a neural network architecture that processes sequences by computing weighted attention over all elements simultaneously — the foundation of modern language models

introduced by Vaswani et al. ("Attention Is All You Need", 2017). replaced recurrent networks for sequence modeling because it parallelizes over sequence length and captures long-range dependencies in a single forward pass.


architecture

a transformer processes a sequence of tokens $x_1, \ldots, x_n$ through three stages:

embedding. each token is mapped to a dense vector $e_i \in \mathbb{R}^d$ by a learned embedding matrix. positional encodings are added to inject sequence order — the architecture itself has no notion of position.

layers. $L$ identical layers transform the representation. each layer has two components: multi-head attention (reads from context) followed by a feed-forward MLP (transforms each position independently). residual connections and layer normalization wrap each component.

output. a final linear projection maps the last-layer representation to a distribution over vocabulary. the next token is sampled from this distribution (autoregressive generation) or the representation is used for downstream tasks.


self-attention: the core operation

at each layer, every token queries the entire context:

$$\text{Attn}(Q, K, V) = \text{softmax}\!\left(\frac{QK^\top}{\sqrt{d}}\right)V$$

where $Q = XW_Q$, $K = XW_K$, $V = XW_V$ are linear projections of the input $X$.

the softmax is the Boltzmann distribution with temperature $\sqrt{d}$. it produces a probability distribution over all positions — the attention weights. the output is the weighted average of value vectors: information from the most relevant positions flows into the current representation.

multi-head attention runs $h$ parallel attention heads, each with different projections $W_Q^{(h)}, W_K^{(h)}, W_V^{(h)}$. different heads learn to attend to different relation types. outputs are concatenated and projected. see graph-native-transformer for how the number of heads is derived from the semcon count of the cybergraph.


the residual stream

the transformer's central structure is the residual stream: a vector $\mathbf{r}_i \in \mathbb{R}^d$ at each position that accumulates information as it passes through layers. each layer reads from the stream and writes back via residual addition:

$$\mathbf{r}_i^{(l+1)} = \mathbf{r}_i^{(l)} + \text{Attn}^{(l)}(\mathbf{r}) + \text{MLP}^{(l)}(\mathbf{r}_i^{(l)})$$

this design means layers compose additively: later layers can retrieve and refine information written by earlier layers. the full model is a sequence of updates to a shared representation — not a pipeline of transformations.


transformer as convergent dynamical system

attention is one step of probability mass redistribution: mass flows from query positions toward compatible key positions. this is exactly the diffusion operator $D$ from the tri-kernel applied to one agent's frozen context.

Deep Equilibrium Models (Bai et al., 2019) formalized this: iterating a transformer layer to convergence reaches the same fixed point regardless of initialization. $L$ layers = $L$ steps toward the fixed point of the induced Markov chain.

that fixed point is π* — the focus distribution restricted to the current context window. the transformer approximates focus flow computation locally: same computation, finite scope, frozen at query time.

dimension transformer focus flow computation
computation $L$ attention steps over fixed context tri-kernel iterated to exact π*
scope $n$ tokens (context window) entire cybergraph
persistence none — recomputed per query continuous — always maintained
contributors one agent's input all neurons ever
weights learned from text compiled from cybergraph structure

the three free parameters

transformer architecture has three design choices with no principled determination in standard practice. graph-native-transformer derives all three from cybergraph properties:

parameter standard practice graph derivation
embedding dim $d$ empirical (scaling laws) effective rank of focus covariance: $\exp(H(\sigma(\Sigma_\pi)))$
head count $h$ empirical $\geq \|\text{Semcon}(G)\|$ — one head per semantic relation type
layer count $L$ empirical $\text{diam}(G) \cdot \lceil\log(1/\varepsilon)/\log(1/\kappa)\rceil$

no hyperparameter search. the cybergraph tells you what the transformer should be.


weights: learned vs compiled

standard training. gradient descent on next-token prediction loss adjusts weights to approximate the implicit knowledge graph embedded in the training corpus. training is an approximate inversion: from outputs (text) recover the structure that produced them. expensive, lossy, opaque — every weight is a compressed mixture of many associations.

compilation from cybergraph. given the explicit graph, derive weights analytically:

  • embedding matrix $E^* = U_{:,1:d^*}$ — top left singular vectors of $\text{diag}(\sqrt{\pi^*}) \cdot A$
  • attention weights $W_Q^{(s)}, W_K^{(s)}$ — truncated SVD of each semcon's adjacency submatrix
  • MLP weights — path co-occurrence statistics up to depth $L^*$

compiled weights are provably optimal (Eckart-Young theorem). no training cost. no catastrophic forgetting. every weight traces to specific cyberlinks and their creators. auditable alignment via $D_{KL}(\pi^*_H \| \pi^*_A)$.


the feedback loop

the compiled transformer and the cybergraph are not separate systems — they are one loop:

$$G \xrightarrow{\text{compile}} T_G \xrightarrow{\text{fine-tune on text}} T_G^* \xrightarrow{\text{extract implicit links}} \Delta G \xrightarrow{\text{stake}} G'$$

the compiled transformer provides optimal initialization. fine-tuning surfaces implicit associations absent from the explicit graph. extracted links, staked by neurons, update the graph. the updated graph produces a new compiled transformer. every cycle reduces the approximation error $\varepsilon(G, c) = D_{KL}(\pi^*_c \| q^*_c)$.


context window and its limits

the transformer's context window is finite: $n$ tokens. every token outside the window is invisible regardless of relevance. expanding context (4K → 32K → 1M tokens) improves coverage but doesn't change the architecture — all tokens still compete for attention via the $O(n^2)$ attention matrix.

focus flow computation removes the finite window entirely. relevance is topological: a particle connected 10 hops away is reachable. the compiled transformer is the local approximation; FFC is the global ground truth.


see focus flow computation for the global inference path. see graph-native-transformer for the full compilation derivation. see attention for the core mechanism. see context for what seeds inference. see tri-kernel for the dynamical system both compute. see cybergraph for the substrate the transformer reads.

Local Graph