As an avid victoria 3 player, I can’t help but make the analogy. We are developing a fundamentally 5x production method for labor markets high level. There are, of course, sub-industries within that to get there gradually, but this is exemplified somewhat in my earlier piece on AI-enabled services which explores this at a more banal low level.

Where the Victoria 3 Analogy Holds

1. Reconciling academia and industry
   1. Many note today that the interplay between academic research and industry is a unique phenom - that NeurIPS should turn into a largely career-focused conference after a few iterations should not be surprising. When nitrogen fixation became a vastly important tool for massive surpluses as opposed to subsistence farming - what did large commercialized agriculture conglomerates in the late 19th century do? When electrical experiements yielded working lightbulbs, and the combustion engine first released, how much industrial funding from large conglomerates went into fantastical innovations like an “automobile in every industry?”
   2. Victoria 3 represents this somewhat with the Industrialist interest group funding higher education buildout and some companies whose presence actively improves research in their specific areas (see Nokia).
   3. The difference today is that the vast availability of private capital and complex early stage financial funding projects means that commercializable research funding, especially for AI, is commonplace. Victoria 3 caps your “innovation” in terms of discovery velocity by tying it to your country’s literacy - the idea being that having sufficient universities to match your country’s “innovation cap” by how many of your pops are literate can roughly track to your rate of innovation.
2. Luddites —> data centers
   1. When downstream economic effects are more evenly distributed, preponderonce of luddites decrease. Urbanization in Victoria 3 is a proxy for easier dissemination of economic changes because it is easier to control the success of industry via production method changing and policy that improves social mobility.
   2. Link to Stalenhag piece
3. Jevons Paradox (overutilized)
   1. Every new paper that Deepseek releases that makes N-1 even more cost-effective, or every new architecture (like bitnet) that gets to clear scalability, will increase demand because this will make complex model workflows costs more available
      1. This doesn’t simply mean the cost of an api making calls cheaper. Its the compounding network effects behind standard practices to implement best in class agentic systems that makes it so you need less and less MLE expertise (and subsequent “R&D” budget to actually build sophisticated working agentic systems consistently)
   2. When electricity first became conceived of as a power source, the only use cases were electric lighting to replace the gas lamps of Victorian England and electric saws for logging. The first private power plant owners, with inefficient hydroelectric systems, never imagined that the commodity they produced would power anything being a cog in the city’s lighting system supply infrastructure. This is a similar dynamic to today’s “mature markets” for AI apps —> coding/search as the first mature use cases and subsequently the first private power plant owners viewing their product not as a general utility provider, but as a cog in the first mature markets’ supply chain (see Anthropic and coding models). Attempts to build out models then for other markets, then, are seen as separate products (Anthropic’s current attempts at computer use and finance agents) even though they should all coalesce to be the same product by this logic.
      1. If we take this historical analogy along its furthest length, we will notice how electricity, when its uses became common in every industry, became largely state-owned. Thought this is for a longer writing on the possiblenationalization of AI companies, this is possibly likely (although we contend with the fact that everybody has access to N-1 models just via having access to electricity and open source).
4. Abstractions of work | Cottage industries and guilds, and resistance to AI adoption taking over skilled work
   1. Every technological innovation has abstracted away some skill/expertise such that the same work can be replicated with less labor, in essence abstracting away low level reasoning. Automatic looms did this with cotton spinners. Automated glass bottle blowers did this with glassworks. Rotary valve engines did this with furniture manufactories and steelworks.
   2. The physical industrial revolution brought about commercialized agriculture and hyper efficient manufacturing via production lines. With this volume complexity came a swell of need for white collar workers with baseline mathematical, reasoning, and context education - education that we trusted universities to bestow.
   3. Today, the digital industrial evolution will do what automated spinning looms did for spinners, but with the office worker. Their work will be abstracted, they will complain and commit to luddite behavior, their productivity will quickly look bimodal compared to the likes of those who choose to upskill more quickly.
      1. There are no set rules and instutitions for becoming “AI-native” today besides those with the most agency, but then again there was never really an institutional rulebook for when John Cockerill set up the first continental industrial societies in Belgium, or when Russians in Baku made Azerbaijan the site of over 50% of the world’s oil production by 1910 (despite oils’ use cases and abundance being discovered far before), or when Qing officials had to figure out western-style munitions manufacturing at China’s Hanyang Arsenal by decree of the self-strengthening movement.
   4. In this way, when data companies today espouse the “future of work” for an AI age, they are really, in essence, describing the upskilling that happens when some technology realeases for a 10x improvement. In the Victoria 3 time age - this is de-peasanting - whereas the availability of drastically better jobs in urban settings, as well as more efficient agriculture such that we could achieve better agricultural gains with less people - meant that its in the players’ economic interest to move everyone to work in the factories. By way of making peoples’ livelihoods inefficient and forcing them to seek better employment - you could improve their spending power as urban consumers for the very goods they worked to produce in factories. In a more banal manner, this is the reason why banning slavery is good in Victoria 3 - slaves aren’t very good consumers and allocations of human labor.
5. Oil and rubber —> data
   1. In Victoria 3, an extremely large inflection point in GDP growth within the game’s timespan is the shift to hyper efficient oil and rubber-based production methods that occurs around the 1890s-1910s. In game, these are represented by the impact of the following technologies:
      • Machined Tooling for Tooling Workshops from Elastics
      • Automated assembly lines for automobiles and the creation of personalized transportation vehicles
      • Oil as a fuel for railways, plastics creation, vacuum cleaning, nitrogen fixation, and as a general hyperefficient fuel for power plants which power electric arc welding, electricity, and conveyor belts on their own
      1. Of course, we have to take into course the impact that production methods efficiency today not only impacts raw output, but also has a much greater effect on labor savings by essentially duplicating workers with agents
         1. This is also somewhat represented in the past by labor automation technologies (steam donkey —> rotary engines —> railroads —> conveyor belts)
   2. We can speed up/bring pre-industrial countries and economies into the industrial age via technology sharing
      1. Though these production methods were first discovered and utilized only by Western European nations, by 1936, the value of oil had a use in every industry and global marketplace demand, as opposed to 1870 with [famous inventor’s discovery].
6. Last mile adaptation of RL products to enterprise workflows
   1. A problem that Victoria 3 abstracts away and does not deal with is how fast dissemination of new production methods occurs. It prefers to abstract this away for gameplay reasons and assumes that production method implementation is instantaneous and that technological know-how is confined within a nation-state. This is obviously inaccurate (nation-states during the Victorian period were neither sophisticated enough to have instantaneous knowlege dissemination on this level (how could you expect oil mining practices in Surgut, Baku, Chita, and Ploiesti to standardize at the exact same time?
   2. Today, the last mile is truly the longest mile. The knowledge workers that a business needs to hire for the latest production methods are extremely costly and difficult to hire, moreso to hiring railroad engineers in 1890s Buryatia when the requisite expertise is all based in St. Petersburg. Our economies have gotten so privatized that the latest production methods are often confined within private companies as well. But crucially different - is that we have multi-hundred employee companies with messes of knowlege bases that are massive ordeals to make AI-ready.
   3. Jaya Gupta recently wrote a piece on “context graphs” which got a lot of traction on twitter. In Victoria 3, the context graphs that enterprises needed were largely physical and could be mapped out geograhically, on org charts, and manhandled with Pinkertons. Today, the context graphs are digital, and require code to manhandle to whip into shape to roll out new production methods.