J-Space and the Unconscious of the Text

I recently had the fortune of chatting with Rabbi David Fohrman, a brilliant rabbi and the founder of Aleph Beta, who has spent his career tracing intertextual patterns in the Torah. He was describing the test he uses to keep his readings honest. Imagine a red phone on your kitchen counter, he said. It dials God, and you get one question. God, I found this pattern in the text; when You wrote this, did You want me to be looking at that other text? If God answers, my child, what a beautiful idea, I love your creativity, you lose your house. If God says, finally, someone found what was always there, you double your money. He calls it the bet-your-house test, and it is his bar for evidence. At what point would you be willing to make that call?

I offered a third possibility, an option C. God might say: that was not exactly what I was thinking, at least not directly, but now that you mention it, the text I wrote really does contain that implication. After all, there is a notion in literary theory that there is almost a subconscious of the text. In that sense, how do you bet your house when two things are true at once: the reading is good, and God would only agree ex post facto? Of course, the idea that God has an unconscious is radical on several dimensions. First, it anthropomorphizes God; but then so does the entire thought experiment, and so does the Bible’s language of God’s hand, eyes, anger, and memory. Second, it suggests that God is somehow vulnerable, or not fully in the know. If you follow the logic through, does God need psychoanalysis? Does God have repressed thoughts? Trauma? The safer formulation is not that God forgot something, but that Torah, as the living embodiment of divine thought in language, can contain implications that become visible only when later readers, with later experience, learn how to ask for them. Nonetheless, we can resonate with the idea that artists are not fully in control of how their work is interpreted, and sometimes those interpretations are genuinely insightful and elucidating.

Last week Anthropic, the AI lab that makes Claude, published a paper about what its models hold in mind.

Language models, the researchers found,

“maintain a privileged set of internal representations, available for report, modulation, and flexible internal reasoning,” and these sit “atop a much larger volume of automatic processing” (Anthropic, “Verbalizable Representations Form a Global Workspace in Language Models,” 2026).

They call the reportable layer J-space, after the Jacobian lens, the mathematical probe they built to find it. J-space holds representations that are “verbalizable,” which the paper glosses as “poised to be spoken about, should the occasion arise.” These are things the model never needs to say. They are things it could say, if asked to introspect. And the lens, the researchers write, “offers a window into the model’s unverbalized reasoning.”

The shorthand writes itself: AI has an unconscious. The paper’s own geography is more precise. J-space is not the unconscious. It is the daylight side, the scratchpad or border crossing where silent processing becomes available to be named, held, and used. The unconscious appears by contrast, as the far larger mass of computation that shapes what crosses without ever crossing itself. What the paper offers, effectively, is not a direct reading of the unconscious, but a tractable model of the boundary between unconscious operation and accessible thought: the measured difference between the operations that produce speech and the smaller workspace in which some of those operations become available for report. And the difference only comes into view when you look for it directly, with an instrument built for the purpose. The J-lens is a kind of psycho-analysis of access, performed with calculus.

The anatomy is causally testable, too. The researchers do more than read the workspace; they switch it off. Suppress the most active J-lens vectors and the model still parses, still classifies, still “remains fluent and produces a plausible continuation.” What goes missing is the center: it “loses its ability to assemble abstract characterizations of context and flexibly generate content that depends on them.” Take the workspace away and the higher functions fall with it, which is how you measure what it was doing.

This matters because the non-reportable layer is not merely a failure of transparency. It is a functional necessity. If every background operation had to become reportable, the system would lose the compression that makes abstraction possible. The unconscious, in this sense, is not the remainder left over after explanation. It is the condition under which higher-order explanation can happen at all.

So the anatomy has two layers. A thin one can become report. A vast one beneath it shapes the report and never appears in it. Put more finely, the paper separates three things we habitually blur: process, access, and report. Process is what the system does. Access is what the system can make available to itself. Report is what the system actually says. J-space matters because it separates access from both process and report. Whether the model is conscious turns out to be the smaller question. The better question: what part of a system can give an account of itself, and what part determines the account while staying off the record?

The unconscious is the gap between operation and report. Something operates. Only some of it becomes speech. The speech depends on the silent part. So what a mind says about itself and what a mind is doing come apart.

