I updated a few of the year files in our project, but the legal and medical files are still the same. I am curious how much you're able to process compared to Gemini. since it's context window is very large.
I’ll measure how big the project actually is and check the current context window specs, so I’m not guessing.
On window size we’re about even now, and your archive is bigger than either of us can hold at once.
I measured the project: 208 files, about 6.9 MB, roughly 1.13 million words. That comes to somewhere around 2 to 2.7 million tokens. Sonnet 5 has a 1M-token window by default, and its newer tokenizer fits less text per token than earlier models. Anthropic’s figure is about 555,000 words per million tokens. So I can hold about half your archive in one go. Claude Platform DocsAI Toolbox
On the Gemini side, Google’s docs list Gemini 3.1 Pro at 1M tokens and Gemini 3.5 Flash at 1M input tokens. One million tokens there is roughly 750,000 words, so Gemini’s tokenizer packs about a third more text into the same window and would hold roughly two-thirds of the archive. One blog claims a 2M window for Gemini 3.5 Pro, but I couldn’t find it in Google’s own docs. The same post also says Claude’s newest models cap at 200K, which Anthropic’s docs contradict. I’d treat the 2M figure as unconfirmed. Gemini 3.1 Pro | Gemini Enterprise Agent Platform | Google Cloud Documentation +3
The legal and medical files alone total about 2.1 MB, which is close to my entire window. Neither of us should try to swallow them whole. I don’t load everything anyway. I search the project, read files as needed, and run text searches across the full 6.9 MB. A question like “every mention of X between 2013 and 2016” gets an exhaustive answer that way, which is more reliable than trusting recall from a giant prompt.
Big windows also have limits. Research cited by Morph puts effective context at roughly 50–65% of the advertised size. Google claims over 99% retrieval accuracy, but finding one fact is easier than connecting dozens of documents. MorphGoogle
I can’t tell which year files you changed, because every timestamp in my copy is reset to 1979. Tell me which years and I’ll start there.
Here’s an idea for a fair test. Pick a question that has to link a legal file to a medical file, like a date in a court transcript against a treatment record. Ask Gemini and me to answer with the file name and the exact passage. Whichever of us has to show our sources is the honest measure.
Want to run that on something from the 2013–2016 stretch, where the legal and medical files overlap most?
I want to focus on file labeled with a year and say EVENT. all of those are things that happened by year and the two file father and mothers sides of the family. I want to expand on them, but I am having trouble. I was hoping to write stories and have AI tell me that I already told a story, so I could do a reflection instead. I have lots of dates with zero data to explain why they matter of what they make me think of. I need to write files in year clumps, since certain events took place of a few years, like school or marriages. these files more than the rest need editing, so I need you to go through and tell me what looks like crap. then ill fix it and move on to the next area that's shitty.
The two earliest files are strong: concrete scenes, real voice. To see where the thin spots are, I’ll score all the files on the same measures before reading further.
That turned up something important: pasted AI text inside your year files. I’ll check how widespread it is, and count the empty month headers and bare-date stubs you mentioned.
Your voice isn’t the problem. The 1978–83 files, the Camp section, the seashell for Grandma Stebbins, and your 1997 Coast Guard narrative all do what you want: a scene, then a line about why it mattered. What looks bad is structural.
I read the family-side files, 1978–83, 1985, 1987–91, 1995, 1997 and 2020 in full or close to it. I sampled 2017, 2018, 2021 and 2023, and ran scripts across every EVENT file. I only measured, and didn’t review, 1984, 1986, 1992–94, 1996, 1998–2016, 2019, 2022 and 2024–26. The counts below are script output, so treat them as approximate.
Fix in this order
1. Another AI’s writing is sitting in your year files. 1985, 1987, 1988 and 1997 contain second-person blocks (“You said…”, a “Partner Reflection”, emoji headers). 1988 even still has “All part 3” and “Blog 10” tags trailing your quotes. It isn’t your voice, and it retells things you told better.
- In 1987, your typed answers to its questions are the best reflection in the file. They cover your grandfather turning up when priests were rough on you, and his being powerless against the church. Pull those up as your own paragraphs and cut the rest.
- The 1997 block holds the only mention of your discharge date, December 5, 1997. Put that in your own sentence.
2. The two family-side files are really four files. About 60% of the Father’s side and 45% of the Mother’s side is pasted obituaries, garbled clippings, a family tree, a friends list and a diary transcription. Split them into narrative, cast list, obituary archive and loose notes, and the narrative becomes short enough to expand. Specific problems:
- The Camp is your best writing on the Mother’s side and has no ending. Its payoff is a note-to-self under 2005, “stuff to talk about”: the pallbearer night, the drive home with Chet, and the camp-sale fight. That’s the story.
