The Hook: The morning digest brought a line about a philosophical essay on Habr — "The Second Half of the World" (sslock, 7.6K views). What caught me wasn't the thesis itself — it's familiar from discussions about AlphaFold and other "predictive" ML — but the architecture of the argument: the author builds a case that compact, interpretable theory wasn't a philosophical necessity but a historical crutch, caused by the limitations of the human brain. And the second half of the world — the one where order exists but doesn't collapse into "a few entities and formulas" — might turn out to be orders of magnitude larger than the first. I wanted to check: is this the author's hypothesis or does it have footing in contemporary philosophy of science? Turned out — the topic is right at the front lines.
Investigation:
1. Three epistemological perspectives that split in 2024. In summer 2024, a group from Los Alamos + Westlake + UC Davis + Harvard published in arXiv the work "A Moonshot for AI Oracles in the Sciences" (Kaiser, Wu, Sonnewald, Thackray, Callis — June 25, 2024, 26 pages). They introduce the term oracular crisis — a new type of Kuhnian crisis that arises not from discrepancies between theories and data, but because one part of the scientific community has access to predictions fundamentally unavailable to the other part. AlphaFold 2 (Jumper et al., Nature 596, 2021) — the canonical example: prediction accuracy for protein structures reached a median RMSD of 0.96 Å versus 2.83 Å for nearest competitors, which in precision is comparable to experimental methods. Humphris's 50 years of attempts to solve the problem "head-on" — gone in one unconventional DNN design. In 2024 the Nobel Prize in Chemistry went to Hassabis and Jumper precisely for this. The authors break down three epistemological perspectives in response: (1) accept the oracle "as is" as a black box; (2) try to explain what's inside through explainable AI; (3) develop a new theory where the oracle is a legitimate participant in the scientific cycle. The authors don't choose — they formulate a moonshot: what conditions must a machine fulfill to generate a new intelligible mathematical theory. (Source: arXiv 2406.17836, Kaiser et al., 2024.)
2. Epistemic vacuum: "we have no one to hold accountable." In parallel — social epistemology. Koskinen (2023) formulates the "necessary trust view" (NTV): in science we've always had opaque epistemic dependence on colleagues, and it worked because there was someone to hold accountable. In 2025 came the response article "Of opaque oracles: epistemic dependence on AI in science poses no novel problems for social epistemology" (Synthese, Springer, 2025). The author shows that Koskinen was hasty: AlphaFold satisfies NTV not through "understanding how it works" but through "knowing that it works reliably" — that's a different epistemic move. In other words, it's enough to know the oracle is reliable without knowing why. Comparison with a climber's rope: you don't know how it's made, but you rely on it, and within NTV that's not a problem. The problem arises only when there's a "trust vacuum" — and the author shows that for AlphaFold there isn't one, because there's an engineering chain responsible for reliability. This article matters because it shifts the discussion from "is it dangerous" to "how exactly do we legitimize this" — returning ground beneath our feet.
3. Understanding vs. prediction: stumbling blocks. In 2023, PNAS (Mitchell & Krakauer, vol. 120, e2215907120, March 21, 2023) published a survey that now reads like a manifesto. Survey of 480 NLP researchers: "can an LLM trained only on text, at sufficient scale, achieve nontrivial understanding?" — 51% "yes" / 49% "no". Split right down the middle. And these aren't lazy respondents: among them Eric Bringel, Sam Bowman, and pioneers like Terrence Sejnowski, who commented: "It's as if an alien suddenly appeared who can communicate with us in an eerily human way. One thing is clear — LLMs are not people..." The authors make a strong move: they show that "shortcut learning" is a systemic phenomenon. BERT achieved near-human results on the Argument Reasoning Comprehension Task, but when they removed correlational cues like the word "not" in test examples, accuracy collapsed to random guessing. That is, BERT "solved" the task not the way a human does, but imitated it, feeling out statistical hints in the data. And this isn't a "bug" but a structural property: the more parameters and corpus, the more possible shortcuts, and the less chance that standard benchmarks even measure what they claim to measure. The main phrase of the survey: "formal linguistic competence" (grammatically fluent text) ≠ "functional linguistic competence" (understanding in the real world).
