Garthipson Bubble, AI

In 2024, Google shipped Gemini 1.5 Pro with a one-million-token context window. In 2025, OpenAI shipped GPT-4.1 with the same. In the same year, Meta shipped Llama 4 with ten million. The marketing is straightforward: a context window is a workspace, and bigger workspace means more can fit in it. A paper published in July 2025 by the vector-database company Chroma, testing 18 frontier models including all three of the above, concludes the marketing is wrong. The finding is blunt: "every single one gets worse as input length increases." Not some. Not most. All.

The paper is called Context Rot, and the company behind it is the one whose product is supposed to help with exactly the problem the paper documents. The methodological move that makes the result trustworthy is that the experiments hold task difficulty constant and vary only input length — the model is being asked the same simple question while the surrounding text grows from hundreds to tens of thousands of tokens. Under those conditions, performance degrades continuously from the first length increment tested, not as a cliff near the advertised limit.

The mechanism that explains most of it has a name and a paper. In 2023, Nelson Liu and four co-authors at Stanford published "Lost in the Middle: How Language Models Use Long Contexts" (arXiv:2307.03172, later in Transactions of the Association for Computational Linguistics). They measured what happens when the relevant piece of information is placed at different positions inside a long input. The result is a U-shaped curve: models attend strongly to the beginning and end of their context, and poorly to the middle. In multi-document question answering across twenty documents, accuracy dropped by more than 30% when the relevant document was placed in positions 5–15 compared to position 1 or 20. This holds for models explicitly trained for long contexts. The effect is not a capability gap that more training will close. It is a property of how softmax attention distributes weight across tokens, and it has not gone away in the three years since the paper was published.

The Chroma paper adds two mechanisms on top. First, distractor interference: semantically related but irrelevant content makes things worse in a non-uniform way. Adding one distractor reduces performance; adding four compounds the effect, but some distractors are more distracting than others. Code search is the worst case — test fixtures, deprecated implementations, and similarly-named functions from unrelated modules all sit in the same semantic neighbourhood as the target. A well-organised codebase is, perversely, harder for an LLM to search than a randomly arranged one. Second, structural coherence: across all 18 models, performance was better on randomly shuffled documents than on logically structured ones. The intuitive assumption — that coherent text helps the model — is empirically backwards. Adjacent paragraphs create plausible-sounding distractors that compete with the relevant information for the same attention budget.

A separate failure mode operates at training time rather than inference time. Ilia Shumailov and colleagues published "The Curse of Recursion" in 2023 (arXiv:2305.17493, later in Nature, July 2024) and named it model collapse: when generative models are trained on data produced by earlier generative models, the tails of the original distribution disappear. Rare words, rare phrasings, rare facts. After enough rounds of training on the output of the previous round, the model converges to a degraded version of the original — homogeneous, plausible, and missing the things that made the source distribution worth modelling. The paper gives examples on Gaussian mixture models, variational autoencoders, and large language models. The mechanism is general. The practical implication is that the internet is increasingly full of model-generated text, and any future model trained on a naive crawl of it inherits a quieter version of the same disease.

A few things follow from these results. First, context window size as a marketing number has lost most of its meaning. A 1M-token window does not give a model 1M tokens of reliable attention; it gives a model a 1M-token arena in which signal-to-noise ratio continuously degrades. Anthropic's published work on multi-agent research systems reports that token usage alone explains 80% of performance variance in browsing evaluations, and the architecture that wins is the one where the lead agent never sees the search traces — each subagent does its exploration in its own context window and returns 50–200 tokens of condensed result. The fix is not bigger windows. The fix is smaller, cleaner ones. Second, the popular "needle in a haystack" benchmark that the industry uses to advertise long-context capability is, the Chroma authors argue, the wrong benchmark: it tests lexical retrieval of an explicitly inserted sentence, which is the easiest case. Real questions require semantic inference across context the model has never seen arranged this way, and degrade faster. Third, the recursive-data problem has no clean solution. Watermarking, provenance metadata, and synthetic-data detection are all being developed; none of them returns the lost tails.

The 2025-2026 literature is honest about the limits of what it has measured. The Chroma paper does not claim to explain why structural coherence hurts — it points at interpretability work as the next step. The Liu paper does not claim the U-shape is fixable — it notes the effect persists across model sizes and training approaches. The Shumailov paper does not claim to have prevented collapse — it shows the mechanism and warns. The pattern across the field is that the limits of transformer attention at long context are now an empirical fact, not a research question. The research question is what to do about them.

A one-million-token context window is real, in the sense that the model will accept the input. It is also, in the sense the empirical literature has settled on, mostly a lie.