Which do you trust more: someone who remembers everything, or someone who knows where everything is written?
Thamus would say the first. You have just watched five arguments, and none of them settles it. Whichever lifeline you used, you are the one who says final answer. Before you give it, decide what a person has to carry in their own head to tell whether what came back is right.
What did the person have to already know to catch the memory answer when it was wrong?
What did they have to know to paste the right text, and the right version of it?
When the regulation was read out with its citation, who in the room could check it?
Pasted context
The model gets this and nothing else. Paste a passage from ecfr.gov, a sponsor’s policy page, an award letter.
Workshop key
The ways ask a language model, and that needs a key. In the workshop, scan the QR code or open the short link and the key is loaded for you. Otherwise paste one here.
How the key is handled
Tab-scoped. The key is held in sessionStorage, which belongs to this browser tab. It clears when the tab closes and is never written to disk.
Static site. This page is served as files from GitHub Pages. No server here stores or sees your key.
Sent only to the model’s gateway. A key starting with mr2_ goes to the University of Idaho’s MindRouter through the AI4RA lakehouse proxy. Any other key is treated as a Google Gemini key and goes to Google. Reading the regulation, the web search and the vector database need no key at all.
URL stripped. A key that arrives as ?key= is read once and removed from the address bar.
Facilitators: pre-fill by sharing a link to this page with ?key=YOUR_KEY appended. Mint a workshop key and revoke it afterward.
The story
In Plato’s Phaedrus, Socrates tells a story about an Egyptian god named Theuth, the inventor of number, geometry, astronomy, and letters. Theuth brings his inventions to Thamus, the king, and argues for each one. When he comes to writing he makes his strongest claim:
This will make the Egyptians wiser and give them better memories; it is a specific both for the memory and for the wit.
Thamus is not persuaded. The inventor of a thing, he says, is not the best judge of what it will do to the people who use it. Then he delivers the verdict that has followed writing ever since:
This discovery of yours will create forgetfulness in the learners’ souls, because they will not use their memories; they will trust to the external written characters and not remember of themselves. The specific which you have discovered is an aid not to memory, but to reminiscence, and you give your disciples not truth, but only the semblance of truth; they will be hearers of many things and will have learned nothing; they will appear to be omniscient and will generally know nothing; they will be tiresome company, having the show of wisdom without the reality.
Plato, Phaedrus 274c to 275b, in Benjamin Jowett’s translation.
The paradox
Thamus warned that writing would ruin memory. He is remembered only because Plato wrote him down. Socrates, who tells the story, wrote nothing. So the dialogue does to Socrates what Plato does to Thamus: a warning against the written record, preserved for twenty-four centuries by nothing but the written record.
The same tension runs through every tool that lets us reach for stored information instead of holding it ourselves: writing, then libraries, then search engines, and now language models. The costs Thamus named are real. So are the benefits he dismissed. The question is never whether to reach for the record. It is who should reach for it, when, and what they must still carry in their own head to judge what comes back.
What it means for a machine
Thamus assumed that memory was the reliable thing and the record was the threat. For a language model it is backwards. Its memory is an enormous, averaged, frozen impression of everything it read, up to a date after which it read nothing. Ask it for the de minimis indirect cost rate and it will tell you ten percent, because that was the rule in most of what it read, and it will not tell you that the rule changed to fifteen percent on October 1, 2024. It is Thamus’s learner exactly: a hearer of many things, appearing omniscient, generally knowing nothing about today.
The remedy is the one Thamus refused. Give the model the written record and let it be reminded. For a machine, Thamus’s flaw is the feature: a model that is reminded is a model that can be checked.
One compliance question. You are the contestant. Five lifelines, and every one of them asks the audience. But you can provide the audience with tools.
You are the contestant
You are the one who has to answer. You can answer from what you know, or you can use a lifeline. The first lifeline is a crowd: a thousand well-read, well-meaning people with no particular expertise. Ask them a question and you get back whatever most of them would say, in one confident voice.
They have read almost everything ever published.
They have read none of your files unless you hand them over.
They answer fast, and they always answer.
There is no minority report. The dissenters are averaged away.
Your lifelines
Answering from your own knowledge uses no lifeline at all, and on the show it is the only answer that counts. All five on this board ask the audience. What changes is the tools you provide the audience before it answers, and what that costs you.
1 No contextAsk the audience2 Pasted contextBring your own notes3 Web searchPhone a friend4 API callAsk the expert5 Vector databaseThe card catalog
Ask the audience
1No context
You ask the crowd and take the majority, confidence percentage and all. Nobody leaves the room and nobody opens a book.
