Archive / Research & notes
A 1945 essay proposed linking personal records by association
Vannevar Bush's essay describes an imagined memex and associative trails, not a machine that was ever built.
3 min readPRESERVED ARCHIVE
How the work got here.
One hundred retrospective records: original ideas, product changes, setup constraints and evidence. Preserved from the September 2026 archive, with its source dates and caveats intact.
No matching pieces. Clear the search or choose another topic.
Archive / Research & notes
A 1962 report framed computers as tools to augment thinking
Engelbart's SRI report defines augmenting human intellect well before any product built on it existed.
3 min readArchive / Automation & agents
A 1960 paper argued computers should share decisions with people
Licklider's Man-Computer Symbiosis proposes cooperative decision-making, not machines resembling language models.
3 min readArchive / Automation & agents
A 1966 paper warned about its own illusion of understanding
Weizenbaum's ELIZA paper explains the pattern-matching script and flags how readily users assumed it understood them.
3 min readArchive / Meetings & email
Ray Tomlinson forgot the words of the first network email
Tomlinson's own account says the first networked email, sent in late 1971, used forgettable test text he never recalled.
3 min readArchive / Writing & documents
Xerox PARC's Bravo tied the screen to the printed page
Computer History Museum materials describe Bravo as the research system that introduced WYSIWYG document editing.
3 min readArchive / Coding & data
VisiCalc paired a visible grid with instant recalculation
Dan Bricklin's own site describes VisiCalc's 1979 debut and the point-and-recalculate model spreadsheets still use.
3 min readArchive / Setup & friction
The Xerox Star sold a desktop metaphor few offices bought
Surviving Star 8010 documentation shows a realized desktop system that its own licensing kept locked to Xerox.
3 min readArchive / Setup & friction
A 1985 congressional report predicted office jobs would stall
The OTA's 1985 study forecasts office automation's effects to 2000 while flagging its own forecasting tools as weak.
3 min readArchive / Evidence
A 1993 paper named the computer productivity paradox
Brynjolfsson's paper defines the productivity paradox and tests four explanations for it, not a resolved verdict.
3 min readArchive / Writing & documents
Grammarly's generative writing tool shipped switched off by default
Grammarly's archived 2023 launch page shows GrammarlyGO arrived off by default for Business and Education accounts, gated behind an admin toggle.
3 min readArchive / Writing & documents
Atlassian's AI layer turns on by default once a plan qualifies
Atlassian's own admin and trust documentation, read on 16 September 2026, show Atlassian Intelligence auto-enabling on paid plans unless an admin turns it off.
3 min readArchive / Writing & documents
Adobe says Acrobat's AI Assistant never trains on your PDF
Adobe's own product and legal pages, read on 16 September 2026, describe Acrobat's AI Assistant and disclose that prompts may still be reviewed for abuse.
3 min readArchive / Writing & documents
A content credential records a process, not whether it is true
The C2PA specification, read on 16 September 2026, says validation confirms a claim's binding to a file, not whether the claim itself is accurate.
3 min readArchive / Writing & documents
Word's Copilot needs a license, a role, and up to a day to appear
Microsoft's own setup documentation, current as of 16 September 2026, lists the licensing, admin roles, and security steps before Copilot appears in Word.
3 min readArchive / Writing & documents
Google Docs put a generative draft button next to spellcheck
Google's 2023 rollout post shows Help me write launched as a Duet AI Enterprise add-on that could not be turned off once enabled.
2 min readArchive / Writing & documents
Notion's Autofill turns a table column into an AI setup decision
Notion's help documentation, read on 16 September 2026, sets Autofill's plan requirements, run triggers, and its own accuracy warning.
3 min readArchive / Writing & documents
Coda folded its AI column into the price of making a doc
Coda's product and help pages, read on 16 September 2026, describe the AI column as bundled into the plan, with a caution about trusting its output.
2 min readArchive / Writing & documents
Ironclad's AI flags contract risk, then hands the call to a lawyer
Ironclad's own product pages, read on 16 September 2026, describe AI-assisted redlining against a trained playbook, with human sign-off kept explicit.
