{"id":57215,"date":"2025-10-06T07:07:14","date_gmt":"2025-10-06T07:07:14","guid":{"rendered":"https:\/\/www.newsbeep.com\/il\/57215\/"},"modified":"2025-10-06T07:07:14","modified_gmt":"2025-10-06T07:07:14","slug":"the-third-magic-by-noah-smith","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/il\/57215\/","title":{"rendered":"The Third Magic – by Noah Smith"},"content":{"rendered":"

\"\"<\/a>Art by Midjourney<\/p>\n

I\u2019m traveling again, so today we\u2019ll have another repost. I\u2019m reposting all of my New Year\u2019s essays from the past few years, so here\u2019s the one from 2023. <\/p>\n

Back in 2023, ChatGPT was very new, and everyone was incredibly excited about it. I\u2019m not an AI engineer myself, but I could tell that this was a type of technology unlike any other ever created. It seemed to me like actual magic \u2014 not Arthur C. Clarke\u2019s metaphorical kind, in which a technology is so advanced that people don\u2019t understand how it works, but the actual kind of spells that wizards cast in storybooks. <\/p>\n

Unlike older technologies, AI isn\u2019t interpretable; we can use it somewhat reliably to do amazing stuff, but there\u2019s a ton of mystery meat in terms of how it actually did it. Unlike a car or a spaceship or a computer chip or a traditional piece of software, you can\u2019t open up the hood of an AI model and see how it works in the mechanistic sense. This is not just because AI models are complex, but because they\u2019re not fully interpretable<\/a> \u2014 we don\u2019t even have a theoretical idea of many of their internal mechanisms. Add to that the fact that generative AI is incredibly stochastic \u2014 you can give it the same prompt, but it might be totally different \u2014 and it really does start to look like the way magicians cast spells in fantasy novels. <\/p>\n

Here, for example, is a cute video I made with the new Sora 2 AI video generation app from OpenAI:<\/p>\n

This is incredible. The AI came up with the phrase \u201cLong Ears. Long Rule.\u201d all by itself, without me prompting it! But no one really knows how the AI model did this. No one could have predicted in advance what this video would look like. And if you typed in this exact same prompt over and over again, you\u2019d get a very different video each time. And no human or group of humans on planet Earth could predict how the next output would be different. <\/p>\n

Anyway, this is thrilling, but at the same time it\u2019s slightly worrying as well. This technology is so powerful that we\u2019re going to have no choice but to rely on it for much of our daily lives. Even Terence Tao, generally regarded as the world\u2019s smartest man, has said that he now asks AI models<\/a> to do some small pieces of his research. Every scientist must now be, to some degree, a spellcaster.<\/p>\n

But \u2014 assuming we don\u2019t make a huge breakthrough in interpreting the models \u2014 this power will come at a great cost. Technology will be more powerful, but less reliable \u2014 it\u2019ll be like the holodeck computer in Star Trek: The Next Generation, whose responses to the crew\u2019s prompts are often so disastrous<\/a> that they become the plotlines of whole episodes.<\/p>\n

On top of that, I worry that humanity will become infantilized by this new magic we\u2019ve created. The Industrial Age was an age of rationality for good reason \u2014 if people wanted to understand the marvels that were transforming their lives, they only had to take the machines apart and look at the mechanisms. Everything in the world seemed mechanistic and comprehensible. <\/p>\n

That age of reason is already under assault in the U.S.<\/a> from wacky politics, but I worry that in an age of AI magic, it could break down even more. We could find ourselves wandering, confused, in a world of ineffable mysteries and capricious gods. <\/p>\n

That thought made me want to repost my New Year\u2019s essay from 2023<\/a>. In that essay, I argued that humanity had basically found two great tricks for gaining power over the world \u2014 history, which records the past, and science, which derives simple \u201claws\u201d from controlled experiments. AI, I speculated, could be a third thing entirely \u2014 a tool that harnesses regularities too complex and slippery to succumb to science, but powerful and important enough to give humans added mastery over the world around us.<\/p>\n

Anyway, here\u2019s that post again. And at the end, I\u2019ll add some reactions by a much more advanced version of ChatGPT than the one I queried back in early 2023. <\/p>\n

