Honey, AI Shrunk the Archive

Jon Ippolito

Professor of New Media
Director, Digital Curation graduate program
The University of Maine
LearnWithAI.org
📝 "Honey, AI Shrunk the Archive"

🐘 @jonippolito@digipres.club 🦋 @jonippolito.net 🐦/🧵 @jonippolito 🔗 https://www.linkedin.com/in/jonippolito
Digital Curation Logo Xvga

LearnWithAI.org

IMPACT RISK framework for understanding AI risks

ai_impact_risk_acronym

AI-Impact-Risk.com

Prior art

Dig Brumfield Teleconf 2025 Ann Xvga

Archive v. AI

Evolution of scholarship

Ai Phone In Archive Library By Cecilia Maas

Cecilia Maas, "Archiving with AI", 2023

Archive v. database

Evolution of scholarship

Database V Archive Backblaze

Lev Manovich, "Database as a Symbolic Form", 1998

Pre-AI archival trajectory

Ai Archive Trajectories Wxga

Post-AI archival trajectory

Ai Archive Trajectories Wxga

"AI is an archive in reverse"—Eryk Salvaggio

How will AI affect scholarship?

  1. 🦜 Animistic AI analogs
  2. ☕️ Mechanistic AI analogs
  3. 📊 Consequences of mechanistic analogy
  4. 🏛 Warping the archival trajectory

Which analog?

Which analog?

1. 🦜 Animistic analogs for generative AI

open_ai_microsoft_sparks_synergetics_india bender_emily_stochastic_parrot

Which analog?

Homogeneity

Deep Neural Network Diagram Duif Cell Diagram Brittanica

Which analog?

Training

Smith Linda Head Mounted Camera Baby 2 By Eric Hanus Vga

Linda Smith, Indiana University

Fedorenko Language Brain Fmri Scan

Evelina Fedorenko, MIT

2. ☕️ Mechanistic analogs

Which analog?

Mechanistic analog 1: Compression algorithm

Museum In Hand Sd Wide

JPEG compression

Jpeg Compression Dspguide B Jpeg Compression Dspguide A

Compression algorithms already work as classifiers

Jiang Gzip Compression As Ai Link

Which analog?

Mechanistic analog 2: A cup of coffee

Steamy Coffee By Waferboard
Temperature Equation Thermodyn Xvga
Temperature Equation Thermodyn Xvga
Temperature Equation Ai Softmax Xvga
Temperature Equation Ai Softmax Xvga
Temperature Equation Thermodyn Xvga
Temperature Equation Thermodyn Xvga
Temperature Equation Ai Softmax Xvga
Temperature Equation Ai Softmax Xvga

3. 📊 Conclusions from Mechanistic AI analogs

Compression smooths over outliers

Before training

Latte Art Before

After training

Latter Art After

Temperature affects probability in surprising ways

doctor_50%_chance_surgery_meme

Opportunistic solutions

julius_caesar_last_breath_getty
Chance of your next breath containing a molecule of Julius Caesar's dying breath: 98.2%
Feature
Definition
Options
Example

📸 Snapshot

A sampling of possible solutions

Many or few good examples

A lungful

🌨️ Density

The concentration of useful answers near the snapshot

Dense or sparse

Caesar's atoms near your lungs

🌡️ Temperature

The variety of possible answers in the snapshot

Hot or cold

Molecular mixing in the atmosphere

The two prerequisites for a "Caesar scenario"

Lungful of molecules

High density

Many possible solutions relative to the number in a snapshot

moving molecules

High temperature

Attempted solutions vary with each snapshot

Ai Molecules Density V Temperature 1080

Problem: insufficient density ("data voids")

verge_obituary_spam

Problem: temperature too low

Ai Covid Research Failure

Problem: temperature too high

Ai Law Court Citations Failure
Ai Density V Temperature Axes Only
Ai Density V Temperature Axes Only

AI task parameter survey

Ai Density V Temperature Sample Tasks
Ai Density V Temperature Sample Tasks
Ai Density V Temperature Emoji
Ai Density V Temperature Emoji
Ai Density V Temperature All
Ai Density V Temperature All

4. 🏛 Warping the archival trajectory

Ai Archive Trajectories Wxga

Improved access to collections

imgs_dot_ai_collection_similar_to_night_watch

Imgs.ai

Image generation in place of search?

leonardo_ai_realtime_generator_horse

Leonardo.ai Realtime Generation

Replacing the entire Internet?

websim_ai_malaysia

Websim.ai

AI-generated artist websites

Ai Generated Art Site Micropolis Ai Generated Art Site Naimark Michael

Archival trajectory

Ai Archive Trajectories Wxga

Archival loop

Ai Archive Trajectories Wxga Jpeg Compression Generation Loss
Back to title screen

High v. low stakes as determinant

High v low stakes (as represented by poker chips)

The problem with high v. low stakes

x-ray AI-generated slack comment

Low stakes

Google Olympics Ai Fan Letter Ad

Opportunistic v. prescriptive tasks in movies

cracking_a_safe
cracking_a_safe
cutting_bomb_wire
cutting_bomb_wire

Opportunistic v. prescriptive tasks in warfare

medieval_soldiers_breaching_fortress_alamy medieval_castle_aerial_drawing_carneycastle

Opportunistic v. prescriptive tasks in healthcare

AI buzzwords you shouldn't use

😵‍💫 Hallucinate

⚒️ Fabricate

👩‍🔬 Model

📊 Estimator

🚗 Guardrail

🩹 Bandaid

🤔 Interpretability

🧐 Accountability

Creativity v. trust

ai_replace_designers_handshake_meme
ai_replace_designers_handshake_meme