Nvidia researchers question the ultra-large context window approach
Understand this story
An Nvidia research team says that language models' context windows of a million tokens or more are often overengineered and costly, and that in multiple tests, periodic summarization plus targeted lookup performed comparably. This judgment targets the rapid expansion of context windows in recent years: GPT-4 had 8,192 tokens at release, while GPT 5.6 Sol already supports about 1 million tokens, roughly 120 times as many. The researchers concluded from this that continuing to expand windows is not a necessary direction.
AI overview based on source reports · Updated 7 hr ago
Coverage timeline
Follow the coverage to explore different perspectives.
- The DecoderNvidia researchers argue huge AI context windows are a dead end
An Nvidia research team argues that very large context windows are often overkill and expensive, since in many tests periodic summaries and targeted lookups performed about as well. The piece notes that context windows have grown from GPT-4's 8,192 tokens in March 2023 to roughly one million tokens in GPT 5.6 Sol, enough for Tolstoy's "War and Peace."
Activity for this story
- Current comparable activity
- 8
- Peak comparable activity
- 10Nov 22 02:00
- Change in the past 24 hours
- –
Trends compare the same fully observed participants, which may cover fewer participants than current activity. Move the pointer or tap the chart to inspect hourly activity; use the left and right arrow keys with a keyboard.