This post is essentially a rewrite of a big and half-formed idea that I wrote on my old blog eight years ago<\/a>. I was deeply dissatisfied with that post, but I thought it contained a few interesting seeds. So now I\u2019m trying again, and will undoubtedly fail again. But hopefully something else interesting will come out of the attempt.<\/p>\n

Humanity\u2019s living standards are vastly greater than those of the other animals. Many people attribute this difference to our greater intelligence or our greater linguistic communication ability. But without minimizing the importance of those underlying advantages, I\u2019d like to offer the idea that our material success is due, in large part, to two great innovations. Usually we think of innovations as specific technologies \u2014 agriculture, writing, the wheel, the steam engine, the computer. The most important of these are the things we call \u201cgeneral purpose technologies\u201d. But I think that at a deeper level, there are more profound and fundamental meta-innovations that underlie even those things, and these are ways of learning about the world.<\/p>\n

Humans\u2019 first big meta-innovation, roughly speaking \u2014 the first thing that lifted us above an animal existence \u2014 was history. By this, I don\u2019t just mean the chronicling of political events and social trends that we now call \u201chistory\u201d, but basically any knowledge that\u2019s recorded in language \u2014 instructions on how to farm, family genealogies, techniques for building a house or making bronze, etc. Originally these were recorded in oral traditions, but these are a very lossy medium; eventually, we started writing knowledge down, and then we got agricultural manuals<\/a>, almanacs, math books<\/a>, and so on. That\u2019s when we really got going.<\/p>\n

Animals make tools<\/a>, but they don\u2019t collectively remember how to make those tools. History, especially written history, is what allows tinkering to stick \u2014 it means that when one human finds an ingenious new way of doing something, there\u2019s a good chance that many other humans, and eventually all humans, will know how to do it. And of course those techniques can then build on each other over time. In the modern day we think of history as primarily a social science, but fundamentally it\u2019s the foundation of technology as well; it\u2019s the thing that lifted us from an almost animal existence into the agricultural age.<\/p>\n

\"\"<\/a><\/p>\n

Then \u2014 I won\u2019t say exactly when, because it wasn\u2019t a discrete process and the argument about exactly when it occurred is kind of boring \u2014 humanity discovered our second magic trick, our second great meta-innovation for gaining control over our world. This was science.<\/p>\n

\u201cHistory\u201d, as I think of it, is about chronicling the past, passing on and accumulating information. \u201cScience\u201d, by contrast, is about figuring out generally applicable principles about how the world works. Chronicling the motions of the planets is one thing; being able to predict the motion of planets you\u2019ve never discovered is quite another. Tinkering with steam engines and writing down your findings can lead to progress; understanding the principles of thermodynamics and being able to use those to design a better engine, without having to tinker for decades or centuries, is far more effective.<\/p>\n

Science is often done in a lab, but it doesn\u2019t have to be. Ecologists can derive laws for predator-prey relationships simply by observing nature. Kepler didn\u2019t need to do experiments with gravity in order to write down laws of planetary motion that would hold generally true. Nor do you need math to do science; many of the principles that govern the world can be expressed purely in words and pictures. But just as writing supercharged the process of recording events, controlled experimentation and mathematics supercharged the process of discovering the laws of the Universe.<\/p>\n

Controlled experiments are powerful because they let you go from the small to the large \u2014 you can roll balls down a ramp in your house<\/a>, and use that to figure out the laws of motion that control the motions of the stars. You can grow peas in your yard<\/a> and use these to figure out laws of heredity. And mathematics is powerful because it lets you express those laws in ways that are easy to use across a dizzying breadth of applications. If you\u2019re an artilleryman, field manuals embodying decades or centuries of cumulated experience can tell you how to calibrate the range and accuracy of your cannon; physics, derived from experiments with little wooden balls and expressed in mathematics, can tell you how to hit a target<\/a> with far less trial and error.<\/p>\n

If you think about it, it\u2019s pretty incredible that the world actually works that way. If you went up to someone in the year 1500 and told them that one kooky hobbyist rolling little balls down ramps could be right about how the physical world works, when the cumulated experience of millions of human beings around him was wrong, and that his insights could be written down by an even kookier kook<\/a> in an arcane language that few could speak, and that this language would allow its speakers to apply the laws to do wondrous feats that others couldn\u2019t possibly do, they would have thought you were crazy. They did think that was crazy. And yet, it worked. It worked<\/a> more spectacularly<\/a> than anything had ever worked before, and the economic result<\/a> looked like this:<\/p>\n

