Marcin Treder — Blog https://treder.design/blog Marcin Treder — Director of UX, Product & Strategy at Google Play, founding CEO of UXPin, and author of The AI Handbook for Product Design. en-us Copyright 2026 Marcin Treder Marcin Treder Sun, 26 Jul 2026 17:10:28 GMT Augmented Intelligence https://treder.design/blog/augmented-intelligence https://treder.design/blog/augmented-intelligence Sun, 26 Jul 2026 17:10:28 GMT Essay Back in 2019 I worked on a new version of my website. Coming up with the design language and my personal brand was a lot of fun. Coding the website was more mixed. Back in 2019 I worked on a new version of my website. Coming up with the design language and my personal brand was a lot of fun. Coding the website was more mixed. There were some good moments (I always liked writing React!), but also some tedious steps, for example, figuring out the build configuration that I could deploy to Netlify. All in all it was a few weeks of work after my regular job. Yesterday I decided to refresh the website, refactor the code, add a newsletter, a proper blog with an admin panel, and a page about my upcoming book. It took no more than 2 hours to complete everything. Between design changes coordinated by Claude Design, code and deployment done by Claude Code, it just felt seamless. I could entirely focus on quality, providing feedback and steering Claude to add features I needed. The work that needed weeks was compressed into hours. It was incredible, but also scary. Just like everyone else I occasionally worry about AI replacing all of us. After all, what I used to do manually can now be done by a machine. Seeing the machine churn out code at the speed of thought can be disheartening. It is a shallow perspective though. The entire job wasn't replaced, just some of the tasks were. I felt augmented by the machine, not replaced. It was a great feeling. A time study from 1957 In the spring and summer of 1957, J.C.R. Licklider kept a log of his own working day. He was a psychoacoustician running research at Bolt Beranek and Newman, curious about where his intellectual effort actually went, and lacking a better subject he used himself. Roughly 85 percent of what he counted as thinking time turned out to have gone into what he called getting into a position to think: tracking down references, plotting graphs or explaining to an assistant how to plot them, and on one occasion spending several hours reconciling six experiments whose authors had each measured speech-to-noise ratio a different way. Once the numbers were comparable, seeing what he needed took seconds. The observation he made next is the one I keep returning to. He admitted that his choices about which problems to attempt at all had been shaped to an embarrassing degree by what was clerically feasible rather than by what he was intellectually capable of. Something close to that happened to me on Saturday. The 2019 work went faster, and more importantly I built things I would never have started, because back then the newsletter and the admin panel and the proper blog all sat on the far side of a line I had drawn without noticing, the line where a decent idea stops being worth the evenings it would cost. Figuring out the Netlify build configuration was my equivalent of plotting graphs by hand, and the hours it consumed mattered less than the ideas it quietly disqualified. Licklider published all this in 1960 as "Man-Computer Symbiosis," proposing machines that would take on the preparatory work so a person could spend the day on the part that required a person. Douglas Engelbart turned the same instinct into a research program two years later under the title "Augmenting Human Intellect," and the mouse came out of it. Making people more capable was one of the two things the field originally set out to do, alongside building a machine that could think on its own, and it carries a longer pedigree than the recent reassurances would suggest. I heard Jensen Huang reach for the same idea in an interview recently, almost in passing, while describing how Nvidia's engineers work now. He called it augmented development, without offering it as terminology, just grabbing the ordinary word that fit what he sees at the desks in his own company. What the payroll records show Optimism is only worth having if it survives the bad numbers, so here is the worst one I know. