{"slug": "we-know-more-than-we-can-tell", "title": "We know more than we can tell", "summary": "LLMs are marketed as productivity tools, but tacit knowledge—gained from experience and impossible to fully articulate—remains the most valuable and scarce resource in companies, as it cannot be copied or transferred. The author argues that the type of knowledge a company relies on should match its goal, with tacit knowledge being crucial for novel, ambiguous, and high-stakes problems.", "body_md": "We are being sold LLMs as incredible productivity tools, and the more I hear that pitch the more I think about Michael Polanyi. He had a simple line: we know more than we can tell. He called the gap tacit knowledge. It is the stuff we pick up from experience and internalize without noticing, so we can't really explain it. How to ride a bike. When a deal feels off. Which of two correct looking designs is the one that quietly rots in six months.\n\nSo instead of asking \"are LLMs good enough to replace humans\", I want a more useful question: what kind of knowledge is actually scarce inside a company, and therefore actually valuable?\n\nA company runs on roughly four kinds. Tacit: learned from experience, hard to put into words, only humans have it. Explicit: credentials, documented procedure, codified and transferable. LLM: average token generation, conditioned on the quality of the input. Deterministic: knowledge encoded directly into programs, exact and repeatable, no judgment.\n\nImagine four companies chasing the same goal, X, each betting everything on one kind. Deterministic wins when X can be fully specified. Encode it once and it runs forever for almost nothing. But most real goals can't be fully specified. Explicit is legible and scales through onboarding, but it is inert, and the moment the situation drifts off the page it has nothing to say. An LLM behaves like it knows your problem for a bit, then that is over and the knowledge is gone. It lives and dies to fulfill the prompt. It never wrestles with whether the goal is good.\n\nTacit is the strange one, and it is the one the marketing quietly prices at zero. A person carrying it is not only working during working hours. We sleep on a problem and come back the next day with a sharper answer. We solve it in the shower. And here is the part that matters most: knowledge that escapes words is knowledge you can't transfer by telling. You can only get it the slow, expensive way, by doing it badly at first until it settles into instinct. That one property, costly to get and impossible to hand over, is what makes it scarce. Explicit copies for free. Code copies perfectly. An LLM's competence can be rented by anyone who types the same prompt. Tacit knowledge has no copy operation.\n\nNobody can explain how the senior engineer knew the system would rot, but everyone can see it didn't rot on their watch. Clear results, hidden method. That is a moat a competitor can't close by listening to you talk. It is close to why companies exist at all. Markets are great at trading what you can write into a contract, and useless at trading what you can't say. So you gather people under one roof and let it grow inside them.\n\nThe usual comeback is scale: don't compare one human to one model, compare a person to a swarm of agents running past any working day, even past a lifespan, building more agents forever. But scale multiplies token generation, not judgment. A thousand agents arguing in a data center still don't care whether the goal was worth it. You scaled the output, not the stake.\n\nSo who wins? It depends on the goal, and any company that picks a knowledge type before understanding its X is doing astrology. If X is broad and shallow, the LLM company wins. If it is coordination and handover and compliance, explicit wins. If it is new, ambiguous and high stakes, tacit wins, and it is the only one that gets better on its own over time.\n\nSo when you hear \"10x productivity\", ask which kind of knowledge your goal is actually bottlenecked on. And if you are making a career bet, build the kind nobody can extract from you. That one is yours.\n\nExciting, and slightly unnerving, time to be working in this field.", "url": "https://wpnews.pro/news/we-know-more-than-we-can-tell", "canonical_source": "https://www.nvegater.com/blog/tacit-knowledge", "published_at": "2026-06-17 15:40:20+00:00", "updated_at": "2026-06-17 15:53:18.277404+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-products"], "entities": ["Michael Polanyi"], "alternates": {"html": "https://wpnews.pro/news/we-know-more-than-we-can-tell", "markdown": "https://wpnews.pro/news/we-know-more-than-we-can-tell.md", "text": "https://wpnews.pro/news/we-know-more-than-we-can-tell.txt", "jsonld": "https://wpnews.pro/news/we-know-more-than-we-can-tell.jsonld"}}