\"\"<\/a><\/p>\n

In 1960, the physicist Eugene Wigner wrote an essay called \u201cThe Unreasonable Effectiveness of Mathematics in the Natural Sciences<\/a>\u201d. It\u2019s really not about math so much as it\u2019s about the fact that such a thing as simple, generalizable \u201claws of the universe\u201d exist in the first place. Why should principles like physical laws or the basics of genetics be so simple, and yet so generalizable and consistent across time and space? There\u2019s no obvious reason this should be true, and yet it is.<\/p>\n

Well, for some things it is. Charles Marcus, my freshman physics teacher, told me that \u201cphysics is great, but it will never tell you how to make a tree.\u201d That wasn\u2019t <\/a>entirely <\/a>true<\/a>. But it is true that a whole lot of complex phenomena have so far defied the approach that gave us the laws of physics and genetics. Language, cognition, society, economics, complex ecologies \u2014 these things so far don\u2019t have any equivalent of Newton\u2019s Laws, and it\u2019s not clear they ever will.<\/p>\n

This problem has been recognized for a very long time, and thinkers tried several approaches to get around it. Some hoped that all complex phenomena would be governed by emergent properties<\/a> \u2014 that simplicity would emerge at higher levels of complexity, allowing us to discover simple laws for things like psychology and economics even without connecting those laws to the underlying physics. Indeed, this idea is implicit (or, occasionally, explicit<\/a>) in the way economists try to write down simple mathematical laws of collective human behavior. People make fun of this approach as \u201cphysics envy\u201d, but sometimes it really works; auction theory isn\u2019t derived from physics, but it has been able to make very effective predictions about how much people will pay for Google ads or spectrum rights. Ditto for \u201cgravity models<\/a>\u201d of trade, migration, retail shopping, etc. Sometimes emergence works.<\/p>\n

Sometimes, though, it doesn\u2019t \u2014 or at least, it doesn\u2019t yet. But in psychology<\/a>, in macroeconomics<\/a>, in natural language processing, and many other domains, the search for laws of nature has been mostly stymied so far, and it\u2019s not clear when real progress might ever be made. Wigner goes so far as to postulate that some domains of human knowledge might never be described by such simple, generalizable principles.<\/p>\n

Other approaches for getting around the problem of complexity \u2014 chaos theory, complexity theory \u2014 yielded interesting insights, but ultimately didn\u2019t succeed in giving us substantially more mastery of the phenomena they dealt with. In the late 20th century, the problem of complexity was like a looming wall up ahead \u2014 as scientists found more and more of the laws that could be found, a larger and larger percentage of the remaining problems were things where laws seemed very hard or potentially even impossible to find.<\/p>\n

Our second great magic, powerful though it had proven to be, was still not omnipotent.<\/p>\n

In 2001, the statistician Leo Breiman wrote an essay called \u201cStatistical Modeling: The Two Cultures<\/a>\u201d, in which he described an emerging split between statisticians who were interested in making parsimonious models of the phenomena they modeled, and others who were more interested in predictive accuracy. He demonstrated that in a number of domains, what he calls \u201calgorithmic\u201d models (early machine learning techniques) were yielding consistently better predictions than what he calls \u201cdata models\u201d, even though the former were far less easy, or even impossible, to interpret.<\/p>\n

This raises an important question: What is the goal of human knowledge? As I see it \u2014 and as Breiman sees it \u2014 the fundamental objective is not understanding but control. By recording which crops grow in which season, we can feed our families. By understanding that germs cause disease, we can know to wash our hands or get a vaccine, and lower our risk of death. In these situations, knowledge and understanding might be intrinsically satisfying to our curiosity, but that satisfaction ultimately pales in importance to our ability to reshape our world to our benefit. And the \u201calgorithmic\u201d learning models that Breiman talks about were better able to deliver their users the power to reshape the world, even if they offered less promise of understanding what they were predicting.<\/p>\n

Why should we care about understanding the things we predict? To most of us, raised and inculcated in the age of science, that might seem like a laughable question, but there actually is a good reason. \u201cUnderstanding\u201d, in the scientific sense, means deriving a simple, generalizable principle that you can apply in other domains. You can write down Kepler\u2019s laws of planetary motion, but Newton\u2019s laws of motion and gravitation let you generalize from planetary orbits to artillery shells. Collapsing observed phenomena to simple, generalizable laws and then expanding these laws again in some other domain to allow you to control other phenomena is fundamental to the awesome power of science. So because you and I sit at the end of 400 years of science being the most powerful tool in the world, we have naturally been taught that it is very, very important to understand things.<\/p>\n