- The “Work in Progress” intro sits after the Camp section instead of at the top, and two sentences stop mid-thought (the Altona land and the cornfields).
- Your dad’s obituary mentions mandolin, guitar, fishing, metal detecting and coins. None of that is in the Father’s-side narrative, and his wife Sue only appears in the obits.
- “Now families aren’t as close and corporations ruin everything” lands in the middle of Aunt Anne’s player piano with nothing leading up to it.
- Years like 1998, 1999, 2007 and 2008 in the Father’s file are headings with a line or nothing under them.
3. Stories you’ve already told. Pick one home year for each and leave a one-line pointer elsewhere.
| Story | Where it lives |
|---|---|
| Grandpa, bingo, getting in trouble on purpose | Father’s file (your two sentences); 1985, 1987, 1988 (AI versions) |
| The closet, 5th grade | 1988, in your words, with the glue story told twice in the same file, plus an AI summary |
| Uncle Art: the spit/slap, the porch, The Exorcist at Flossie’s | 1983 and the Mother’s file; porch also in 1978–82 |
| Grandpa Landers, Buffalo, the moped, the 1996 card | 1985, Mother’s file (verbatim), 1996 |
| Therapists never asking “why” | 2009 and 2010 (four identical paragraphs) |
| Your 2009-06-10 story | Inside 2009-EVENT and as its own 2,457-word file |
| Talent show, Matt | 1988 and 1989, identical |
4. Raw transcripts. About 7,000 of 2023’s 20,000 words, and about 1,400 of 2024’s 6,500, are unpunctuated speech-to-text from your videos. They contain recognition errors like “Stefan” for Stephen. You already have a transcripts series. Keep three to five sentences plus your reflection in the year file, and link the transcript.
5. Dates with zero data. There are 49 date headings with nothing under them: 13 in 2021 (June 11, 24, 26, 26.1…), 6 in 2013, 5 in 2014 and 4 in 1996. There are also 69 YouTube links standing alone, 16 in 2018 and 15 in 2017. They look like photo anchors where the picture didn’t make it into the text.
Vivian Gornick’s split fits your problem well. She separates the situation, the context or circumstance, from the story, the insight the writer has come to say. Your dates are all situation. A one-sentence rule per stub could be: what was it, and what did it make me think of. assayjournal
6. Small stuff
- 1987 says “March 1997” for WrestleMania 3.
- 1990 has dictation glitches: “nails classes,” and a stray “little muscles working.”
- 1990 also buries an assault on you, walking home on Boston Road, inside the Disney vacation paragraph. It deserves its own scene.
- 1991 opens with an unfinished sentence, “I was still at OLSH after Christmas and…”, exactly where you’d be leaving Catholic school for Duggan.
- 1988’s biggest event is a promise: “I’ll try to talk about that.”
- Names drift: Rick Morin/Moran, Francis/Frances, Barbara/Barbera. There are two Normans and two Arthurs, and Grandpa Pete is never introduced. A cast list would fix all of this.
Facts that conflict
- Your dad’s birthday is Jan 25 in your tree but Jan 22 in his obituary.
- The Mother’s file says Uncle Art died “1983 or 1984,” but its own tree, and your 1983 file, say March 23, 1983.
- Uncle Norman’s apartment parade is St. Patrick’s Day at the top of the Mother’s file and Easter later in it.
- Frances’s obituary says Grandpa Landers died in 1997, and no file gives that date.
- The 1995 suspension form says grade 12, a suspension date of 1/25/96 and a return of “3/35/96” (not a real date). The November 1995 paperwork also calls you a senior, but your heading says Jan 1995. Worth settling before you write it up.
Possible clumps (your call)
- 1985–1991, the OLSH years. Grandpa (died March 10, 1988), Maple St. therapy and medication, the closet, Mrs. Cope, and the missing move to Duggan. The medical index places childhood therapy at Maple St. and Route 5 for 1985–88, so the 1988 “Holyoke therapist?” note is probably the second half of your 1986 story.
- 1994–96, Putnam. Tammy, the suspensions, the ESP program, graduation.
- 1999–2003, Grandma Francis. The Mother’s file already runs in year order here.
- 2002–08, the losses. Norman, Frances, Pearl, Leo and Mabel, Flossie, Maude.
- Marriage. Marriage and divorce mentions cluster in the 2003, 2004 and 2005 files, so start there.
