How 'lossy encyclopedia' became the metaphor that explains why AI disappoints and delivers
Silicon Valley's search for the right words to describe artificial intelligence reveals society's deeper struggle with transformative technology
A programmer wanted ChatGPT to generate boilerplate code for connecting a Raspberry Pi to a specific display driver. The request seemed simple enough, create a skeleton project for Zephyr operating system with st7789 SPI display drivers configured. ChatGPT failed spectacularly.
This kind of failure happens daily across Silicon Valley and beyond, yet it still surprises users who've been told they're wielding "artificial general intelligence." Simon Willison, watching this particular frustration unfold on Hacker News, offered both an explanation and a new way of thinking about the entire category of technology, "An LLM is a lossy encyclopedia."
Those four words, published on his blog in September 2025, do more than describe a technical limitation. They capture how society learns to speak about transformative technology—moving from breathless hype toward practical wisdom. Willison's phrase has become the metaphor that technologists didn't know they needed, honest about both promise and peril.
The practitioner who found the words
Willison earned his authority the hard way. As co-creator of Django, one of the world's most popular web frameworks, he possesses unquestionable technical credentials. More importantly, he has spent three years documenting daily AI experiments on his blog, creating a unique record of one expert's journey from enchantment to enlightenment.
This isn't theoretical analysis. Willison uses AI tools every day for coding, writing, and research. He has tested dozens of models, pushed them to their limits, and catalogued their systematic failures. When GPT-4 hallucinates nonexistent Python libraries, he notices. When Claude invents plausible-sounding but wrong historical facts, he catches it. This sustained engagement has given him something most AI commentators lack, genuine intuition about what these systems can and cannot do.
The path has been humbling. Early excitement ("This is revolutionary!") gradually gave way to measured assessment ("This is useful, but..."). Unlike casual users who try ChatGPT once and dismiss it, or enthusiasts who remain uncritically dazzled, Willison represents the sweet spot of informed scepticism. He's learned to work with AI's capabilities while compensating for its systematic blindness.
His metaphors consistently try to recalibrate expectations. In September 2022, he formalised "prompt injection"—a critical security vulnerability that most AI companies prefer not to discuss. He advocates for "slop" as the term for low-quality AI content flooding the internet. Now "lossy encyclopedia" aims to help users understand both what they're getting and what they're missing.
Willison's formulation enters a battlefield. The fight over how to describe AI isn't academic—it shapes user behaviour, investment flows, and regulatory responses. Corporate marketing departments promote "artificial general intelligence" and "digital employees" because these metaphors justify massive valuations and suggest imminent transformation.
Academic critics respond with deflating comparisons, "stochastic parrots," "autocomplete on steroids," "plagiarism machines." Each camp uses metaphors as weapons to advance their agenda. The result is a public thoroughly confused about what they're actually dealing with.
The stakes couldn't be higher. Metaphors carry implicit instructions for use. Call something a "digital assistant" and people delegate important decisions to it. Frame it as a "pattern-matching machine" and users become appropriately sceptical. Label it an "oracle" and people defer to its pronouncements. Each framing leads to different behaviours, different risks, different outcomes.
Research confirms this isn't just linguistic preference. A 2024 study found that describing AI as a "collaborative partner" versus a "sophisticated tool" significantly influenced how much users trusted its output. Partners get the benefit of the doubt; tools get verified.
Willison's "lossy encyclopedia" attempts something more subtle than cheerleading or fear-mongering. It acknowledges sophisticated capabilities while warning about systematic gaps. It suggests appropriate use cases while discouraging dangerous over-reliance.
The evolution of understanding
The journey from Ted Chiang's "blurry JPEG of the web" to Willison's "lossy encyclopedia" reveals how collective understanding matures. Chiang's 2023 metaphor, brilliant and influential, focused on explaining AI's technical architecture. Language models compress vast amounts of text into neural weights, necessarily losing information in the process. Think JPEG compression, but for knowledge.
Willison's refinement adds something crucial, practical wisdom about when compression artifacts matter. Encyclopedias are reference tools with well-understood limitations. You consult them for general information, historical context, broad overviews. You don't rely on encyclopedias for legal advice, medical diagnoses, or precise technical specifications. The "lossy" qualifier warns that even appropriate uses require verification.
