Bush observed that the scientific literature was expanding exponentially while the researcher's capacity to consult it remained constant. Without compression mechanisms, valuable knowledge would be buried under accumulating publication. He proposed several forms of compression: microfilm reduced physical volume, mechanical indexing accelerated retrieval, and associative trails condensed expert navigation into shareable paths. But Bush's most important insight was that compression is not merely technical—it is cognitive and social. The researcher needs not just smaller storage or faster access but meaningful organization that matches how problems are actually approached. Contemporary AI represents the most powerful compression mechanism ever created: billions of documents condensed into neural network weights that respond to natural-language queries with synthesized outputs. The compression is lossy—context is simplified, nuance is flattened, sources are obscured—but the accessibility is unprecedented.
Bush's compression problem was quantitative in 1945 and remains so in 2026, though at vastly different scales. In 1945, a research library might contain hundreds of thousands of volumes; locating relevant material required hours of card catalog consultation and physical retrieval. In 2026, the digital archive contains billions of documents; locating relevant material without AI assistance is effectively impossible. The compression problem Bush diagnosed has been solved and re-created at higher magnitude: AI compresses the corpus into navigable form, but the corpus grows faster than compression mechanisms improve, and the cycle continues.
Compression always involves trade-offs between fidelity and accessibility. Microfilm sacrificed readability for storage density. Abstracts sacrificed context for brevity. Keyword indexing sacrificed semantic richness for search speed. AI compression sacrifices source transparency for response fluency—users receive synthesized answers without seeing which documents contributed which claims. This opacity is a feature, not a bug: the compression works precisely because it conceals the mechanical complexity that alphabetic retrieval made visible. But the concealment creates verification problems: users cannot easily check whether the compressed output accurately represents the uncompressed sources.
The Vannevar Bush — On AI simulation emphasizes that compression is stratified: the neural network compresses the training corpus into weights, the weights compress into activations during inference, the activations compress into tokens, and the tokens compress into the user's understanding. Each compression layer loses information the previous layer contained. The question for augmentation is whether what is lost is mechanical detail (which humans are glad to delegate) or substantive nuance (which humans need to preserve). The answer is both, and distinguishing mechanical from substantive loss requires judgment AI cannot automate.
Bush's thinking about compression emerged from observing scientific photography. Microfilm could reduce a page to a square millimeter—spectacular compression, but useless unless retrieval was equally fast. The constraint was not storage but access: compression without navigability merely relocates the problem. Bush's solution was dual—compress storage through microfilm, compress navigation through associative trails. The combination would make vast knowledge accessible at scales that linear reading could never achieve.
The concept drew on information theory avant la lettre. Bush did not have Claude Shannon's formal framework (published in 1948), but he understood intuitively that information could be represented in multiple formats with different density-accessibility trade-offs. The memex was an information-theoretic device: it would transform high-fidelity, low-density storage (printed volumes) into lower-fidelity, high-density storage (microfilm) while providing high-fidelity, high-speed access through mechanical projection and associative navigation. The design balanced compression and decompression to optimize the researcher's workflow.
Volume growth requires navigational innovation. When the corpus expands faster than reading speed, access depends on compression and search rather than comprehensive consultation.
Lossy compression is acceptable for discovery. Bush recognized that researchers need good-enough access to relevant material more than perfect access to comprehensive material—a principle AI maximizes.
Compression shifts cognitive burden. Every compression mechanism trades one form of labor for another—alphabetical indexing eliminates browsing but requires knowing the right term; associative trails eliminate categorical translation but require building the trail.
Social compression through expert trails. The most powerful compression is not mechanical but curatorial—expert researchers creating trails that compress navigation through a literature into followable paths.
Compression enables new questions. When knowledge becomes accessible at previously impossible scales, researchers can ask questions that comprehensive reading would never permit—comparative studies, cross-domain synthesis, identification of gaps that local reading obscures.
AI's compression mechanisms are debated along familiar lines. Advocates emphasize democratization: AI makes expert-level knowledge accessible to anyone with internet access, compressing years of reading into seconds of query. Critics emphasize loss: the compression obscures sources, flattens context, and produces confident-sounding answers that may misrepresent the nuanced, contested, or ambiguous state of actual knowledge. The debate parallels arguments about every previous compression technology—microfilm, abstracts, executive summaries. Bush's framework suggests both are partly right: compression is necessary for access, and every compression loses something. The question is whether what is lost is recoverable when needed and whether users develop the judgment to recognize when recovery is necessary.