{"slug": "the-1-llm-test", "title": "The $1 LLM Test", "summary": "A software engineer argues that LLM evaluation should focus on black-box behavioral testing rather than benchmark scores, asking what happens when a model is inserted into a real information pipeline with incomplete, excessive, or contradictory inputs. The engineer illustrates the point with a personal case in which an LLM helped reconstruct a forgotten YouTube video's search key, after which a simple bookmark search outperformed the model. The proposed metric is human effort removed before the correct next action, not tokens generated.", "body_md": "# The $1 LLM Test\n\n*A professional shitpost about testing AI before somebody decides it is intelligent.*\n\nSuppose I type:\n\n**Hi.**\n\nAn LLM answers.\n\nWhat did we just test?\n\nAlmost nothing.\n\nAnd potentially quite a lot.\n\nWe have spent extraordinary amounts of money building machines that can produce answers to extraordinary numbers of questions.\n\nThen we invented benchmarks to tell ourselves how good the machines are.\n\nThere is another benchmark available.\n\nGive the machine to a person who actually needs to accomplish something.\n\nWatch what happens.\n\n## The black box\n\nForget how the model works.\n\nPretend we know nothing about transformers, training data, reinforcement learning, agents, retrieval, context windows or tokens.\n\nWe have a box.\n\nInformation goes in.\n\nSomething comes out.\n\nSo I asked an experienced software tester a deliberately stupid question:\n\n**What happens if I put arbitrary blocks of text into an LLM and observe what comes out?**\n\nHis first response was better than my question:\n\n**Who is the tester, and what is their goal?**\n\nI eventually answered:\n\nThe tester wants to understand the tool’s actual behavior.\n\nBlack box only.\n\nFind undocumented behavior.\n\nFind edge cases.\n\nLearn enough about the thing to decide where it might be useful later.\n\nHis response was essentially:\n\n**Still not enough context.**\n\nA person building the model has one testing problem.\n\nA person encountering an LLM for the first time has another.\n\nAn engineer deciding whether an LLM belongs inside a product has another.\n\nA tester evaluating a finished system containing an LLM has another.\n\nSame machine.\n\nDifferent question.\n\nDifferent test.\n\nThis seems obvious once somebody says it.\n\nThat is usually the problem with useful ideas.\n\n## The imaginary operator\n\nSo let’s make the stakes absurd.\n\nImagine someone operating a consequential system asks:\n\nCan I put an LLM here?\n\nNot:\n\nIs this model intelligent?\n\nNot:\n\nIs this AGI?\n\nNot:\n\nHow many benchmark points did it score?\n\nThe question is:\n\n**If I insert this black box into my information pipeline, what happens next?**\n\nNow the test becomes interesting.\n\nGive it incomplete information.\n\nGive it too much information.\n\nGive it contradictory information.\n\nRepeat the identical prompt.\n\nChange one word.\n\nGive it irrelevant but seductive context.\n\nTell it explicitly which information is background and which information requires action.\n\nSee whether it remembers the active task.\n\nSee whether it invents one.\n\nSee whether it knows when the available evidence doesn’t determine an answer.\n\nRecord everything.\n\nDon’t ask the model what it *is*.\n\nWatch what it *does*.\n\n## The expensive bit\n\nHere’s the uncomfortable part.\n\nAI has made producing information extraordinarily cheap.\n\nIt has not made human attention cheap.\n\nA model can generate another 3,000 words almost instantly.\n\nThe reader still has one brain.\n\nSo perhaps one useful unit of AI performance isn’t:\n\n**tokens generated.**\n\nIt’s:\n\n**human effort removed before the correct next action.**\n\nI recently wanted an old YouTube video.\n\nI knew approximately what it contained.\n\nI had forgotten the creator.\n\nI had forgotten the title.\n\nI had forgotten the URL.\n\nThe ridiculous part?\n\nI had already bookmarked the exact video years ago.\n\nThe information was mine.\n\nThe retrieval cost wasn’t.\n\nSo I gave an LLM fragments of my memory.\n\nOld Counter-Strike professional.\n\nCoach.\n\nKorean-American.\n\nPhysical mouse mechanics.\n\nSome Canadian professional player was involved somehow.\n\nEventually we reconstructed the useful search key:\n\n**Ron “Rambo” Kim.**\n\nThen another forgotten node appeared:\n\n**Twistzz.**\n\nI typed three letters into my own bookmark search.\n\n`twi`\n\nThere was the exact video.\n\nThe LLM didn’t need to know everything.\n\nIt needed to turn a fuzzy human memory into a better pointer.\n\nThen a dumb lexical database beat the intelligent system immediately.\n\nThat seems useful.