Previously, in The Five AI Waves, Pt 1: AI and Electricity, we compared AI and electricity as disruptive technologies and economic platforms, and we noted they are the two main economic pillars of the 21st century intelligence era economy.
Part of the reason for the analogy between electricity and AI is to tease out useful markers of what we can expect with the adoption of AI. While the adoption timescale for AI is vastly accelerated over that of electricity, there is a common thread: Both are platforms for diverse applications, and as such, we can expect waves of technology innovation and adoption for distinct applications.
Hence the Five AI Waves, based on five broad categories of AI applications and uses:
-
Interactive AI chatbots that answer queries and provide interaction.
-
Generative AI creation tools that generate images, music, video, writing and other content.
-
Agentic AI with AI agents that perform tasks, including software development, and automatically complete workflows.
-
Inventive AI that performs scientific analyses as a research scientist, generates mathematical proofs, develops new molecule or drug candidates, or invents new technologies or innovations.
-
Physical AI in an AI robot, self-driving car, or other physical AI application.
These five applications rest on different kinds of AI models and infrastructure. The maturity of the AI technology stack is different for each application. AI chatbots could use LLMs, but Generative AI requires multi-modal generative AI models, either diffusion or transformer-based but trained specifically for audio, video, or image output. AI agents need an AI ecosystem of a harness, skills, tools, memory, and an AI model capable of using all of them.
AI adoption so far has for the most part been the use of generative AI accessed via ChatGPT or other chatbots. Over a billion users have interacted with AI through ChatGPT or other AI chatbot interfaces. This has been the ‘low hanging fruit’ for AI use cases because AI chatbots require the lowest threshold of autonomous reliability; the user remains inside the loop.
AI agents face a more difficult standard. An agent must make a series of decisions, use tools correctly and recover when something goes wrong. Errors can compound across a long workflow. Scientific AI must meet an even higher standard: its results must survive mathematical proof, experimental testing or scientific review.
Physical AI has the greatest challenge. It must operate safely in an unpredictable world where mistakes can damage equipment, property or people.
The AI chatbot is the easiest to deploy at scale because the user can supervise it one interaction at a time. Creative AI can also be implemented as an assistant for human creators. However, AI applications with the greatest potential are those that automate more activity. They are further behind precisely because they are more demanding. As AI improves in reliability and intelligence, it unlocks more applications and automation.
The bottom line is that there is no single AI adoption curve. For each of these five categories of AI application, they are each climbing their own adoption curve according to their own technical requirements, economics and tolerance for error.
While we enjoyed simpler AI assistants such as Siri for over a decade, the ChatGPT moment and GPT-4-level LLMs kicked off the AI chatbot and assistant wave with truly intelligent AI.
The technology needed for this conversational AI has been LLMs and AI reasoning models. The basic capability in AI chatbots was to answer queries, compose emails, and summarize texts, but as AI improved, the utility has scaled as well, overcoming handicaps such as hallucination with grounding and lack of complex thinking with RL training for reasoning and tool use to leverage tools.
The outcome for this is personal interaction and utility. Accelerating and improving tasks with partial assistance and copilot in chat interactions. With reasoning and grounding, these adaptive AI reasoning assistants can take on research. With further improvement, recent AI models have evolved to support agentic AI, the next category.
In terms of adoption, conversational AI has traveled farthest along the curve of AI applications as it has been directly developed on LLMs. It is now in mainstream use, with over one billion people having interacted with AI globally:
Stanford’s 2026 AI Index estimates that generative AI reached 53% of the global population within three years. The Federal Reserve reported that by November 2025 approximately 50% of Americans used generative AI outside work and 41% used it for work.
Multi-modal generative AI, or creative AI, turns intelligence into an artifact: an image, article, advertisement, song, video, or three-dimensional world. Unlike a chatbot answer, the output is intended to become part of a larger product or communication.
Even before ChatGPT, OpenAI released Dalle and Dall-E 2, diffusion-based text-to-image generation models. These early models generated low resolution images and had artifact errors, but they followed intention and conveyed meaning. With many iterations and improvements since, image generation became a mainstream activity, starting with hobbyists using Midjourney and Stable Diffusion in 2023.