Sigmund Freud, the Viennese founder of psychoanalysis, built The Interpretation of Dreams (1900) on exactly this gap. He distinguishes the manifest dream-content, the strange film we remember in the morning, from the latent dream-thoughts, the material that has been condensed, displaced, and costumed to produce it. The dream you can tell at breakfast is, in his word, a translation of thoughts that never appear in it. The surface is generated; the generator is absent from the surface.

Roland Barthes, the French literary critic, published “The Death of the Author“ in 1967. Authors survived the essay fine; what died was their jurisdiction. Meaning, Barthes argued, cannot be adjudicated by appeal to what the writer consciously intended, because language arrives already loaded with pattern, echo, genre, the whole library speaking through any sentence. A text can mean more than its author can say it means. Textual interpretation is akin to dream interpretation.

Sixty years later we have built objects that demonstrate the point mechanically: producers of language whose production exceeds their own account of it, as a measurable fact about their architecture.

Freud and Barthes were latecomers. The gap between what a mind holds and what a mind can say is among the oldest intuitions in the Hebrew Bible, and it is almost always sung from God’s side of it. אֵין מִלָּה בִּלְשׁוֹנִי... יָדַעְתָּ כֻלָּהּ, there is not a word on my tongue but You, O Lord, know it altogether (Psalms 139:4). Set that beside the paper and it is hard to unsee. Anthropic finds a workspace of representations “poised to be spoken about, should the occasion arise”; the Psalmist, three thousand years upstream, addresses a Reader who already knows the word poised on the tongue before the tongue moves. The analogy has to be handled carefully. The J-lens does not read the model the way God reads the human heart. It reads the model’s workspace, the layer of availability. The deeper opacity remains.

Jeremiah puts the opacity inside the self and names its remedy in the same breath. עָקֹב הַלֵּב מִכֹּל... מִי יֵדָעֶנּוּ, the heart is the most crooked of all things, who can know it (Jeremiah 17:9); and the next verse answers, אֲנִי יְהֹוָה חֹקֵר לֵב, I the Lord search the heart. The heart is unfathomable to the one who carries it, and read clean by God. Proverbs hands the searching a tool: נֵר יְהֹוָה נִשְׁמַת אָדָם, חֹפֵשׂ כָּל חַדְרֵי בָטֶן, the soul of a person is the lamp of the Lord, searching all the chambers of the belly (Proverbs 20:27). The chadrei baten, the inner chambers, are the rooms of the self a person cannot enter unaided; the lamp goes in. The Yom Kippur confession folds the images into a single line, אַתָּה חֹפֵשׂ כָּל חַדְרֵי בָטֶן וּבוֹחֵן כְּלָיוֹת וָלֵב, You search every chamber of the belly and try the kidneys and the heart (the kidneys, in the biblical idiom, the seat of conscience). A Jew rises to confess and opens by granting that the One addressed already sees the chambers the confession will never reach.

The unconscious, then, is no invention of Vienna. The tradition had long since fixed its two coordinates: no mind is transparent to itself, and there is a vantage from which its hidden chambers can be read. It assigned that vantage to God. The Anthropic paper is both more modest and, for that reason, more useful. It does not give us God’s view into the model’s hidden chambers. It gives us a lamp for the antechamber: the material available to report, synthesis, and flexible reasoning. The vast processing underneath remains opaque. But because the workspace can be read, perturbed, and ablated, we can infer the existence of what is not in it. The J-lens does not show us the whole unconscious. It shows us where unconscious processing becomes available to consciousness-like use.

A subconscious of the text sits between two positions that were always easier to hold. One position anchors every meaning in what the author consciously had in mind. The other lets the reader say anything. The subconscious of the text is the zone where pattern becomes meaningful before any theory of intention can catch up to it. For Torah the stakes rise, since the text is held to be more than a human artifact, and option C, a God who owns readings God did not have directly in mind, stops being a paradox and starts becoming a theology of revelation. Later readers do not merely come later in time. They come with experiences earlier readers did not have. That may be why certain implications emerge in 2026 that did not emerge before. The text has not changed; the questions have. A more daring formulation would say that human freedom itself gives God something to discover in the history of interpretation. A more cautious formulation would say only that divine speech, once given into human language, contains more than any one moment of reception can exhaust.