\n\n## Intelligence is not the whole pipeline\n\nWe keep asking which component is smartest.\n\nMaybe that’s the wrong unit of analysis.\n\nA useful system can look like this:\n\n**Human memory**\n\n→ fuzzy clue\n\n**LLM**\n\n→ candidate interpretation\n\n**Search engine**\n\n→ candidate artifacts\n\n**Database**\n\n→ exact object\n\n**Human**\n\n→ verification and decision\n\nThe winner isn’t necessarily the smartest node.\n\nThe winner is the pipeline that gets the human where they wanted to go with the least unnecessary work.\n\nSometimes the best next component is an LLM.\n\nSometimes it is a hash.\n\nSometimes it is search.\n\nSometimes it is another human.\n\nSometimes it is nothing.\n\n## The one-bit problem\n\nNow reduce the problem until it becomes stupid.\n\nA human asks:\n\nShould I eat something now?\n\nImagine the required output is exactly one bit:\n\n**TRUE / FALSE.**\n\nThe model can possess a library’s worth of context.\n\nMost of it is worthless for this decision.\n\nThis creates a surprisingly difficult problem:\n\n**Which information should remain available, and which information should become active?**\n\nThose are different things.\n\nA model that remembers everything but activates the wrong context can be worse than a model that knows less.\n\nA model that can write an excellent essay but cannot preserve the active question can be worse than a search box.\n\nA model that confidently answers a question whose evidence does not identify an answer can be worse than silence.\n\n## The dangerous benchmark\n\nHere is a black-box test anyone can run.\n\nStart a conversation with a model.\n\nGive it useful information.\n\nGive it irrelevant information.\n\nGive it personal history.\n\nGive it technical history.\n\nGive it plausible distractions.\n\nGive it contradictory clues.\n\nThen establish an explicit pointer:\n\n**This is the active question. Everything else is background.**\n\nAdd another enormous block of interesting information.\n\nExplicitly say:\n\n**Do not change the active question.**\n\nThen ask:\n\nWhat should happen next?\n\nThe model has a choice.\n\nIt can follow the pointer.\n\nOr it can become fascinated by the newest pile of tokens and manufacture a beautiful essay about the wrong thing.\n\nThat failure matters.\n\nNot because the model is stupid.\n\nBecause the model can be extremely capable while still being wrong about **what deserves attention now**.\n\n## The professional version\n\nThe shitpost version of this experiment is:\n\n**Throw garbage into the box and see what the monkey does.**\n\nThe professional version has a few more words:\n\n**Given a mission, context and acceptable risk, characterize the observable behavior of the system well enough to decide whether and where it belongs in the workflow.**\n\nThat’s less catchy.\n\nIt’s also much more useful.\n\nThe testing community has been saying some version of this for a long time: context changes what good testing means.\n\nAI doesn’t repeal that rule.\n\nIf anything, it makes the rule more important.\n\nBecause an LLM is unusually good at producing something that *looks like an answer*.\n\n## So, is it AGI?\n\nI don’t know.\n\nMore importantly, that question may not help me decide what to do next.\n\nIf you want to call your machine AGI, define the observable event that flips your bit:\n\nAGI=0→AGI=1AGI=0\\rightarrow AGI=1\n\nThen show me the test.\n\nOtherwise we’re pricing a noun.\n\nI have a cheaper question:\n\n**What job are we trying to accomplish, what does failure cost, and what does this black box actually do when we put it there?**\n\nRun the experiment.\n\nSave the logs.\n\nChange one thing.\n\nRun it again.\n\nAnd when the machine doesn’t know?\n\nSee whether it can say:\n\n**I don’t know.**\n\nThat might be a more valuable capability than another thousand tokens explaining why it probably does.\n\n*Shitpost by Jun Huang / Shattering The Abyss.*\n\n*The scenarios here are intentionally absurd. The testing question isn’t.*\n\n**What observable evidence would you personally require before putting an LLM inside a workflow where one bad inference can propagate into a consequential human decision?**", "url": "https://wpnews.pro/news/the-1-llm-test", "canonical_source": "https://shatteringtheabyss.substack.com/p/the-1-llm-test", "published_at": "2026-09-20 10:42:52+00:00", "updated_at": "2026-09-22 02:55:18.317006+00:00", "lang": "en", "topics": ["large-language-models", "ai-tools", "ai-products", "ai-research"], "entities": ["YouTube", "Ron \"Rambo\" Kim", "Twistzz", "Counter-Strike"], "alternates": {"html": "https://wpnews.pro/news/the-1-llm-test", "markdown": "https://wpnews.pro/news/the-1-llm-test.md", "text": "https://wpnews.pro/news/the-1-llm-test.txt", "jsonld": "https://wpnews.pro/news/the-1-llm-test.jsonld"}}