This progression mirrors how societies domesticate transformative technologies. Early metaphors emphasise novelty and revolutionary potential. Later ones become practical and specific as users accumulate real experience. Personal computers evolved from "electronic brains" to "productivity tools." The internet graduated from "information superhighway" to "global network."
But here's what makes AI different, the gap between marketing promises and user reality has never been wider. Companies tout "PhD-level intelligence" while users encounter error rates exceeding 15% even in the best models. OpenAI's cutting-edge reasoning system hallucinates 33% of the time on basic knowledge questions. The smaller version hits 48% errors.
The grassroots wisdom emerging
Visit any forum where experienced users discuss AI and you'll find a different conversation than the one happening in boardrooms. Programmers share techniques for using AI as a coding assistant while avoiding its tendency to invent nonexistent functions. Writers describe AI as useful for brainstorming but dangerous for fact-checking. Medical professionals warn colleagues against trusting AI for dosing information after encountering potentially lethal inaccuracies.
This grassroots wisdom contradicts official messaging. While companies promote autonomous AI agents, practitioners advocate "human in the loop" approaches. Where marketing emphasises speed and efficiency, users stress verification and oversight. The Hacker News discussion around Willison's blog post reveals sophisticated individuals who have learned to dance with AI's capabilities while avoiding its traps.
One programmer describes treating AI like "an enthusiastic intern who knows everything but might be wrong about any particular thing." Another compares it to "a savant who's read the entire internet but sometimes confuses what they've read." These aren't dismissive comparisons—they're practical frameworks for productive collaboration with an alien intelligence.
The pattern is clear, actual users are developing measured assessments while official channels maintain transformational rhetoric. Willison's metaphor resonates because it aligns with user experience rather than investment narratives.
The domestication imperative
Humans cannot help but domesticate alien technologies through familiar comparisons. It's how we've always made sense of the impossible. Early automobiles were "horseless carriages." Television was "radio with pictures." Mobile phones were "portable telephones."
These analogies serve essential functions during technological transitions. They make unfamiliar capabilities comprehensible. They suggest use cases and warn about limitations. They help users develop practical intuitions for navigating new tools.
But analogies also constrain thinking. "Horseless carriage" framing kept early automotive design trapped in buggy configurations. "Electronic typewriter" thinking delayed recognition of computers' broader potential. "Information superhighway" metaphors emphasised consumption over the internet's interactive possibilities.
AI metaphors carry similar risks. "Artificial intelligence" suggests human-like reasoning that current systems fundamentally lack. "Digital assistant" implies reliable delegation to demonstrably unreliable systems. "Oracle" framing encourages deference to outputs that systematically require verification.
Willison's "lossy encyclopedia" tries to thread this needle. It's specific about capabilities and limitations. Encyclopedias are useful but bounded. They excel at general knowledge but fail at precision tasks. Adding "lossy" warns about systematic gaps while preserving recognition of genuine utility.
As AI systems become ubiquitous, the metaphors we use will increasingly shape their impact. These frameworks determine whether users verify outputs, how businesses integrate new tools, which regulatory approaches governments pursue. Getting the metaphors wrong could lead to dangerous over-reliance or wasteful under-utilisation.
The "lossy encyclopedia" metaphor matters because it emerges from sustained practical engagement rather than commercial interest or theoretical speculation. It acknowledges impressive capabilities while providing specific guidance about appropriate use. It's neither dismissive nor uncritically enthusiastic—exactly the kind of measured assessment needed as AI transitions from experimental novelty to everyday tool.
Whether this particular metaphor succeeds matters less than the pattern it represents, society's gradual development of more sophisticated frameworks for artificial intelligence. We are learning to speak AI's language, developing conceptual tools for navigating both promise and peril.
Voices like Willison's provide essential guidance in this collective sense-making project. His metaphors don't just describe technology—they teach us how to think about using it wisely. In a world drowning in AI hype and counter-hype, that measured wisdom has never been more valuable.
