▶ Article reader Listen + reading controls ⌄

LISTEN + READ YOUR WAY

Article reader

Preparing the reader…

▶ Play article ← → Stop

0:00 0:00

Speed 0.75× 1× 1.25× 1.5× 1.75×

Reading settings

Typeface Field Notes Atkinson Hyperlegible Lexend OpenDyslexic

Text size

A−

100%

A+

Reading surface Field Notes dark Paper light Warm sepia High contrast Text spacing Standard Comfortable

Reset reading preferences

Long context changes the knowledge-work interface

Anthropic has released Claude 2 with a context window that can accept roughly 100,000 tokens—enough for hundreds of pages of material in one prompt. The immediate attraction is obvious: a user can bring a long report, technical documentation, or even a book into a conversation without dividing it into tiny fragments.

More context changes what a language model can see. It does not guarantee that the model will attend to the right thing.

A larger window reduces one kind of fragmentation

Knowledge workers often lose meaning when documents are separated from their history and surrounding evidence. A requirement makes more sense beside its rationale. An incident report makes more sense beside prior events. A contract clause may depend on definitions many pages away.

A long context window can keep more of those relationships available to the model. That may improve synthesis, comparison, extraction, and question answering. It also lowers the engineering burden of building a retrieval system for every initial use.

For exploratory analysis, this is a meaningful improvement.

Availability is not attention

Placing a fact in the context does not establish that the model will use it reliably. Important material may be diluted by volume, contradicted elsewhere, or expressed in a format the model handles poorly. The output can still sound complete even when a crucial exception was missed.

The human user faces a parallel problem. If the model produces a concise answer from 300 pages, the convenience can make verification harder. The person may not know which passage mattered, which material was ignored, or whether the model blended incompatible sources.

That turns provenance into an interface requirement. Every consequential synthesis should link claims back to specific passages, preserve document identity and date, and distinguish direct extraction from generated interpretation.

Context should be curated, not dumped

Organizations may respond to larger windows by sending every available document. More is not always better. Irrelevant, obsolete, duplicated, or unauthorized material can reduce quality and increase exposure.

Knowledge management still matters. Teams need rules for:

• which sources are authoritative for a task;
• which version is current;
• who may access each document;
• how conflicting sources are represented;
• what metadata travels with the text; and
• when a source should expire or be reviewed.

This is the same problem organizations already face in shared drives and document repositories. A large language model (LLM) makes the consequences more immediate because it can synthesize disorder into a confident response.

Nonaka's work on organizational knowledge creation also reminds us that documents do not contain all relevant knowledge. Experienced practitioners recognize context, exceptions, and relationships that were never written down. Long context expands access to explicit knowledge; it does not eliminate the need for tacit judgment.

Design a reading partnership

A useful long-document workflow should make the division of labor clear:

1. The person defines the question and relevant decision.
2. The system maps the document set and identifies likely evidence.
3. The system produces a synthesis with passage-level references and uncertainty.
4. The person inspects critical evidence and corrects interpretation.
5. The final work product retains the source trail and human decision.

Evaluation should include deliberate “needle” cases: important qualifications hidden in appendices, conflicting versions, tables, unusual terminology, and statements that should cause the system to decline a conclusion.

Claude 2's larger context makes language models more natural partners for serious document work. It also raises the standard for product design. When the machine can read more than the person can readily check, the interface must help the person see what mattered.

The opportunity is not to stop reading. It is to create a better partnership between machine-scale review and human-scale judgment.

Sources and research trail

• Anthropic, “Claude 2” (July 11, 2023).
• Anthropic, “Introducing 100K Context Windows” (May 11, 2023).
• Nonaka, “A Dynamic Theory of Organizational Knowledge Creation” (1994).
• Argote and Ingram, “Knowledge Transfer: A Basis for Competitive Advantage in Firms” (2000).
• Lee and See, “Trust in Automation: Designing for Appropriate Reliance” (2004).