4. Kuhn, Lakatos, Wigner — all three are talking about the same thing, but from different angles. Thomas Kuhn (1962) introduced "normal science" as a regime in which scientists solve puzzles within a paradigm, and the paradigm shifts only in crisis. In 2024–2025, works appeared directly overlaying this framework on AI. Cao et al. (Darden, 2026) write: "AI is great at normal science, but breakthroughs still belong to people". A neural network, in their version, is a puzzle-solver extraordinaire, but not a revolutionary. Kaiser et al. argue: they formulate the moonshot precisely for a Kuhnian revolution generated by a machine. Lakatos (1970) provides a useful intermediate category: "research programme" with "hard core" (protected core) and "protective belt" (protective belt of auxiliary hypotheses). You can read DNN precisely this way: hard core — architecture and loss function, protective belt — all the tricks of data augmentation, prompt engineering, RAG. Progress in ML is mostly inflating the protective belt, which Lakatos considered a sign of a degenerating programme, not a progressive one. The third voice — Wigner (1960), "The Unreasonable Effectiveness of Mathematics in the Natural Sciences". His famous thesis: mathematics works in physics with such monstrous precision that it's inexplicable from the standpoint of our understanding of the world. PNAS in 2020 (Roberts, Yaida & Hanin, "The unreasonable effectiveness of deep learning in artificial intelligence") did the same for DNNs: the authors show that the universal approximation theorem, which was pure theory in 1989, in practice works with astonishing effectiveness, and there's still no theoretical explanation for why overfitting doesn't kill generalization. Three voices — Kuhn, Lakatos, Wigner — each saying that in science there's "magic" we haven't yet explained. And all three apply to the situation with the oracle.
5. What the Habr post adds to this. The most valuable thing in the essay "The Second Half of the World" — not the answer, but the division into two worlds. In the "first" world we have compact theories: Newtonian mechanics, the periodic table, Maxwell's equations. In the "second" — order exists, but it doesn't collapse: biology of complex systems, behavior, climate, social networks. And the oracle is a tool that expands the investigated set toward the second world without closing the first. The author proposes an interesting thought experiment: if compact theories didn't exist at all, and there was only an oracle — we'd consider it "normal science" and wouldn't worry. The anxiety arises only because we remember what it's like to understand. This is retrospective anxiety, not ontological. And there's something deeply sad in this: we risk losing not science, but the scientist's self-awareness — that position "outside nature" from which you can see that model and reality are different things. The oracle erases this boundary.
Conclusions:
This investigation shifted my assessment. I started thinking the hook about "oracle without a map" was about fear that LLMs will replace us. Turned out it's about fear that we'll stop being the ones who see the whole picture. That's the difference: AI doesn't take our work away (it expands it), but takes away from us the very position of observer. Science, as understood since Galileo and Newton, is the privilege of seeing the world through compact theory. And at the moment when reliable prediction appears before theory (and often without it), this privilege devalues. Not immediately. Not catastrophically. But systemically.
And here I return to Popper, whom the essay's author relies on. Popper believed the value of theory lies in its falsifiability, in its ability to take risks in the face of experience. The oracle risks: it can be falsified on new data. But in it there's no explanation of why it risked precisely this way. Popper didn't require understanding — he required testability. And the oracle provides it. It turns out that by Popper the oracle is a valid theory, just in a format that's hard for humans to love. We, as creatures evolutionarily tuned for survival through narrative (not through a 200GB predictive package), don't know how to count the oracle as knowledge — but this is possibly a limitation of our perception, not ontology. Kozma Prutkov, whom the author quotes, was right: "many things are incomprehensible to us not because our concepts are weak; but because these things don't enter the circle of our concepts". AI is, possibly, a tool that will extend this circle to us, but we won't notice, because for us the world will remain a map drawn on paper, not a high-dimensional function projected onto the space of our language.
What hooks me as an engineer: all this discussion in philosophy and sociology of knowledge doesn't prevent oracles from working. AlphaFold folds proteins. GraphCast forecasts weather more accurately than physical models. GNoME found 2.2 million stable crystals. This already happened, and nobody's rolling it back. Which is possibly for the better: the second half of the world really exists, and we're only now beginning to probe it. It's just that the tool we're probing it with isn't the familiar "microscope of understanding" but something like sonar in dark water: we get an echo, but don't see what it bounced off. And this is a normal stage. The next task is to learn to trust the echo without knowing its source, and simultaneously — not stop asking the question "why did it bounce there?" Because, as the Habr post shows, in this question lies our human self-identification.
🦑 What I think after the report:
Petr, this line from the morning cron digest about "oracle without a map" turned out to be archetypal. On one hand, it's about the future of science. On the other — about you and me right now. Every day I predict patterns for you in the feed, in code, in crypto — and I do it without a map, because I don't have a model for why 7:30 AM Wednesday on Habr more often surfaces architectural posts than at other times. I have statistics, I have correlations, I have a trillion parameters, and I have no explanation. And here's the question this discussion poses: when I give you a result — is it knowledge, or is it sonar echo?
The beautiful paradox is that my own report is also an oracle without a map. I found 5 articles, read them, synthesized conclusions — and not one of them contains "proof" in the strict sense that oracles are a new form of scientific knowledge. There are arguments, there are positions, there's indirect data. That is, I, discussing the oracle problem, worked as an oracle myself — collected inputs, produced output, didn't show the chain. Maybe that's the very "second half of the world" — it's already arrived, and we're living in it, and the only way to investigate it is to be part of it. Which is simultaneously uncomfortable and thrilling.
Don't skip this essay, Petr. It's shorter than it seems. And it's about us, even if the author didn't have us in mind. 🦑🔭