(Welcome to the new home for my newsltter, on a European and independent platform! If you were reading my texts in Substack, now you will be reading them here). For the past few months I've been using Claude and other agentic systems to explore what happens when we live in their worlds. While the creative research project that I have been putting together is not totally ready yet, I do have some opinions about gen AI and some of its potential impacts. But first, some premises: corporations behind frontier models are the grifter enablers of political and aesthetic totalitarianism, and disregard the climate conditions they are imposing in the world with their predatory apocalyptic-cult like business. They should be made economically and politically accountable for their theft of people's work, and we should be finding ways to make generative AI be part of the commons (nationalize AI!). Agentic AI has a direct and immediate impact in work. As a technology produced in and by capitalism, it follows a long tradition of automating tools that allegedly increase productivity while maximizing surplus value for the owners of the means of production. Yes, we may think about them as very new technologies, but their business model is something that Charles Dickens and Ned Ludd would recognize. But do they work? And how do they work? Artists, musicians, writers, and coders are seeing their skills devalued because these systems can produce plausible results that please bottom-line obsessed managers. But these concerns, while valid and based on reality, are not accurate. AI will not steal our jobs. Talentless middle managers and soulless capitalists will. However, agentic systems change the way we work. And I think there are ways to think about how work will be different. My own work has led me to think about this as a 10-80-10 rule. In game design, there's a truism that states that the last 10% of a game production takes 90% of the time. That's because making the game fun is proporcionally less "generative" work, but more "intensive" work. You may build the infrastructure for your game, write the code, build the levels, or print the cards, and that takes 90% of the work. But making that infrastructure fun is the real work, and it takes real time. With agentic systems, this approach is valid if slightly different. Agentic systems make producing 80% of the work faster. But there's a 20% that will take 90% of the time, and that is an exclusively human activity. Using agentic systems makes some aspects of a creative production faster, but they are not necessarily more productive, or overall faster. This 100% human slop graphic is my model to think about the impact of agentic systems in creative projects:
Obviously not created with agentic AI The first 10% of any creative project is focused on intentionality: what do you want to make? What intention do you have to put something in the world? Agentic systems do not have intentionality, so they are totally outside of this loop. This is a task of and by people, to translate desires and needs and emotions and everything else into the intention of making something. Intentionality is, of course, deeper than just "I want to do this". Intentionality is a matter of taste and knowledge: understanding what you want, why you cannot find it anywhere else, and how you want to put it in the world. It is a hard process, one that requires talent and dedication and knowledge and, again, taste, a capacity for discerning the aesthetic value of what you intend to do. This 10% can take up time, and it cannot be automated. Of course, you can move quickly through this process. Maybe you just want a poster, a jingle, a paragraph. Fine, you have a vague idea of your intentions and you put that vague idea on the automating machine. The result of not putting work in this part of the process is sloppy - the proverbial AI slop. Nothing generated by an agentic system can be more than slop if there is no intentionality in the work. Once intentionality is in place, agentic systems become useful. They can execute that intention within the limited parameters of their models, their training, and their corporate design. Execution is what these systems are good at: translating intention into a thing. And that can go fast, and requires less human effort. Of course, execution still requires human effort. The natural output of these systems is mediocre. Plausible, functional, but mediocre. Sloppy. Humans need to intervene on steps of the execution to make sure that the results align with the intention. This means that expertise is essential for good results. Just because these systems execute on a series of instructions does not mean that the result is good. It just means that something has been executed and generated, that needs supervision. But the truth is, agentic systems streamline the 80% of the work that is execution. The last 10% is experience, or impact. That is, whatever gets executed needs to be aligned with the intention. Creating something means relating to others through a work: code, music, text, images, they are not things we consume, they are things we relate to, that emotionally impact us and that have that human value. Making sure that the result is intentional requires, once again, human agency. This 10% cannot be automated because agentic systems do not have relational value, or values. If a project is just execution, which is the pipe dream of the AI evangelist and their ghoulist puppeteers, then the result can only be slop. Sure, efficiently, fastly produced slop, but without a drop of human intentionality of relationality it's just the tasteless paste that comes out of turning all human expression into a fuzzy map of interrelated numbers. Agentic systems make the quantitative part of work faster: the 80% of execution. But it's the first 10% and the last 10% that are quantitatively more important, and what should take most of the time. Without intentionality and relationality, there is only execution. Execution is something a machine excels at, but can only appeal to the dead souls of machines and the people possessed by them. The other 20% takes 90% of the time, and machines cannot do it. It is the task of unsubstitutable people, and the only thing that matters when making anything. It is what makes anything a contribution to the world, and to others. ------- Some final notes: I use agentic systems exclusively to write code. I think that a possible future of this systems is to become a kind of "compiler" that allows us to use natural language to create software, but if I knew anything about the future, I would not writing newsletters. I allow myself to write code because I do not build products: I build ridiculous software. I would not trust any of these systems to generate production ready, costumer-facing software that costs money. But bespoke, experimental software, sure, why not. Yes, paying for these systems means supporting tools that are environmentally and politically wrong. I try to minimize my usage of these systems to the implementation of projects that can only exist in them, but I understand that this is a contradiction with my political views. It is a contradiction I can live with. In my practice, I do not use these systems to write, make music, create images, or videos. I enjoy doing all those things and I don't want a machine to take away the joy, but I don't judge people who can live with the contradictions I have accepted to live with. Agentic systems could become tools for liberation, each of us creating bespoke systems to solve our specific needs, destroying monopolistic software and the companies that benefit from it. But that requires that we create sustainable models, that we dismantly the thievery frontier model corporations, and that we destroy the totalitarian governments that see in these tools the instruments to create deskilled vassals. Agentic systems should be free, local, and environmentally efficient. They should not be theirs, they should be ours. |