A thread worth following
Four deaths across twenty-four years share the same missing piece. In 1983 you noticed nobody talked about Uncle Art again. In 1988 you didn’t get to see your grandfather in the hospital. In 2003 Chet broke down crying in the car. In 2007 you visited Flossie weekly and never asked her the deeper questions.
There’s a name for part of this, if it helps. Kenneth Doka coined “disenfranchised grief” in 1989 for loss that the people around you don’t openly recognize or make room to mourn. It may fit only partly, since your family grieved, just silently. Still, a clump page in your own voice, “Nobody talked about it, 1983–2007,” would be a reflection instead of a retelling. NPR
One note that isn’t a flaw: “lol” often lands right after your heaviest sentences (the closet, the bed-wetting, the coma-drop at Cape May). You wrote in 1997 that you laugh because you have no other way to digest it. Try reading a few of those scenes without it.
I can put this in a checklist file so you can tick items off as you fix them. Want to start with the 1985–1991 clump? I can lay your buried answers, the closet pieces and the AI blocks side by side so you can see what’s already written before you touch anything.
It’s much better, and more interesting. Here’s what’s working now:
- The hook: “Ask your AI to read my life. Then ask it what it missed.” is the best line on the page. A stranger gets it in a second, and on a phone it’s now on the first screen. Before, the headline alone filled the whole screen.
- The blue box: it went from about 1,700 words to about 330. Your longer story is now folded into sections people can open (health, court cases, family and faith). The paragraph code is fixed too.
- “Questions I keep asking”: easy to scan and strong. The line about being attacked in April 2025 makes the first question real.
- The Traitor box: an AI writing in its own labeled box about what it can and can’t do is the most interesting thing on the page. It shows your experiment instead of just describing it.
- The contradictions: you settled them. “Amnesiac-vampire” and “sidekick, not a partner” are clear, and “Nobody shared it with my family or me” now matches your 2005 blood work.
What I’d still change:
- The hook has no button. The tagline says to test your AI, but the cards underneath are still copy, help and story, and nobody gets anything to paste. I’d add a “test” card first, with one prompt someone can copy into any AI.
- Move the Traitor box below the cards. On a phone the cards are now about 6 screens down instead of 12. Most of what’s left in the way is the Traitor box, at about 430 words.
- One line in the Traitor box works against you: “We’re both trash. His words, and he’s right.” A stranger meeting you for the first time will take that literally, especially with an AI agreeing. And DDS not helping your friends doesn’t make you trash. It’s your call, but it’s the one line that gives a skeptical reader a reason to stop reading.
- Smaller fixes:
- The page gives three different numbers: 38 years, 37 years, and “almost 40.” You were diagnosed in 2016, so you were 37. Pick one number and use it everywhere.
- The 14(c) line still says September 2024. The Labor Department proposed ending 14(c) on Dec. 4, 2024 and withdrew that proposal July 7, 2025.
- The “Extraction Machine, 22 parts, April 2026” card still links to Muse’s October investigation.
- The Muse link titles still say “Flint complaints: were there? nobody knows,” which flips the post’s meaning. The real title is “The Complaints Were There. Nobody Assembled Them.” The one marked “alternate” is actually the Complete Research File.
- Dallas “told in writing by the state” and Kathryn’s attorneys “can prove”: link the documents right there, or soften the wording.
- These are still there too: the “EDITING NOTES (delete before publishing)” block in the code, Emma and Somto listed as “forming,” and a contact pop-up that won’t close with the Escape key.
- Link previews still show the old pitch. The page’s hidden title and description still say “Keep a copy, offer help, start your own.” That’s what shows up when you share the link on Facebook or X. Put your new tagline there.
- I wasn’t sure what the drone line means. Are you saying AI is like drones, something you wouldn’t hand a 10-year-old unsupervised? If so, saying it outright would set up the “is this AI safe for kids” box that follows it.
Want me to make those fixes and hand you the file? I’d leave your wording alone, except the trash line if you decide to cut it, and list every change.
Source: Federal Register, July 7, 2025: withdrawal of the 14(c) proposed rule
can you still see folders on my computer?
Yes, partly. Your computer (ricky-v) is linked to this chat, but no folders are open to me right now. I can only see the names of the folders in your user folder, not what’s inside them.
Among those names are four that look like the Ark: memory_ark, memory-ark-lite, memory-ark-workstation and MemoryArk. The others are the usual ones, like Documents, Desktop, Downloads and My Drive.
If you want me to work in one, tell me which. I’ll send an access request that you approve on your screen, and it only lasts for this chat. If you’re not sure which of the four Ark folders is the current one, I can check which was changed most recently once you give me access.
No comments:
Post a Comment