Chapter 08The whole argument, in one page

The argument, in one page

Seven chapters, seven mechanisms, one sentence: a memory of fixed capacity needs an eviction policy, and a learned selective read is the best eviction policy we know how to build. Everything else was implementation.

The chain

The series has a shape that is easy to miss while you are inside it: each chapter's fix is the next chapter's problem. Not a list of improvements — a chain, where every link is forced by the one before it.

EACH FIX CREATES THE NEXT CONSTRAINT cache growswith N; decodereads all of it FROM 01 02 · move φ offthe product;fixed d×d state writes collideonce N > d;nothing leaves 03 · read beforewrite; erase thiskey exactly recurrence readsits own state —not parallel 04 · WY form;C sequential steps→ one solve can replace, butcannot releasecapacity 05 · add decay α;replace andrelease one half-life forevery channel ofthe state 06 · α → Diag(a);a half-life perchannel no gating fixes capacity — d slots is still d slots 06–07 · stop compressing everything. keep exact attention every 4th layer.
The forcing structure. Red is a constraint, blue is the mechanism that answers it. Nothing here was optional: linear attention's fixed state is what makes interference possible, interference is what the delta rule exists to fix, the delta rule's self-referential update is what makes chunking hard, and so on to the end. The final move is the only one that reverses direction — after six chapters of compressing the memory, K3 concedes that some of it must not be compressed at all.

Everything, in one table

MechanismThe stateUpdate ruleEviction policyWhat it cost
Softmax attention
01 · GPT-2, 2019
KV cache, \(O(N)\), one private slot per token append none — nothing is ever evicted \(O(N)\) memory, \(O(N)\) bytes read per decode step
Linear attention
02 · Katharopoulos 2020
\(S \in \mathbb{R}^{d_k\times d_v}\), fixed \(S \mathrel{+}= k^{\top}v\) none — writes superimpose interference beyond \(d\) associations; weaker feature map than \(\exp\)
Delta rule
03 · Schlag 2021
same \(S \leftarrow (I-\beta k^{\top}k)S + \beta k^{\top}v\) rank-one, content-addressed: erase exactly what this key held a sequential dependency that blocks parallel training
Chunkwise / WY
04 · Yang 2024
same unchanged — this is an algorithm, not an architecture one \(C\times C\) triangular solve per chunk; \(C\) chosen for tensor cores, not for FLOPs
Gated DeltaNet
05 · Mamba-2 + delta
same \(S \leftarrow \alpha(I-\beta k^{\top}k)S + \beta k^{\top}v\) rank-one replace and full-rank release one shared half-life for the whole state
KDA
06 · Kimi Linear 2025
same \(S \leftarrow \mathrm{Diag}(a)(I-\beta k^{\top}k)S + \beta k^{\top}v\) per-channel half-life; \(d_k\) independent horizons numerical range — needs sub-chunking and a hand-written kernel
MLA hybrid
06 · 3 KDA : 1 attention
mixed: 3 fixed, 1 growing but latent-compressed both the fixed layers evict; the attention layer does not ≈¼ the cache traffic, and the concession that compression has a floor
MoE
07 · 898 experts, 18 active
parameters, not activations router selects top-k selects which knowledge to load, per token every expert must stay resident; routing crosses devices
AttnRes
07 · every 12 layers
block outputs along depth \(h_\ell = \sum_i \alpha_i f_i(h_i)\), \(\sum\alpha=1\) selects which earlier computation to read ≈2% latency; reported 1.25× compute advantage

One operation, three axes

Read down the eviction column and the same shape appears nine times. A query. A set of keys. A normalised weighted read of the associated values. What differs between MLA, AttnRes and an MoE router is only what the keys are indexed by.

THE SAME OPERATION, INDEXED THREE DIFFERENT WAYS TOKEN AXIS · MLA tokens 1 … N “Of everything I have read, which words matter here?” keys indexed by position DEPTH AXIS · ATTNRES blocks 1 … 8 “Of everything I have computed, which layer's?” keys indexed by depth PARAMETER AXIS · MoE experts 1 … 898 “Of everything I know, which weights apply here?” keys indexed by expert In all three: a learned query, a similarity score, a softmax, a weighted read. The axis changes. The operation does not.
Attention was never about sequences. It is a general mechanism for reading selectively from a set of things when you cannot afford to read all of them. Once that is clear, MoE stops looking like a scaling trick and AttnRes stops looking like a residual-connection tweak: they are the same primitive, pointed at the parameter axis and the depth axis instead of at the token axis.
The claim, stated plainly

A fixed-capacity associative memory holds on the order of \(d\) separable facts, because there are only \(d\) orthogonal directions in \(\mathbb{R}^d\). Once it is full, a purely additive write adds interference rather than information. So capacity forces an eviction policy, and a good eviction policy has to be learned — hand-designed ones (FIFO, uniform decay, fixed windows) throw away the wrong things. Gating, routing, and the delta rule are three shapes of the same learned decision, and attention is the most effective selective read we currently have to implement it.

What is not settled

Three things in this lineage are load-bearing and not yet proven, and it is worth naming them rather than reporting the numbers as conclusions.

1 · “Better than full attention”

Kimi Linear's headline is that a hybrid outperforms full attention under controlled comparison, not merely that it is cheaper. That is a strong claim, and controlled comparisons at one scale have a poor record of surviving to the next. The efficiency argument stands on its own — fewer bytes read per token is not disputable. The quality argument should be held more loosely until it replicates independently at several scales.

2 · Whether decay is enough position information

Chapter 6 argued that per-channel decay is a learned, multi-scale relative position bias, which is why the attention layers can drop positional encoding entirely. That is elegant, and it is also an empirical bet. Exponential decay is monotone in distance; it cannot express “exactly 4 tokens back” the way a rotary phase can. Whether that matters appears to depend on the task in ways not yet well characterised.

3 · Selection fails silently

Every mechanism here replaces exhaustive computation with a learned choice. Exhaustive computation is wrong loudly; a learned choice is wrong quietly. A router that consistently misroutes a rare domain, a channel whose decay was fit badly, an AttnRes query that never learns to look back — none of these produce an error, they produce a slightly worse model in a way no aggregate loss number distinguishes from “needs more training.” The field has gotten much better at building selection mechanisms than at auditing them.

Where the pressure goes next

Trace which resource was binding in each chapter and the trajectory is legible. GPT-2 was bound by cache bytes. Linear attention traded that for a bound on capacity. The delta rule and gating spent the next four chapters getting the most out of that capacity, and hit the wall that \(d\) slots is \(d\) slots. K3's answer was to stop trying: keep some exact attention, and buy the rest of the capacity as sparse parameters rather than as state.

Which relocates the binding constraint to somewhere new. A 2.8T-parameter model with 2% activation is not compute-bound and not cache-bound — it is bound by having 2.8 trillion parameters resident somewhere, and by the interconnect that moves tokens to whichever device holds the experts they routed to. That is a systems problem, not an attention problem, and it is where the next several years of this story will be written.

The other pressure is the quiet one from the box above. Seven chapters of increasingly sophisticated learned selection have made model behaviour depend on choices that are hard to observe and harder to verify. Building better selectors is going well. Knowing what they selected, and whether it was right, is not yet a solved discipline.


Read the chapters

01 · GPT-2 and the price of perfect recall  ·  02 · Linear attention  ·  03 · The delta rule  ·  04 · The chunkwise form  ·  05 · Learning to forget  ·  06 · Kimi Delta Attention  ·  07 · Kimi K3