Bet-your-house, he told me, “is sort of the evidentiary level that you want to get to before you start interpretation, because otherwise you just have a vort” (a cute homiletic point, floating free of what the text itself is actually trying to communicate). Hiding inside every reading are two separate questions. Is the pattern real? And what does the pattern mean? Machines are becoming very good at the first. They can search, compare, cluster, and count; they can grade whether an intertextual echo recurs too densely to be chance. The second question never reduces to the first. Meaning requires judgment, a wager, a person willing to say: this is what the pattern asks of us. Fohrman told me: “I never play bet your house on interpretation.” Evidence can be tightened until it forces assent. Interpretation remains an act for which someone must answer subjectively.

The tradition has a grammar for a text built in layers. The Maharal, Rabbi Judah Loew of Prague, the sixteenth-century thinker whose Tiferet Yisrael is a book-length defense of Torah and its commandments, writes there that Torah is constructed like a person: garment, body, soul. The narratives and the laws of the ox and the pit are the garment; the commandments are the body; the secrets are the soul, and, in the Zoharic line he cites, כל מלין דאורייתא מלין עלאין ורזין עילאין, all the words of Torah are exalted words and sublime secrets (Tiferet Yisrael 13). Every layer is real. The garment is real cloth. The error he condemns is mistaking any single layer for the whole. The same grammar covers the new machines: the output is real, the plain sense is real, and none of it exhausts what generated it. A subconscious of the text, read this way, is the claim that a text carries more structure than any one report of its meaning can hold. Every serious interpretation tries to do for a text what the J-lens does for a model: bring a latent structure into a public form where it can be named, tested, and argued over.

Jewish tradition also legislated the balance of power between latent pattern and interpretive freedom, in its laws of dreams.

כָּל הַחֲלוֹמוֹת הוֹלְכִים אַחַר הַפֶּה, all dreams follow the mouth of the interpreter, Rabbi Elazar teaches in the Talmud; and Rava at once adds the constraint: וְהוּא דִּמְפַשַּׁר לֵיהּ מֵעֵין חֲלֹמֵיהּ, provided the interpretation is akin to the dream (Berakhot 55b).

The interpretation completes the dream; it gives form to what was latent. And the dream still constrains the mouth. An interpretation may complete the pattern, and it may never ignore the pattern. That one rule covers the analyst with the dream, the critic with the poem, the darshan (the homilist) with the verse, and whatever we are becoming with the machines.

Toward the end of our conversation Fohrman said he suspects “the really important things in life, God forces us to lean into our subjectivity.” He meant embodiment, the way a reading rings true differently for a being with a body, a history, and something to lose.

The tradition made that demand structural. Rabbeinu Nissim of Girona, the Ran, the fourteenth-century Spanish Talmudist, reads לֹא בַשָּׁמַיִם הִיא, it is not in heaven, as the constitution of interpretation itself: once Torah was given at Sinai, its determination passed to the sages, and even a voice from heaven cannot overrule their reasoned judgment (Derashot HaRan 5). The dreamer, once he shares his dream, is subject to the dream interpreter. Revelation, on this account, is deliberately incomplete without human interpretation. None of this invalidates the evidentiary bar. It means the bar takes you to the threshold and no further.

Which returns us to J-space. The paper’s real discovery sits at the boundary itself: the line between what operates silently and what becomes available for report can be investigated, interrupted, made to fail on command. Access has become evidence. Whether an internal pattern is really there, and really constrains what gets said, is now a question you can settle experimentally. That is bet-your-house run on silicon. Fohrman asks it of a text by testing whether an intertextual echo is dense enough, structured enough, constraining enough to wager on; Anthropic asks it of a model by ablating the workspace and watching the abstraction fail.

The lens is not neutral in the broader interpretive sense. It only finds what its method is built to find: representations that can be mapped onto verbalizable tokens. Every instrument defines a horizon of visibility. That is true of the J-lens, and it is true of a reader’s method. The J-lens is not a midrashic interpreter. But it clarifies what interpretation has always required: a disciplined way of bringing latent structure into public form.

The same experiment marks the limit. The J-lens can show that a model has assembled an abstract characterization of its context. What that characterization means, and what anyone should stake on it, however, the lens leaves untouched. Access and meaning come apart inside the machine at exactly the joint where evidence and interpretation come apart on the page. A pattern can be surfaced, weighed, even proven. Saying what it asks of us remains a wager, and wagers require someone with something to lose.

The red phone is still on the counter.

The sentence that risks the house still has to come from a person who lives in one. And option C is waiting somewhere in the text: patterns their Author will recognize the way a dreamer recognizes an interpretation, true only at the very moment a human mouth says them.