Embedded Files
Right now, in 2026, knowledge work is being erased, bottom up, by machines
Right now, in 2026, knowledge work is being erased, bottom up, by machines
This is not intended to be hyperbolic. In the absence of massive global governmental intervention to stop AI development and deployment, knowledge work as we know it will be gone within a generation, possibly even within a few years. Just as the industrial revolution replaced most physical labour with machines, the cerebral revolution will replace most thinking labour with AIs.
Programming, my industry, is the canary in the coalmine
Programming, my industry, is the canary in the coalmine
If I returned to my engineering management job, I would no longer hire junior programmers. I would hire senior developers and give them all the AI tools money can buy (which is not even much money. I’m using a revolutionary tool to achieve things I simply couldn’t do before for $17 a month). This is initially going to be a problem for junior developers, who will not be able to find work. This could become a problem for companies who can’t hire seniors, because a senior is just a junior with some experience, and there won’t be any of that any more. However, that won't be a problem because AI is getting better at a continuous rate and I don’t see any reason for it to slow down. For ordinary software projects, I’d estimate in perhaps three years AI will take over senior software development roles (architecting projects, reviewing code, setting technical direction) for most projects and increasingly few technical (and product) staff will be managing increasingly powerful AIs.
Other knowledge-based industries will be similarly disrupted in the near future
Other knowledge-based industries will be similarly disrupted in the near future
I saw a video of some architects laughing at an AI-developed plan. It was, to be fair, terrible. However, I laughed at some terrible AI code a couple of years ago. I’m not laughing now.
Design is already under pressure. AI-generated images were not great two years ago. Now Nano Banana is producing images that I think are pretty hard to tell from reality. Design has always been undervalued (fiverr was founded fifteen years ago with the premise of creating any piece of art or design for $5) but for the majority of people, AI is now much cheaper than an artist and also much quicker to iterate than a real person. This was free and took me just a few minutes to create¹:
Other design disciplines like architecture will be similarly pressured, as will writing-based jobs like marketing. Tasks like copy editing are already pretty much automatable. A retired friend of mine who was senior in communications was telling me they tried prompting ChatGPT to write a statement similar to their previous work and were amazed at how good it was on its first try.
AI-first entertainment already exists in the form of AI-slop, but the slop is getting better and better, and whilst it isn’t gourmet, it is cheap, plentiful, and somehow makes you want more. It’s UPF for your eyes. It’s a perfectly engineered cheesy tortilla chip compared to a real Mexican meal. I imagine Netflix will move first here and I wouldn’t be surprised if within ten years we have personal TV shows tailored to our beliefs and vices.
After design, I imagine law (which is already very structured) and very expensive will be the next domino to fall. Here, practitioners may have been burned by early tools that hallucinated cases, but law-specific tools will not do that (and even consumer grade tools have basically learnt to search the web and not hallucinate nearly so much any more.)
Education is on a huge inflection point.
Education is on a huge inflection point.
Teaching is a huge piece of knowledge work and is ill-prepared for the AI revolution. On the one hand, cheating on homework and coursework is going through the roof. Bad AI usage is still (currently) detectable, but careful use of AI is impossible to detect and will remain so. Non-AI-using students will struggle to compete with peers using AI, and I suspect we will see a combination of more written exams, and also more open projects which showcase what the student (with the AI) can achieve.
But without discipline in how an AI is used, learning is harmed by AI usage. Formal teaching will probably require training AIs which are less obsequiously helpful and more Socratic in how they answer user queries.
However, if an education system can figure out how to use AI to personalise learning and tutor every individual student, that form of mastery learning is known to vastly improve outcomes. If we're lucky, the children of the 2030's will be learning from something like Neal Stephenson's Young Ladies Illustrated Primer, an immersive book and virtual world which teaches through a child's interests and develops a curriculum exactly suited to their learning.
This is a new cerebral revolution, where not labour but thought is being replaced.²
This is a new cerebral revolution, where not labour but thought is being replaced.²
Improvments to education are broadly irrelevant if we cannot figure out how humans can thrive in this new world.
I’m ignoring existential risks here. I think they are real but I do not have a good basis for evaluating the level of risk. If Anyone Builds It, Everyone Dies is on my reading list (and I’ve followed Eliezer Yudowsky for a long time) but I still assign a low and uncertain probability to the worst-case scenarios.
Previous industrial revolutions have replaced labour, in a way which is often disruptive to the people doing that labour at that time, but has generally been positive for humanity (perhaps aside from the environmental externalities). Computer scientist Alan Kay once famously described the computer as "the bicycle of the mind". But the modern AI is not a bicylce for a mind; it is a facsimile of a mind.
The replacement of thought means that, instead of humans moving up the value chain and becoming more useful, humans will instead be pushed out of the value chain completely, as they simply cannot do anything a machine cannot do cheaper.
This will present a society-changing political question: do we find ways to support a growing unemployed population by harnessing the output of machines for the common good, probably through a universal basic income funded by AI-specific taxation? Or do we see an increase in wealth at the very top and poverty for everyone else? The current trends are not very encouraging, but this will be the defining political issue of the 21st century.
I do not believe the genie can be put back in the bottle. On the political front, I think now is the time for the middle class to begin demanding UBI and real taxation of multinationals, whilst they still have power, lest we end up back in a feudal system with 99% of the population living in penury. On the technical side, I’m riding the wave because I don’t see any other option right now.
¹ Actual prompt I used: "I want to make a three panel cartoon which apes several different styles. The panels should be laid out side by side. In the first panel, an unemployed artist with a long face and spiky hair holds a sign saying "Will draw cartoons for food", in the style of american dailies. Low levels of detail, not extraneous words or illustrations. The second panel is the same artist in the street and dressed raggedly, but in the style of Dilbert, specifically the eyes have no pupils. Don't put the artist in a suit. Keep a street background, not an office. A unix guy with a white beard, braces and quite rotund is reaching out to hand the arist something and saying "Here's a nickel kid, buy a nano banana". In the third panel, style of Futarama (slightly futuristic and bold color) the artist (still with the long face and blonde spiky hair) sits on a throne attended by adoring but classy robot girlfriend, feeding him grapes, with rows of robots furiously drawing onto stacks of real paper." It was the about the fourth revision of the prompt I tried (and I updated my ideas as well as tuning the output as we went.)
² Actually, that is not true. In software development we have been building tools that do a some of our thinking for us for a long time, starting at least with Admiral Grace Hopper and others first compilers in 1950’s. But current LLM-based AI tools seem qualitatively different.
² Actually, that is not true. In software development we have been building tools that do a some of our thinking for us for a long time, starting at least with Admiral Grace Hopper and others first compilers in 1950’s. But current LLM-based AI tools seem qualitatively different.
Report abuse