How Transformers Stack Meaning Like Finnish Words

Arshavir Blackwell, PhD

Welcome to Inside the Black Box. I'm Arshavir Blackwell. Today we're exploring a mental model—an analogy—that I think can sharpen your intuition about how large language models work. This analogy is Finnish morphology. Not how LLMs *process* Finnish—that's a different topic—but how the *structure* of Finnish, the way it builds meaning, mirrors something deep about how transformers build internal representations. This is a thinking tool, not a technical claim. Let's see how far it takes us.

Arshavir Blackwell, PhD

Finnish is what linguists call an agglutinative language. That means it builds words and meaning by stacking morphemes—small meaning-bearing units—in sequence. Each morpheme does exactly one job. Take the word "talo." It means house. . Add the suffix "-ssa" and you get "talossa"—in the house. That "-ssa" is the inessive case marker. It's doing one job: marking location. Add "-si" and you get "talossasi"—in *your* house. That "-si" is a possessive suffix. One job: marking ownership by "you." Add "-kin" at the end—"talossasikin"—and now it means *even* in your house, or *also* in your house. The "-kin" is a focus particle adding emphasis.

Arshavir Blackwell, PhD

Four pieces, stacked in order, each transparent in its function. English needs a whole phrase: "even in your house." Finnish collapses it into one word. Here's what makes Finnish useful as a mental model: you can *point* to each piece and say what it contributes. The stem provides the core meaning. Each suffix modifies that meaning in a specific, identifiable way. And the final word—the surface form—integrates everything into a single unit. That structure—base meaning, sequential modification, integration into a final form—is what I want you to hold onto. Because something analogous, though of course not exactly the same, happens in transformers.

Arshavir Blackwell, PhD

When a transformer processes a word like "house," it starts with a token embedding—a high-dimensional vector that encodes the model's baseline sense of that word. Think of this as analogous to the Finnish stem: the core lexical meaning. But that embedding doesn't stay static. As it flows through the transformer's layers, it gets modified. The mechanism doing this modification is multi-head attention. Here's the key parallel: each attention head acts like a specialized operator that adds contextual information. Just as Finnish suffixes each contribute one grammatical feature, attention heads each contribute one type of contextual modification.

Arshavir Blackwell, PhD

Some attention heads track syntactic role—is "house" the subject, the object, or part of a prepositional phrase? Others track referential information—whose house? How definite? How recently mentioned? Still others add what we might call pragmatic coloring—is this literal ("the house collapsed") or metaphorical ("the capital ‘H’ House voted")? Researchers in mechanistic interpretability have actually found attention heads that specialize in specific functions, as we have noted in previous episodes. There are heads that track subject-verb agreement. Heads that handle coreference—figuring out what pronouns refer to. Heads that implement negation.

Arshavir Blackwell, PhD

These are functional specializations—much like Finnish suffixes have functional specializations. Now, the *mechanism* between Finnish and LLMs is quite different. Finnish stacks suffixes through linear concatenation—you literally attach one morpheme after another. Transformers stack features through vector addition in high-dimensional space. Finnish morphology is symbolic and discrete. Attention is geometric, continuous, and can act at arbitary far distances. But the compositional *logic* is the same: start with a base, apply a sequence of modifying operations, and produce an integrated result. [Arshavir Blackwell, PhD] To make this analogy more concrete, let's talk about the residual stream. In Anthropic's "Mathematical Framework for Transformer Circuits," the residual stream is described as a communication channel that runs through the entire model. Each layer reads from this stream and writes back to it. Here's the crucial point: information in the residual stream *accumulates*. When an attention head adds its contribution, it doesn't replace what was there—it adds to it. The features build up, layer after layer.

Arshavir Blackwell, PhD

This is directly analogous to agglutination. In Finnish, you start with "talo" and accumulate suffixes: "-ssa" adds location, "-si" adds possession, "-kin" adds emphasis. Each suffix modifies what came before without erasing it. In a transformer, you start with the token embedding and accumulate attention contributions: one head adds syntactic role, another adds referential properties, another adds discourse context. Each modification enriches what came before without erasing it.

