There is a widening hierarchy of participants and drivers in AI. This is not unusual, as almost all technological revolutions show this type of fractional distillation.
1) AI creators (working on models & other source techniques)
2) AI experts (implementing structured projects and products)
3) Experts-in-the -loop (using tools like Claude code etc)
4) Humans-in-the loop (using chatbots)
5) Humans who don’t use AI
It is particularly acute in AI, as there are a huge conceptual and competence gaps between these different types of participants.
AI creators
Those who work in the large foundational model companies, such as OpenAI, Google, Anthropic, Meta and others are experts both in the fundamental mathematics and techniques. It is a mistake to focus entirely in the training of models here as the field has widened out into advanced techniques before and after training that are proving effective.
This is the most opaque group, as the surrounding system that makes those models useful, reliable and economical viable in the real world are highly technical. Companies now rely on techniques to ground responses in external knowledge, tool use and function calling to let models act rather than just speak. They use agentic architectures to plan and execute multi-step tasks and long-term memory to maintain context and personalisation across sessions. There’s also serious evaluation and guardrails, synthetic and curated data pipelines, inference optimisation to cut costs and latency, even human-in-the-loop training. The centre of gravity has shifted to these more mature enhancing techniques because real value now depends on accuracy, trust, controllability, cost and integration into workflows. These problems that are solved at the system and product level rather than in the training run.
AI experts
These are the companies and researchers who are building the next layer, on the basis of the fundamental technology. They may enable model orchestration, routing tasks across multiple specialised models as a service to customers. They are also creating vertical products in law, healthcare, training, education etc.
Many are involved in RAG and similar techniques to create useful projects and products across and within organisations. This is similar to the growth of internet companies in the early 2000s on top of the internet, now a massive portion of the world economy. I’ve been involved in this layer for over a decade. It’s pioneering, exciting and difficult but without this layer of experimentation the core technology would remain aloof and unusable. It takes commitment by CEOs and a real push to get through prototypes to product.
Experts-in-the-loop
As the technology matures a new group emerged which used more advanced techniques and tools, often the ‘pro’ models. This can be the sophisticated use of AI to handle and use data. But the dominant use has turned out to be coding and on a wider scale media production (text, graphics, images and video). My book on AI and Productivity’ explains why this emergent group is the key to solving the productivity problem and overcoming productivity paradoxes using AI.
Experts bring domain knowledge, this is usaally a mix of tacit and explicit knwoe3ldge and skiils that allow them to get the best out of the available tools. Coders make goof users fo Clade Code, graphic artists of image creation tools and will video creators in creating video. Experienced researchers will get the best out of the technology for research and so on.
Humans-in-the loop
By far the largest group are the ordinary users. One could split them into the two lower fremium groups of paid and free users but both sit on a rising tide of quality and functionality. They tend to single promtt and use the technology as a form of performance support, namely solving an immediate problem; finding something they don't knopw, communications issue, summarising, translating, rewriting and so on. They tend to prompt simply and as the technology improves get more and more bang for their buck.
Humans who don’t use AI
This last group are the disinterested or those who have no real need, which is fine. It also includes those who actively dislike the technology. They have all sorts of objections, some rational, others irrational and ill-informed.
Least informed but loudest are this last group, at least on social media. They often invoke an eye-roll from the rest, not because they don’t have a point but because the claims are often unsubstantiated or just plain wrong. Typically they will refer to content from old models, rely on anecdote and make claims about the technology hitting brick walls.
Conclusion
We should not be surprised by any of this. The fractional distillation into these sedimental layers of expertise is both a normal and necessary part of the evolution of a technology. In truth, as the technology matures and improves, this hierarchy tarts to massively expand in the middle layers, while the bottom layer shrinks. Think printing, cars, television internet. In this sense resistance is futile.
Ultimately the hierarchy turns in to fully realised rout and it becomes a normalised consumer product woven into the fabric of our lives. That will happen faster tyan most people think, within a few years. It is a hierarchy that automates, thereby blurring the lines, flattening it out and dropping the doubters.
At this stage of AIs development, only 3 years in, the last group is still the largest and therefore the loudest, that does not mean it is the most significant, merely the least informed and practiced.