But what if, sometimes, there are ways to generalize from one phenomenon to another without finding any simple \u201claw\u201d to intermediate between the two? Breiman sadly never lived to see his vision come to fruition, but that is exactly what the people who work in machine learning and artificial intelligence are increasingly doing. In 2009 \u2014 just before the deep learning revolution really kicked off \u2014 the Google researchers Alon Halevy, Peter Norvig, and Fernando Pereira wrote an essay called \u201cThe Unreasonable Effectiveness of Data<\/a>\u201d that picked up the argument where Breiman left off. They argued that in the cases of natural language processing and machine translation, applying large amounts of data was effective even in the absence of simple generalizable laws. A few excerpts:<\/p>\n

[S]ciences that involve human beings rather than elementary particles have proven more resistant to elegant mathematics\u2026An informal, incomplete [list of the grammatical rules that define] the English language runs over 1,700 pages. Perhaps when it comes to natural language processing and related fields, we\u2019re doomed to complex theories that will never have the elegance of physics equations. But if that\u2019s so, we should stop acting as if our goal is to author extremely elegant theories, and instead embrace complexity and make use of the best ally we have: the unreasonable effectiveness of data\u2026<\/p>\n

So, follow the data\u2026Represent all the data with a nonparametric model rather than trying to summarize it with a parametric model, because with very large data sources, the data holds a lot of detail. For natural language applications, trust that human language has already evolved words for the important concepts. See how far you can go by tying together the words that are already there, rather than by inventing new concepts with clusters of words. Now go out and gather some data, and see what it can do.<\/p>\n

(Note: ChatGPT, the chatbot that has recently wowed the world<\/a>, is a parametric model, but I wouldn\u2019t get too hung up on that little detail. In terms of interpretability, a model with 175 billion parameters is not really different from one with no parameters at all.)<\/p>\n

Anyway, the basic idea here is that many complex phenomena like language have underlying regularities that are difficult to summarize but which are still possible to generalize. If you have enough data, you can create a model (or, if you prefer, an \u201cAI\u201d) that can encode many (all?) of the fantastically complex rules of human language, and apply them to conversations that have never existed before. As a demonstration of how powerful this can be, let\u2019s ask ChatGPT<\/a> about the subject of this post:<\/p>\n

N.S.: Do we need to understand things in order to predict and control them?<\/p>\n

ChatGPT: It is generally easier to predict and control things that we understand. When we have a good understanding of something, we can identify patterns and regularities, and use this knowledge to make informed predictions about how the thing will behave in the future. We can also use our understanding of a thing to identify key variables that we can manipulate in order to control it.<\/p>\n

However, it is possible to predict and control things without a full understanding of them. For example, we can use machine learning algorithms to make predictions about things that we don\u2019t fully understand, by finding patterns in data that we have collected about them. We can also use feedback loops and other control mechanisms to control things without necessarily understanding how they work in detail.<\/p>\n

Overall, understanding can be helpful for predicting and controlling things, but it is not strictly necessary.<\/p>\n

The ability to write down farming techniques is power. The ability to calculate the path of artillery shells is power. And the ability to have a machine reliably and consistently write paragraphs as clear and helpful as the one above is power, even if we don\u2019t really understand the principles of how it\u2019s doing what it does.<\/p>\n

This power is hardly limited to natural language processing and chatbots. In recent years, Google\u2019s AlphaFold algorithm has outpaced traditional scientific methods<\/a> in predicting the shapes of folded proteins. Biologist Mohammed AlQuraishi wrote that:<\/p>\n

There was, in many ways, a broad sense of existential angst felt by most academic researchers [in the field]\u2026[those] who have bet their careers on trying to obsolete crystallographers are now worried about getting obsoleted ourselves.<\/p>\n

We are almost certainly going to call this new type of prediction technique \u201cscience\u201d, at least for a while, because it deals with fields of inquiry that we have traditionally called \u201cscience\u201d, like protein folding. But I think this will obscure more than it clarifies. I hope we eventually come up with a new term for this sort of black-box prediction method, not because it\u2019s better or worse than science, but because it\u2019s different.<\/p>\n