On common knowledge the audience is right most of the time, which is why people trust it. It fails on the obscure question and on the question whose answer changed after the crowd stopped reading, and the percentages look just as confident either way.
Benefit
Instant and free.
Cost
Frozen on a date, and it cannot tell you so.
Bring your own notes
2Pasted context
The lifeline the show never allowed. You hand the crowd a page before you ask, and the crowd answers from that page and nothing else.
It is the most common thing people do in a chat window. The crowd does not know where the page came from or how old it is; you do, or you should. The two fill links in the paste box load 2 CFR 200.414(f) as it reads today and as it read in 2023, so you can hand over either and watch the crowd answer faithfully from the one you chose.
Benefit
You know exactly what it read.
Cost
You have to find the right text, and the right version, yourself.
Phone a friend
3Web search
You hand the crowd a phone. It calls a friend who googles the question while the clock runs. Back come eight headlines with the first lines under each, from whatever pages ranked that day, and the crowd answers from those.
Nobody reads an article through. The show dropped this lifeline in 2010 for exactly this reason: the friend had become a search engine. On this page the friend is a SearXNG metasearch on the AI4RA MCP server, merging Google, Bing, DuckDuckGo, Brave and others.
Benefit
Reaches anything public, including what changed last week.
Cost
Anyone’s page, any year. The crowd referees sources it cannot judge.
Ask the expert
4API call
You hand the crowd the number for the records office. The crowd agrees on two or three words to ask for, the clerk finds the section and reads one paragraph aloud, and the crowd answers from that paragraph, citation and date included.
What it actually is: the model fetching from an external source while it works. The model is given tools it can call, the way an app calls an API. Here the tools come from an MCP server (Model Context Protocol, an open standard for handing a model tools) that wraps the eCFR’s own public API at ecfr.gov. The model calls ecfr_search with its two or three words, the server returns the matching sections of 2 CFR 200, the model calls ecfr_get_regulation on the top one, and when the section is long it picks the one paragraph to fetch. Only then does it answer, from text that arrived seconds ago.
The same pattern works for any source with an API: SAM.gov, NIH RePORTER, a sponsor’s policy site, your own grants system. It is the slowest lifeline and the one with the most steps that can fail, and the only one whose answer changes when the source changes rather than when the model does. Every call it made is listed under the answer.
Benefit
Current, cited and checkable to the paragraph.
Cost
Slowest and most fragile: several calls, each of which can fail or read the wrong paragraph.
The card catalog, or the 50:50
5Vector database
You hand the crowd a card catalog. Before the show, 2 CFR 200 was cut into 1,103 passages and each got a card, filed by meaning rather than alphabet (a vector from an embedding model). The crowd files the question the same way, pulls the four closest cards, and answers from those passages.
Like the 50:50, it narrows the field rather than answering, and it can leave you with cards that circle the answer without landing on it. The catalog lives in your browser; only the question goes out, to be filed by the same model that filed the cards. It is only as current as the day it was filed.
Benefit
Fast and cheap: the whole corpus searched by meaning, the ranking done in your browser.
Cost
Only as current as the day it was built, and a near miss looks like a hit.
Not on the board
Structured data over SQL. When the record is a table rather than prose, the crowd should write the query and let a database compute the answer. Data Crawler Carl does that for any CSV, in the browser, and is the sixth way in the workshop.
Files in reach of a local agent. Tools that run on your own machine, such as Claude Code, Cursor or Copilot, let you name a file with @path/to/file and the agent reads it into context. Same as handing over your notes, with the fetching done for you, and the same obligation: you still have to know whether that file is the governing one.
Long context. Hand the crowd the whole of 2 CFR 200 and ask. It fits in some models’ windows now, at a cost in time and money per question, and the crowd still has to find the paragraph.
What leaves your browser
Your question and the retrieved text go to the model’s gateway: the University of Idaho’s MindRouter through the AI4RA lakehouse proxy with an mr2_ key, or Google with a Gemini key.
Your question goes to the AI4RA MCP server for the web search and for reading the regulation. The server keeps no keys and logs none.
For the vector database, your question goes to the same gateway to be embedded, by the model that embedded the passages. The passages and the ranking stay in your browser.
Same model, same question, default settings, one system prompt, every time; under every answer on the page, What the model was handed shows exactly what came back. This is a static page on GitHub Pages with no server of its own. Source, build scripts and the index are on GitHub; the MCP server is ai4ra-mcp.
