conversational interfaces can make systems self-explanatory, minimizing the need for training. With good design and automation, even complex systems can be easy to use.
Adaptation
Application of complex systems drives adaptation by surfacing problems and opportunities for improvement. AI can accelerate and improve adaptation by monitoring performance, recognizing problems, and collecting human feedback. Improving systems requires exploring options and making choices — which is to say, design — whether this leads to a software tweak or a full round of system development, production, deployment, and further adaptation.
Applications of AI to implementation will reduce costs and accelerate deployment of new sociotechnical systems at scale, slowly at first, but more rapidly as AI capabilities grow stronger and more general. The expansion of general implementation capacity will transform possibilities, hence options, interests, and goals — and anticipating future change can change the present.
I’ve recently refactored and renamed this sequence for greater clarity. The definitions and boundaries are soft.
Note that a really ambitious design task might call for generating grand, superintelligent-level schemes that would span decades and mobilize vast resources. Generating grand plans, however, is a bounded, episodic task; as always, competitive optimization tends to favor models that enable faster, less expensive iteration. As with generating text or images, generating plans does not require models with internal goals or external agency. And like generated images, most plans will be examined, compared, and discarded.
This is not the agentic intelligence we were looking for, in part because intelligence isn’t a thing.
Questions of “when” are secondary to questions of “what” and “whether” (see “Enough AI, and a threshold”).
I say “converse”, but this should be understood as a multimodal exchange that may include not only words but gestures, images, videos, VR, and interactive models.
Generative design models for complex systems will require multiple components that fill diverse roles in scaffolded, iterative, interactive architectures. Development of generative design models for systems will build on generative models for their parts.
For example, a single continuous-motion assembly machine can assemble a single product at rates >300 million per year (fast machine here), but these machines are complex and can do nothing else.