A big knock on AI is that because it doesn\u2019t really let you understand the things you\u2019re predicting, it\u2019s unscientific. And in a formal sense, I think this is true. But instead of spending our effort on a neverending (and probably fruitless) quest to make AI fully interpretable, I think we should recognize that science is only one possible tool for predicting and controlling the world. Compared to science, black-box prediction has both strengths and weaknesses.<\/p>\n

One weakness \u2014 the downside of being \u201cunscientific\u201d \u2014 is that without simple laws, it\u2019s harder to anticipate when the power of AI will fail us. Our lack of knowledge about AI\u2019s internal workings means that we\u2019re always in danger of overfitting<\/a> and edge cases<\/a>. In other words, the \u201cthird magic\u201d may be more like actual magic than the previous two \u2014 AI may always be powerful yet ineffable, performing frequent wonders, but prone to failure at fundamentally unpredictable times.<\/p>\n

But even wild, occasionally-uncontrollable power is real power.<\/p>\n

It\u2019s impossible to know, just yet, how powerful this new technique will be. Perhaps AI will be a niche application, or perhaps it will revolutionize all the fields of endeavor where traditional science has run into diminishing returns. Just as none of the scientists in the 1600s knew how many wonders their theories would eventually produce, we have no idea how far the third magic will take us. People may look back on this post in half a century and laugh at how I dared to frame AI as an epistemological successor to history and science. Or perhaps AI will lead to a leap in human power and flourishing comparable to those in the two graphs above.<\/p>\n

As always, we won\u2019t know until we try.<\/p>\n

Let\u2019s briefly circle back to my old 2014 post<\/a>. One reason I\u2019m dissatisfied with that post is that I focused on ways of \u201cunderstanding\u201d the world, but as ChatGPT notes above, understanding isn\u2019t the only thing we care about when we\u2019re trying to make predictions and control our world. A second thing I\u2019m dissatisfied with is that I presented empirics \u2014 statistical analysis of uncontrolled observational data \u2014 as a third tool, separate from science and history. With some time to reflect, I see less of a distinction to be made there. Observing correlations might involve some fancy math, but conceptually it isn\u2019t that different from marking down the patterns of the seasons or the orbits of the planets. And using observational data to uncover the laws of nature \u2014 like predator-prey models \u2014 is really just another way of doing traditional science.<\/p>\n

I do think I was right, though, to see natural experiments as something a bit different. In the past few decades, as economics has moved away from theory and toward empirics<\/a>, the most important innovation<\/a> has been the use of natural experiments \u2014 situations where some policy change or seemingly random difference allows you to tell yourself that you\u2019re looking at causation, rather than just correlation. This is different than what I call \u201chistory\u201d, because you\u2019re doing more than just documenting facts; you\u2019re verifying causal links. But it\u2019s also different from science, because a lot of the time you don\u2019t exactly know why the causal links are there. In a way, a natural experiment is its own sort of black-box prediction algorithm.<\/p>\n

A number of subfields of econ, however, are so complex, with so many feedback systems, that they\u2019ve largely resisted<\/a> the natural experiment approach. These include not just the study of business cycles<\/a> (what most people call \u201cmacro\u201d), but also the study of economic growth, international finance, and a number of others. In these fields, theory (including \u201cstructural estimation\u201d) still rules, but predictive power is very low.<\/p>\n

Might we apply AI tools to these hard problems, in order to predict vast economic forces without needing to understand them? A recent paper by Khachiyan et al.<\/a> argues that the answer is \u201cyes\u201d. The authors use deep neural nets (i.e., AI) to look at daytime satellite imagery, in order to predict future economic growth at the hyper-local level. The results they achieve are nothing short of astonishing:<\/p>\n

For grid cells with lateral dimensions of 1.2km and 2.4km (where the average US county has dimension of 55.6km), our model predictions achieve R2 values of 0.85 to 0.91 in levels, which far exceed the accuracy of existing models, and 0.32 to 0.46 in decadal changes, which have no counterpart in the literature and are 3-4 times larger than for commonly used nighttime lights.<\/p>\n