In the last 3 years, AI generative models have been extended to video, 3D rendering, and audio. While hobbyists have been the early adopters of these tools, AI-assisted writing, image generation, design, music and video have entered mainstream professional creative workflows.
Adobe’s 2026 survey of more than 16,000 creators found that, among creators who had used or tried creative AI, 75% considered it integrated or essential to their work and 93% said it accelerated production. This is mainly through assisted creation, as 57% said outputs still required moderate or extensive editing.
AI-assisted writing and image generation are well established, and assisted creation is substantially more mature than autonomous creation. AI image and video generation tools can generate marketing collateral automatically, and AI can generate useful components and drafts at low cost. However, low-cost AI content yields low-value “AI slop” if not managed with judgment.
High-end video, music, film and brand-critical production require human oversight because consistency, copyright, provenance and creative judgment still matter. Human creativity in the era of generative AI increasingly means letting AI work out technical details while human creators exercise editorial control: Judging well among design and content choices, maintaining consistency, and deciding what deserves attention. The key human contribution is good taste.
The AI creation wave is giving us more content than we can consume. The valuable skill for creators will not be generating content but maintaining creativity and quality that makes it worth someone’s attention.
AI agents have been around for a long time, but for several years, AI models powering them weren’t reliable or smart enough to be effective. In 2025, AI models increasingly improved in their intelligence and features to support agentic AI; features such as Skills, memory, tool use, and better context management made things better.
A pivotal moment for the development and adoption of AI agents was the release and viral uptake of OpenClaw, the always-on AI assistant created by Peter Steinberger. This was released in late 2025, just as Claude Opus 4.5 showed itself to be good enough for long-horizon tasks.
With that and the success of Claude Code, AI agents ‘crossed the chasm,’ with early adopters “tokenmaxxing” with their AI agents, especially for software development. Coding is increasingly being done by AI agents with human review and oversight, with AI agents such as Claude Code, Cursor, Cognition’s Devin, Codex, and many others.
AI agent work is expanding beyond software to general knowledge work, and AI agents for general workflows include Claude Cowork, Codex, Grok Bot, OpenClaw, Hermes Agent, and Buzz. Several AI agents include productivity features and connectors such as financial spreadsheets to deal with specific domains, and there are also vertical enterprise solutions for workflow automation in particular domains, such as Harvey for legal work.
The desired outcome from using AI agents is productivity and output, gaining efficiency through automation and expanding what an individual can do. A founder that can’t code can work with an AI coding agent to develop software. In an enterprise trying to streamline what they already do, automation of tasks with AI agents yields higher efficiency and lower costs. An example of the latter is how customer service costs went down at Booking.com.
Since a key benefit of AI automation is cost-reduction, enterprises will be cost-sensitive and only pay for the intelligence they need. AT&T is using hybrid routing to route 25% – 40% of employee AI usage routes to lower-cost open AI models, which cut AI coding costs by 56% for only a 2% quality drop. Many tailored AI agent workflows will be built around non-frontier AI models, with end-users determining the lowest cost AI model that can satisfy each specific use case.
AI agent adoption has crossed the chasm from early adopters to the early majority ‘takeoff’ stage. McKinsey’s August 2026 reports says AI agents have begun moving from demonstrations into useful production deployments, especially in large organizations and in software development, with 31% of large organizations scaling AI coding agents but only 22% of smaller organizations scaling AI agents.
While discrete AI agent task workflows are coming into the mainstream, early adopters are pushing further, managing fleets of AI agents and pushing the automation envelope. In coming years, mainstream adopters will catch up, and enterprise organizations will adjust to continue to scale AI agents.
“When we look back at this time, I think we will realize that we were standing in the foothills of the singularity.” – Demis Hassabis
Invention and scientific discovery are special kinds of intellectual creation. Creating a scientifically correct research paper, a mathematical proof, a new drug molecule candidate, or an improved engine design requires the most difficult, creative and rigorous thinking.
It is the hardest nut for AI to crack. It requires superintelligent frontier AI models or specialist AI models for mathematics and science. An early example of the specialist AI model was AlphaFold; a more recent example is AlphaEvolve. Anthropic’s Fable 5 and OpenAI’s Astra advancing mathematics shows today’s frontier AI models are capable of groundbreaking results.
