mimics inside @lovelymimic - Tumblr Blog

Born on technology’s third base

If you'd like an essay-formatted version of this post to read or share, here's a link to it on pluralistic.net, my surveillance-free, ad-free, tracker-free blog:

https://pluralistic.net/2026/08/21/world-historic-forces/#multicore

Any frank assessment of your own achievements starts with an equally frank assessment of the world-historic forces that attended those achievements. For example, I often tell young people who want to get into tech, "Well, if you don't have the foresight and work ethic to be born in 1971, I can't really help you."

When it comes to tech, being born in 1971 – to a computer scientist father, no less – conferred a tremendous advantage for my career chances. My dad – a refugee – came to Canada at a time when post-war public services meant that he could become the first person in his family to go to university, all the way to a doctorate.

That set me up for life in a house where tech and education were all around me. Both my parents are teachers, both from working class families where no one had ever gone beyond high school, who found themselves in a time and place where it was easier than at any time in history for people from backgrounds like theirs to attend university. I got to go to university, too, at a time when education was cheap enough that I could drop out of four schools before figuring out that it wasn't for me, and still be debt-free, largely thanks to income from a series of part-time jobs.

When I dropped out of my final degree program, it was to take a job in tech at a time when anyone with a little creativity, work ethic, aptitude and curiosity could walk into a career. Millions of us did it, and I ended up working as a freelancer, then founding a startup, and then going to EFF. I know I work hard, I know I apply myself to understanding the world around me, but also…when it comes to this kind of career, I was born on third base.

There's plenty of this to go around. Think of boomers who bought their "starter home" with the income from their first job and traded it in for a succession of larger, nicer homes, each of which skyrocketed in value. Some of those people fancy themselves to be veritable Warren Buffets for having had the shrewd financial insight that buying a house and living in it was a good idea. The truly smart ones know that they just got lucky.

There are world-historic forces all around us, creating moments and circumstances that contribute to the life-thriving of those of us who are lucky enough to be suited to the moment we find ourselves in.

Take computing: for decades, computing was ruled by Moore's Law, an unbroken run in which computers got faster and cheaper every year. If you were interested in the kinds of computing applications that were well-suited to serial computation – programs that worked best when run on a single computer – you were in luck. Even if your application or field of study was expensive and difficult to realize on today's computer, you could just stand still for a year or two and a much faster computer would park itself on your doorstep, ready to solve your problems.

When Moore's Law tapped out – when the pace at which transistors got smaller and computers got faster slowed and plateaued, and the expense of even modest performance gains climbed infinitywards – computing changed with it. Parallel computing – putting more cores on a chip, more chips on a board, more boards in a system – took off, as chipmakers and system builders switched from a focus on building their computers tall to building them wide.

As parallel computing took off, so did parallel applications. This is the beginning of the graphics revolution, as GPUs – components made up of many, many low-powered computers – became more central to academic research and commercial product roadmaps. But it wasn't just graphics that saw a huge lift here: any task that could be parallelized got easier and cheaper to perform every year, in a steady trend that has run to this day. This is the era of performance gaming, VR and AR, cryptocurrency, and, of course, AI.

In What Technology Wants, Kevin Kelly introduces the idea of the "adjacent possible" through the example of the helicopter. Da Vinci sketched a "helicopter" – blades in the shape of maple keys attached to a kind of wine-press screw – in the 15th century. In the centuries that followed, many other people had the insight that twirling blades of that shape on a screw of some type could provide lift for some kind of heavier-than-air craft. But it wasn't until strong alloys, internal combustion engines and light, energy-dense refined hydrocarbon fuels came on the scene that the helicopter became possible, whereupon it was all but inevitable, with several people independently inventing the helicopter all at once:

https://memex.craphound.com/2010/10/13/kevin-kellys-what-technology-wants-how-technology-changes-us-and-vice-versa/

In the same way, the computing industry's focus on parallel computing made life easier for people who burned to do something parallelizable. Then the achievements of the parallel computing partisans drove more investment in improvements to parallel computing hardware and theoretical work on how to parallelize other problems. This feedback loop raised the profile of parallel computing applications, attracting more bright and ambitious people to those applications, whose even more impressive accomplishments brought more people into the field, more capital into hardware development, and more resources to parallelization research.

The point being that world-historic forces, combined with accidents of history, shape the outcomes of individuals, companies and disciplines. It's much easier to be an accomplished graphics wizard in an era in which GPUs are doubling in power every year than it is in an era when linear computing is getting the lion's share of investment and improvement.

These forces and accidents have acted on AI in ways that profoundly shaped its development. The latest AI boom started when a group of machine learning researchers tried a minor variation on existing techniques and saw a major improvement in the outcomes. This is one of the most exciting kinds of breakthrough: if tweaking a single variable in a small way produces a large improvement, then it may be that further tweaking will produce even more improvements.