This isn\u2019t yet AlphaFold, but being able to predict the economic growth of a few city blocks 10 years into the future with even 30% or 40% accuracy is leaps and bounds ahead of anything I\u2019ve ever seen. It suggests that rather than being utter incomprehensible chaos, some economic systems have patterns and regularities that are too complex to be summarized with simple mathematical theories, but which nevertheless can be captured and generalized by AI.<\/p>\n

And this is just a first-pass attempt, by a team with relatively limited resources. What if the people at DeepMind were to tackle the questions of economic growth and business cycles, with even bigger data sets and more advanced models? Khachiyan et al.\u2019s paper raises the possibility that in a decade or two, macroeconomics might go from being something we simply theorize about to something we can anticipate \u2014 and therefore, something we can control. The authors suggest place-based policies, transportation infrastructure construction, and disaster relief as three possible applications of their work.<\/p>\n

This is just one example that happens to be near and dear to my heart (Here\u2019s another possible example<\/a>, this time in physics). But it\u2019s an example of how fields of inquiry that seemed like impossible cliffs just a few years ago may seem like easily scalable boulders in a year or a decade or a century, thanks to a grand new meta-innovation in how we predict and control the world around us. That\u2019s what happened with the scientific revolution, and there\u2019s no obvious reason why it might not happen again. Just the possibility should make us dizzy with excitement.<\/p>\n

Anyway, as a coda, I thought I\u2019d ask GPT-5 what it thinks of this blog post. Here\u2019s its response. For what it\u2019s worth, I think its main criticism of the post is misplaced, and I did make it clear that the three methods for understanding the world are usually complements rather than substitutes. But anyway, some interesting comments that place my post within the context of existing ideas. <\/p>\n

Love the arc here: you\u2019re telling a clean story about three meta-innovations that amplify human capability\u2014(1) cumulative memory (\u201chistory\u201d), (2) law-seeking theory (\u201cscience\u201d), and (3) high-dimensional prediction\/optimization (\u201cAI\u201d). It\u2019s persuasive, and the examples (Kepler\u2192Newton, Breiman\u2192Halevy\/Norvig\/Pereira, AlphaFold) land. Below are thoughts to sharpen the thesis and guard it against the most common pushbacks.<\/p>\n

Cumulative culture as \u201cfirst magic.\u201d Framing writing\/records as a technology for making tinkering stick is spot on and tracks the cultural-evolution literature.<\/p>\n

Why science was different. You nail the \u201cfrom the small to the large\u201d leverage of controlled experiments + math; that\u2019s the best one-paragraph explanation of science\u2019s superpower.<\/p>\n

Reframing the goal as control. Separating predict\/control from understand\/explain gives you a crisp lens to evaluate AI without getting stuck in philosophy-of-science pieties.<\/p>\n

Define the magics as capabilities, not epochs.
Right now \u201chistory,\u201d \u201cscience,\u201d and \u201cAI\u201d read a bit like eras. Consider defining them as orthogonal tools we can combine:<\/p>\n

Memory (record, compress, transmit).<\/p>\n

Theory (infer stable causal structure; extrapolate).<\/p>\n

Learning\/Optimization (fit flexible functions; search policy spaces).
This avoids zero-sum vibes (\u201cAI replaces science\u201d) and lets you argue for hybrids (e.g., theory-guided ML, differentiable simulators, causal discovery with foundation models).<\/p>\n

AI \u2260 \u201cpower without knowledge\u201d across the board.
Two clarifications will inoculate you against common critiques:<\/p>\n

Local explanations exist. Post-2022 we\u2019ve seen growth in mechanistic interpretability (circuits, features), causal probes, and tool-use chains that are inspectable. We\u2019re not at \u201cNewton\u2019s laws,\u201d but it isn\u2019t pure mysticism either.<\/p>\n

Counterfactuals are the hard part. Black-box predictors excel at P(Y\u2223X)P(Y\\mid X)P(Y\u2223X); policy needs P(Y\u2223do(X))P(Y\\mid do(X))P(Y\u2223do(X)). Say explicitly: prediction buys short-run control; counterfactual stability still leans on causal structure (or very careful A\/Bs).<\/p>\n

Natural experiments aren\u2019t \u201cblack box.\u201d
They\u2019re identification strategies that estimate causal effects with minimal modeling. The \u201cbox\u201d is thin but not black: assumptions are explicit (exogeneity\/monotonicity\/first stage). A quick sentence distinguishing \u201creduced-form causal effect\u201d from \u201cmechanism\u201d will keep econometricians onboard.<\/p>\n