Thanks to AlphaFold and related genomics and protein AI models, AI is now demonstrably useful for protein modeling, molecular design, and bioinformatics. It’s showing up in drug development pipelines. For example, AI-discovered and AI-designed drug, rentosertib, reached a randomized Phase IIa trial last year.
We are in the foothills of what is eventually possible. We have seen impressive demonstrations, but broad scientific autonomy remains far behind. Stanford characterizes the list of experimentally confirmed AI discoveries as short. Stanford counted approximately 80,150 AI-related natural-science publications in 2025, with AI appearing in 5.8% to 8.8% of research output depending on the field.
However, useful AI assistance to science and invention does not require AI superintelligence. AI systems can already solve smaller technical challenges that accelerate research. Scientists are using AI systems to search literature, generate hypotheses, improve algorithms, design experiments, analyze results and explore candidate solutions.
AI is also becoming standard research infrastructure, for example, with operational AI weather models and biology foundation models providing faster and more accurate simulation than possible before.
Inventive AI is already a powerful collaborator and assistant for invention and science, but it cannot yet replace experimental validation or automate the scientific process. As an AI assistant, inventive AI is in early stages of adoption. The immediate future is not autonomous AI science; it is a human scientist supervising a team of increasingly capable AI research agents. Even without being fully autonomous, inventive AI will be transformative and accelerate science and technology profoundly.
Fully autonomous scientific discovery requires super-intelligent AI that does not yet exist, but this future may arrive sooner than you think.
Physical AI is where intelligence meets the real world. It includes industrial robots, autonomous vehicles, drones, embedded systems such as AI in your fridge, and, eventually, adaptable humanoid robots.
The capabilities of physical AI are highly jagged, and the category does not occupy one position on the adoption curve. Robotaxis have reached commercial scale in selected cities. Older generation industrial robots controlled with hard-coded software (not AI) are mature. Physical AI-based general-purpose robots remain in pilots or limited deployments.
The physical AI technology for intelligent robotics, including humanoid robots, is shown with Gemini Robotics ER 2 which requires two AI models: A high-level embodied reasoning model that understands the physical world and plans tasks; and a lower level intelligent VLA (vision-language-action) model that converts directions physical movements at the actuator level.
Progress in AI models is helping advance AI for robotics, but training for robotics still requires its own data and its own special-purpose modeling.
The central problem in physical AI is that success in a controlled environment does not guarantee reliable performance in the real world. Robots can work well when the environment is predictable and bounded and the task is repeated. Factories and warehouses provide this structure. Homes, construction sites and public spaces do not.
This challenge creates a gap between staged demos and lab results and real-world capability. Stanford found that robots succeed on only 12% of real household tasks, despite achieving 89.4% in simulation benchmarks, illustrating the gap between controlled lab demos and reliable operation in unpredictable environments.
Because physical AI promises the automation of physical work, its economic impact may ultimately be the greatest of all applications of AI. However, progress will be slowest here because of reliability and data challenges; atoms are less forgiving than tokens and the real world is not a bounded. The decades-long gap between demo and commercial use for self-driving cars is a cautionary example.
The Physical AI wave is arriving not at once but as sub-waves, where bounded and restricted robotic applications reach adoption first. Industrial automation will expand in scope and flexibility by integrating AI, autonomous transportation is scaling, and broadly capable general robots are only beginning their adoption curve.
These five AI application waves are not exhaustive, mutually exclusive or strictly sequential. New AI applications will emerge, and the same AI model may participate and support several waves. The categorization provides a framework for understanding how AI adoption will play out, and how AI can be mature in one application yet primitive or insufficient for another.
The chatbot was the first mass-market AI application because conversation tolerates supervision to overcome lack of reliability. Creative AI extends intelligence into content. AI agents turn it into action. Inventive AI applies it to the creation of new knowledge, while physical AI brings it into the world of atoms.
Each step raises the required level of reliability, autonomy and integration. That is why the waves are moving at different speeds.
While AI chatbots have produced the broadest impact thus far due to their wider adoption, the beneficial impact of these later waves, particularly AI agents and physical AI or robotics, will be much greater when they automate significant economic activities.
We have spent the first years of generative AI watching AI itself improve. AI has arrived. The way forward will be defined by what we build with it.
The chatbot was the opening act. The larger transformation begins when AI creates, acts, discovers and moves through the physical world.
