Brief IA
›

Nvidia: Researchers Challenge Very Long Contexts

Models & LLM·Tom Levy·

Nvidia: Researchers Challenge Very Long Contexts

Nvidia: Researchers Challenge Very Long Contexts
Key Takeaways
1Researchers from Nvidia question the usefulness of very large context windows.
2GPT 5.6 Sol manages around one million tokens.
3Periodic summaries and targeted searches have achieved similar performance in many tests.
Why it matters : Researchers consider very large context windows to be often excessive and costly.

Researchers at Nvidia believe that the very large context windows of language models represent a dead end. These models are now capable of absorbing entire novels in one go.

Increased Capabilities, but Contested

The working memory of a model is referred to as the "context window": it gathers the elements used to formulate its response. Its capacity has significantly progressed in recent years. At its launch in March 2023, GPT-4 accepted 8,192 tokens. GPT 5.6 Sol now accepts around one million, which is about 120 times more, enough to contain Tolstoy's novel, War and Peace.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

However, the Nvidia team argues that such capabilities are often excessive and costly. In many tests, periodic summaries and targeted searches have shown similar performance.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.