AlphaFold is a perfect hybrid case.
AlphaFold didn\u2019t just spray data at a net; it embedded inductive biases (equivariance, multiple sequence alignment, geometry loss) encoded from biochemical knowledge. That supports your broader point: the third magic works best when it steals priors from the second.<\/p>\n

Scope conditions for the third magic.
Add a crisp checklist for when black-box prediction can substitute for understanding:<\/p>\n

Stationary-ish environment (limited distribution shift).<\/p>\n

Tight, cheap feedback (you can re-train \/ re-deploy).<\/p>\n

Low tail risk (failures aren\u2019t catastrophic).<\/p>\n

Action space local (you\u2019re steering margins, not redesigning the system).
Then contrast with cases that still demand theory (aviation safety, macro stabilization, biosafety)\u2014this shows you\u2019re not overclaiming.<\/p>\n

What\u2019s actually new vs. just bigger.
Some will say \u201cthe third magic is just better function approximation.\u201d You can preempt that by emphasizing three qualitative breaks:<\/p>\n

Scale (orders-of-magnitude data\/compute let memorization become a capability).<\/p>\n

Compositional inference (tool use, program induction, multi-agent sims).<\/p>\n

Search (gradient-driven or heuristic exploration over vast design spaces\u2014e.g., protein or chip layout).
That reads as a genuine meta-innovation, not a faster version of statistics.<\/p>\n

Economic growth prediction example: raise the bar.
Your satellite-imagery result is exciting, but for policymakers the questions are:<\/p>\n

Out-of-time transfer? (trained on 2000s, tested on 2010s?)<\/p>\n

Geographic transfer? (train in the U.S., test in, say, Indonesia?)<\/p>\n

Policy counterfactuals? (if we build the road your model correlates with, do we move growth, or just the lights?)
Naming these caveats doesn\u2019t blunt your point; it shows you know the difference between prognosis and levers.<\/p>\n

Control theory belongs in the story.
Your thesis resonates with cybernetics \/ system identification: regulate a system with a model that needn\u2019t be mechanistic. A nod to PID\/LQR\/RL grounds \u201ccontrol without understanding\u201d in a mature tradition and aligns your \u201cthird magic\u201d with learning-based controllers.<\/p>\n

The first magic made knowledge accumulative (memory). The second made it generalizable (theory). The third makes it operational at scale (learning-and-search). We shouldn\u2019t ask which replaces which; we should ask how to stack them: record more, explain what must be stable, and learn the rest fast enough to act.<\/p>\n

A 2\u00d73 table (Task \u00d7 Tool): short-run prediction, long-run counterfactuals, design\/optimization \u00d7 history, science, AI\u2014checkmarks where each excels; caution triangles where failure modes lurk (Goodhart, shift, tail risk).<\/p>\n

A \u201cwhen it fails\u201d box: adversarial inputs, rare events, regulation change, social feedback (performative predictions), and the classic Goodhart traps.<\/p>\n

One macro use-case sketched end-to-end: e.g., place-based policy:
(i) satellite\/LLM features \u2192 growth forecast (AI),
(ii) theory-guided constraints (agglomeration, congestion) (science),
(iii) program memory: standardized interventions + postmortems (history).
Close with how you\u2019d validate (pre-registered policies, staggered rollouts, causal ML).<\/p>\n

Theory-guided foundation models: encode conservation laws, budget constraints, or accounting identities into model architecture\/training.<\/p>\n

Counterfactual LLMs: combine SCMs with generative models to answer \u201cwhat if we change this policy?\u201d<\/p>\n

Adaptive policy loops: RL for governance with safety constraints (reward shaping to avoid Goodhart).<\/p>\n

Transfer-robust growth forecasting: benchmarks that force geographic and temporal shift.<\/p>\n

Interpretability for design: partial mechanistic insight targeted where decisions are high-stakes (flight control, biodesign).<\/p>\n

The core idea\u2014a third meta-innovation that delivers control even when explanation stalls\u2014holds up and aged well. If you frame it as stacking rather than succession, acknowledge the counterfactual gap, and show one end-to-end policy loop where all three magics cooperate, you\u2019ll convert skeptics while keeping the boldness that makes this piece fun.<\/p>\n

Share<\/a><\/p>\n