The minor variation that produced the major improvement in AI performance was scale. Prior to the "deep learning" era, AI research relied on a mix of hand-built models of reality and training data that computers fitted into those models. Deep learning swapped the painstaking work of describing reality in software for a brute-force approach: throw lots more training data at the system and then throw lots more (parallel) computing power at that data and let the computer figure it out without your having to explain how the world worked.

The early gains from this approach were very exciting: they dangled the promise of software that could essentially "teach itself" how to do complicated, valuable things in a series of accelerating returns. The fact that the early improvements in AI systems that used this technique were so much greater than anyone would have expected based on AI research up to that point dangled an even more exciting promise: that the improvements would continue to scale faster than the inputs.

Researchers and investors came to expect an AI that was "untouched by human hands," that taught itself how the world worked. This was the self-licking ice-cream cone of machine learning, the world of "theory-free inference" that had fueled the Big Data industry. With theory-free inference, you don't have to figure out how the world works in order to act upon it: you can just gather up all the data about how things happen in the world, use statistical methods to find the correlations, and then intervene to change the outcomes. You don't have to know why a molecule improves a medical condition – it's enough to discover that fact, produce that molecule, and administer it to people with that condition.

Lots of stuff in the world works this way. Our understanding of the causal relationships that make up reality has massive holes in it that we fill with mere correlation. Correlations are easier to discover than causes, and while correlation is (famously) not causation, causes and effects are correlated, and if you can evince the effect you're seeking without understanding precisely what happened to make that effect appear, well, at least you got the effect you were seeking.

Theory-free inference is a very pragmatic way to approach the world: "I don't need it good, I need it Thursday." Scientists burn to know why a molecule stopped you from dying, but you are likely satisfied to not be dead. What's more, our ability to observe correlations will always race ahead of our understanding of causality, so the power of theory-free inferences pushes out the frontier of things we can act on, beyond the realm of the understood.

Which is all to say: it's reasonable to be excited about a breakthrough in theory-free inference. But just like a boomer who thinks that buying a house to live in makes them a shrewd real-estate speculator, someone who achieves great things through theory-free inference runs the risk of missing the limitations to those techniques.

And they are limited. Theory-free inference is good at predicting what your spouse will type into their phone based on all the things they've ever typed into their phone. You are also good at guessing what your spouse will say based on the things they've said before. The difference is that when your spouse says something entirely unexpected and unprecedented to you (say, "I want a divorce"), the fact that you have a theory about why your spouse said all the things they said up to that moment can help you understand why they've said this new thing. But a machine learning model that relies on theory-free statistical modeling to predict your spouse's next words will be entirely at sea. Theory-free inference works well, but it fails badly.

The problem is that the AI sector has raised literally trillions of dollars by assuring investors that the era of hand-made, causal world models that let computers act on the world is hopelessly inefficient and outdated. But there are many, many tasks that are vastly more efficient and reliable when done through conventional computer programs, rather than through "AI."

As Gary Marcus describes in a recent Organized Money interview, an LLM can recite the rules of chess, but it can't play chess because – lacking a theory of how chess works – it will just emit statistically likely chess moves, even if those moves cause pieces to illegally move through other pieces. The first conventional chess-playing programs ran on electromechanical proto-computers, and they played a better game of chess than an LLM that uses billions of times more computing power and energy:

https://www.organizedmoney.fm/p/an-ai-expert-explains-the-hype

The AI companies have proved that there are many domains and applications where we can swap scale for understanding. But, having ridden some world-historic forces and adjacent possibles to great fortunes and stature, they cannot be dissuaded from their conviction that theory-free inference and scale can do everything. They can't be convinced that in many cases, the things that scale and theory-free inference can do are much better accomplished through causal understandings and conventional computing techniques.

From a research perspective, it is interesting to learn about the potential and limitations of a model trained on the entire internet. From a societal and industrial perspective, it is often grossly wasteful, inefficient and unreliable to swap scale for understanding.

The AI sector was born of world-historical forces that favored massively parallel computing, forces that had also conjured up an internet with trillions of documents that could be fed into those massively parallel computers to conduct theory-free inference. Like every success, AI was born on third base.

As rent-burdened millennials who abandoned avocado toast and fancy coffee and still can't afford a downpayment will tell you, the fact that being born in 1945 made it easy to trip and land on a couple million dollars' worth of real estate wealthy by the time you reached retirement age tells us nothing about how to solve the housing crisis of 2026.

By the same token, continuing to give trillions to AI companies because they experienced early success with theory-free inference at scale tells us nothing about how to solve the vast range of problems that theory-free inference at scale sucks at. Doubling down on AI to overcome its increasingly obvious limitations is like doubling down on building post-war suburbs to fix today's housing market.

It's possible to achieve impressive feats because you're smart and hard working and also because you were in the right place at the right time. Historical contingency produced the AI bubble, and it is producing the conditions for that bubble to pop.

Trending Blogs

Last Seen Blogs

mimics inside