Attention Sink: The Fluke That Made LLMs Actually Usable

Source: Attention Sink: The Fluke That Made LLMs Actually Usable, bycloud, 9:15, uploaded 2025-05-14, playlist index 263.

Large language models depend on attention to relate the current token to the tokens that came before it. Bycloud’s video starts with a puzzle inside that familiar mechanism. Long context looks like a natural extension of attention, yet most alternatives struggle to scale. The technique that does scale may owe part of its success to a behaviour researchers discovered by accident: the model gives a large share of its attention to a token that carries little meaning.

The first token and the long-context problem

The video places the puzzle in the history of the transformer. Attention has remained the dominant mechanism for language models since 2017, even as researchers have gained more computing power and designed many alternatives. The context window has grown from the roughly 4,000 tokens used during the pre-training of the example model towards 8,000, 36,000, or 64,000 tokens in newer systems.

That growth creates a practical problem. Storing the relationships for every token becomes expensive when the conversation or document grows. A simple way to extend a model beyond its training length is sliding-window attention. For an 8,000-token context, the model could keep only the most recent 4,000 tokens and move that window forward as new tokens arrive.

The experiment fails as soon as the first token leaves the window. The model loses coherence and becomes unable to generate reliable sentences. Meta researchers found that the model could keep working if the window always included the first token alongside the recent tokens. This restored coherent generation, although it did not restore the information that had fallen outside the window.

The result suggests that the first token has a function beyond the meaning of the text it represents. Bycloud connects this finding to research Meta shared in 2023, which found that many attention heads across transformer layers assign a disproportionate amount of attention to the first few tokens. The video reports figures of roughly 60 to 80 percent for the beginning-of-sequence token, or BOS token. The token marks the start of a sequence and usually carries little semantic information, yet the model keeps returning to it.

Attention as a mixture

The 2025 Google paper Why Do LLMs Attend to the First Token?, which the video uses to explain the result, treats the sink as a learned solution to over-mixing. The explanation begins with the ordinary attention calculation.

For each current token, the model assigns an attention score to the earlier tokens. The scores add up to one, so they determine the proportion of each earlier token’s value that enters the current representation. A high score gives one token a strong influence. A low score leaves it with little influence. The result is a weighted average of the values carried by the earlier tokens.

Bycloud explains this with a smoothie. Chocolate, mango, and ginger each have a strong flavour. Combining too many strong flavours can produce an unpleasant mixture, so a drink needs something that dilutes them. In the attention calculation, a meaningless token can play that role. When the model has no useful match for the current token, it can direct attention towards the sink instead of spreading the remaining weight across several semantically strong tokens.

The sink therefore acts as a place where attention can go when the model needs a neutral state. During training, the model learns that this low-information token helps keep meaningful representations from being averaged together too aggressively. The model’s generation stays coherent because the attention heads can decline to mix more information when the available matches would make the representation less useful.

This also explains the sliding-window result. Removing the sink forces the attention scores onto the tokens that remain. The model loses the neutral destination that lets it preserve distinctions among older signals. Keeping the sink in the window gives the model a stable place to put excess attention, even though the rest of the window moves through the sequence.

Why the sink sits at the beginning

An autoregressive model generates the next token from the tokens that came before it. The first token stays visible to every later position, which makes it the only consistent candidate for a universal sink. Its special-token status also gives it little semantic competition with the words that carry the content of the prompt.

The video describes two related benefits. When the current prediction contains noisy or weak evidence, the model can focus on the sink and keep its state close to neutral. The sink also reduces the loss of information that comes from repeatedly blurring distinct tokens together. A representation can retain the difference between token 50, which concerns one subject, and token 99, which concerns another, because the model has somewhere else to place attention.

This behaviour does not give the model unlimited memory. The moving window still drops older tokens. The sink protects the mechanics of coherent generation, while the information outside the window remains unavailable. That distinction matters because a fluent continuation can survive even when the model has forgotten much of the earlier conversation.

A fluke with a useful explanation

Bycloud frames attention sink as a fortunate accident in the design of the attention mechanism. The video uses “fluke” to describe the fact that a token with little semantic content became a structural part of a scalable context strategy. The model learns the behaviour during training, so the accident lies in the properties of the mechanism and its training setup rather than in a researcher manually adding a special rule after the fact.

The video comes from a technical blog by Ethan Smith, Softmax Attention is a Fluke, which it recommends for the mathematical treatment. The captions stay at the level of the mechanism and the smoothie analogy. They do not give the models, datasets, attention heads, or measurements behind the reported 60 to 80 percent figure. The two papers named in the description provide a route to that detail, while the video itself supports a narrower account: the first token can serve as a stable low-information destination that helps attention avoid destructive mixing.

Further reading / references

• Efficient Streaming Language Models with Attention Sinks, the 2023 paper named in the video description.
• Why do LLMs attend to the first token?, the 2025 Google paper used to explain over-mixing.
• Softmax Attention is a Fluke, Ethan Smith’s mathematical blog post and the video’s direct inspiration.