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen went through ADP payroll records covering millions of American workers and found that early-career workers aged 22 to 25 in the most AI-exposed occupations have seen a 16 percent relative decline in employment since generative AI became widely adopted, controlling for firm-level shocks, while more experienced workers in the same occupations held steady or grew. The adjustment ran through hiring rather than dismissal, which makes it easy to miss, since entry-level roles are being left unfilled rather than cut. This doesn’t sound great and anyone arguing for augmentation has to sit with that before saying anything else. This is not the whole story though. Those employment declines concentrate in occupations where AI substitutes for the work, while in occupations where it complements the work, employment grew. The authors drew that line empirically, from observed patterns in how people query Claude within each occupation, which means the variable predicting whether a category of work expands or contracts is the shape of the interaction between the person and the model rather than the raw capability of the model itself. It’s nuanced, and I certainly don’t want to be dismissive of people who are in fact being replaced by AI, but whenever you are experiencing this sunken feeling “am I going to get replaced?” think whether AI is able to handle your entire job, or just some of the tasks. Interactions between humans are an essential part of many jobs even if today we tend to downplay their role because we’re busy with technicalities. That’s certainly true for most of the tech roles. The shape of that interaction gets designed Anthropic published a comparison in June 2026 that I have not stopped thinking about since. Looking at the same task across two of their own products, they found that the median chat session producing a blog post ran to thirteen rounds of back-and-forth, while the median Claude Code session producing a blog post contained a single human prompt. The gap held when they compared conversations served by the same underlying model, and they concluded that the product being used mattered more than the model serving it. I spent most of Saturday on the more automated of those two surfaces and had the more augmented experience anyway, because I kept steering and rejecting and treating each output as a draft rather than a delivery. The surface leaned one way and I leaned harder the other way, which is the honest version of what I am claiming here. Neither the model nor the interface fully determines the mode you end up in, though both of them tilt the floor, and the people building the interface are the ones deciding which way it tilts. Defaults do most of that work. A system that hands you a finished artifact invites a different relationship than one that hands you something you can argue with, and the distance between those two products is usually a few weeks of engineering and a decision nobody remembers making. What happened to bank tellers Economists have an optimistic story about all this whose ending rarely gets told. David Autor used automated teller machines as his example of technology complementing labor: ATMs went from roughly 100,000 to 400,000 machines in the United States between 1995 and 2010, and teller employment rose across the preceding three decades rather than collapsing, because cheaper branches meant more branches and the job shifted toward work the machines could not do. Teller employment peaked around 2007 and has fallen since, with the Bureau of Labor Statistics now projecting a further 13 percent decline between 2024 and 2034. The tasks that protected the job in 1995 were themselves automated twenty years later. For a long time I read that as a warning, and lately I have come to read it as a description of how the mechanism works. Augmentation describes a relationship that holds for as long as somebody keeps choosing it, which also means it becomes available to be chosen again every time the models improve. The 1995 arrangement expired more or less on schedule. Nothing prevented anyone from designing its 2015 replacement except that comparatively few people set out to. Conclusions The machine absorbs the tasks that never really needed a human being, and the hours it frees go back into the work that does need one – deciding what is worth building, and then building it well. On Saturday I got two hours of exactly that, and I finished the day better at my own job than I have ever been. Just think about it… was the complexity of the build configuration that I had to deal with 7 years ago an essential aspect of building a website, or just an artifact of imperfect tools and processes? I have no doubt that it’s the latter. And now without this “task” I have more time for the actual job. This task reduction will run throughout every job leaving humans to do the thing that we were deprived of doing – focusing on the essence of the value we want to provide.

References Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30. https://doi.org/10.1257/jep.29.3.3 Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence (revised 13 November 2025). Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine Engelbart, D. C. (1962). Augmenting human intellect: A conceptual framework (Summary Report AFOSR-3233, SRI Project 3578). Stanford Research Institute. Licklider, J. C. R. (1960). Man-computer symbiosis. IRE Transactions on Human Factors in Electronics, HFE-1, 4–11. Massenkoff, M., Lyubich, E., Sacher, S., Hitzig, Z., Zhang, S., Heller, R., & McCrory, P. (2026). Anthropic Economic Index report: Cadences. Anthropic. https://www.anthropic.com/research/economic-index-june-2026-report U.S. Bureau of Labor Statistics. (2025). Occupational outlook handbook: Tellers. https://www.bls.gov/ooh/office-and-administrative-support/tellers.htm

]]>