Arshavir Blackwell, PhD

By the final layer, the token's vector representation carries traces of everything that happened to it. It's no longer just "house"—it's "house as modified by all the contextual operations applied through the network." And then, just as Finnish collapses all those stacked suffixes into one pronounceable surface form, the transformer collapses all those accumulated modifications into one probability distribution over the vocabulary—and predicts the next token.

Arshavir Blackwell, PhD

So why is this analogy useful? What does it buy you that you didn't have before? First: it gives you intuitions about what to look for in interpretability research. If transformers implement something like compositional grammar, then we should expect to find functional units—attention heads or circuits—that behave like grammatical operators. And indeed, that's what researchers are finding. Heads that reliably fire for possessives. Heads that track agreement. Heads that handle scope. Not exactly like Finnish, of course, but still a useful model for guiding our intuitions.

Arshavir Blackwell, PhD

The Finnish mental model tells you these functional units *should* exist and *should* be discoverable. It gives you a hypothesis to test. Second: it helps you think about model failures. When a model consistently struggles with a particular construction, you can ask: which compositional operation is failing? Is there an attention pattern that should be adding some feature but isn't? Is there interference between operations that need the same representational resources?

Arshavir Blackwell, PhD

Finnish handles this transparently—if a learner messes up "-ssa," you know they're confusing location marking. Transformers aren't that transparent, but the analogy suggests there might be analogous points of failure.

Arshavir Blackwell, PhD

Third: it's a communication tool. If you're explaining transformers to someone unfamiliar with the architecture, Finnish provides a concrete, intuitive entry point. Most people can understand the idea of suffixes that modify meaning. From there, you can bridge to attention heads as "geometric suffixes"—operators that modify vectors rather than concatenating morphemes.

Arshavir Blackwell, PhD

Now, I want to be clear about the limits. This is a mental model, not an isomorphism. Finnish morphology operates locally. Suffixes attach to adjacent material. Attention can reach arbitrarily far back in context—across sentences, across paragraphs.

Arshavir Blackwell, PhD

Finnish suffixes are discrete and categorical. You either have "-ssa" or you don't. Attention is continuous and graded. You can, so to speak, have half of a ‘-ssa.’. There is, conceivably, such a thing as half of a rule, if we are allowed to conceptualize attention heads as approximate stand-ins for rules. But remember that heads only approximate rules—heads attend to things with varying weights, and those weights are learned, not specified by a grammar.

Arshavir Blackwell, PhD

And crucially: we don't understand what most attention heads are doing. Many are polysemantic—they activate for multiple unrelated features. Many operate at levels we can't easily interpret semantically. So while Finnish morphology is transparent—you can point to a suffix and name its function—transformer attention is largely opaque. We're working backward from behavior to mechanism. The analogy gives you a *framework* for thinking, not a *map* of what's actually there. It's a flashlight, not a blueprint.

Arshavir Blackwell, PhD

Let me close with the deeper point. Both Finnish and transformers solve the same fundamental problem: how do you take finite building blocks and compose them to express unbounded meaning?

Arshavir Blackwell, PhD

Finnish (and other agglutinative languages) does it with a limited set of morphemes—maybe 15 grammatical cases, a handful of possessive suffixes, various particles—combined according to rules that speakers internalize. Transformers do it with a limited set of parameters—attention weights, embedding dimensions, feedforward layers—combined according to patterns learned from data.

Arshavir Blackwell, PhD

Finite resources. Infinite expression. That's the power of compositionality. The Finnish analogy suggests that what we see as morphological rules—these symbolic, categorical operations—may be surface manifestations of a deeper computational principle. A principle that transformers have rediscovered, not from grammar books, but from the statistical structure of language itself. That's the real payoff of this mental model. It hints that transformers aren't just pattern-matching on text. They're implementing something like compositional reasoning—stacking operations, accumulating features, integrating everything into predictions. What Finnish makes visible through morphemes, transformers make invisible through geometry. But the compositional logic might be the same.

Arshavir Blackwell, PhD

To recap: Finnish morphology, where meaning stacks through visible suffixes, provides a mental model for how transformers build representations through layered attention. It's not a perfect analogy. It's a thinking tool. But it gives you intuitions about functional specialization, compositional structure, and what to look for when models fail. I'm Arshavir Blackwell. This has been Inside the Black Box.