2 min readArchive / Writing & documents
Docusign now reads the contracts it used to just route for signing
Docusign's own IAM page and Iris blog post, read on 16 September 2026, describe post-signature extraction drawn from a company's full agreement history.
3 min readArchive / Meetings & email
Otter's meeting-notice email is a toggle, not a default
Otter's own settings documentation shows which recording notices are opt-in and which an admin can force.
3 min readArchive / Meetings & email
Fireflies follows each meeting platform's own consent rule
Fireflies' own compliance guide shows recording consent differs by video platform, with separate notices a host must enable.
3 min readArchive / Meetings & email
Granola's notice is a watermark, because there's no bot to see
Granola's own documentation shows local audio capture replaces a visible bot with two optional, off-by-default notices.
3 min readArchive / Meetings & email
Gmail's Help me write launched behind a paid Workspace add-on
Google's own 2023 launch note and current help page show the drafting feature's access rules broadened, but never disappeared.
3 min readArchive / Meetings & email
Outlook's Copilot splits into a licensed tier and a basic one
Microsoft's own pages show thread summarizing exists without the paid add-on, while priority access and Graph data stay behind it.
3 min readArchive / Meetings & email
Front drafts replies and shows exactly which past ticket it used
Front's own help pages show suggested replies cite the source conversation, and can be reviewed or sent autonomously under Autopilot.
3 min readArchive / Meetings & email
Calendly's routing form is conditional logic, not an AI agent
Calendly's own help pages describe hand-built rules that route a qualified visitor to a person's calendar, nothing more.
3 min readArchive / Meetings & email
Fathom keeps transcripts free and sells the summary layer
Fathom's own pricing page shows recording and transcription are free forever, while action items and cross-team search are paid.
3 min readArchive / Meetings & email
Read AI turns a meeting's talk time into a shared score
Read AI's own privacy policy shows its Read Score is shared with other participants, unlike its private Speaker Coach metrics.
3 min readArchive / Meetings & email
Slack's huddle notes need someone to switch them on each time
Slack's own FAQ shows AI huddle notes require a manual toggle, but the resulting canvas is visible to the whole channel.
3 min readArchive / Automation & agents
OpenAI's first browser agent paused itself before purchases
OpenAI's own announcement and safety card show why Operator asked a human to confirm before finishing consequential actions.
3 min readArchive / Automation & agents
Anthropic called its own screen-controlling model error-prone
Anthropic's launch post and tool documentation both name specific computer-use failure modes rather than smoothing over them.
3 min readArchive / Automation & agents
ChatGPT's agent mode still asks before it spends money
OpenAI's launch post and its connector-permissions documentation describe where agent mode pauses for a person and where it does not.
3 min readArchive / Automation & agents
Make added a step that picks its own next move
Make's own documentation distinguishes an AI-agent module from a fixed scenario step and names what still needs a human check.
3 min readArchive / Automation & agents
Relay.app built its workflow tool around a human approval step
Archived vendor pages show approval, task and data-input steps as Relay's core feature, before the company wound the product down in 2026.
3 min readArchive / Automation & agents
Lindy's AI teammate waits for approval before anything ships
Lindy's own site and documentation describe a no-code agent builder that still routes consequential actions through a named approver.
3 min readArchive / Automation & agents
UiPath aimed generative AI at the labor of building RPA bots
UiPath's own 2023 announcement describes Autopilot as text-to-workflow generation for existing robotic automation, not a new agentic product.
3 min readArchive / Automation & agents
Salesforce's Agentforce shifted its pitch from assisting to acting
Salesforce's own announcement describes autonomous service and sales agents that hand off to a person with a summary, not resolve everything alone.
3 min readArchive / Automation & agents
Workato tied its new agent builder to its old connector permissions
Workato's own documentation grants agents only OAuth-authorized, admin-approved connector access, with human sign-off routed through Slack or Teams.
3 min readArchive / Automation & agents
ServiceNow's AI agents triage tickets but wait on set guardrails
ServiceNow's own product pages describe agents built to resolve or triage routine tickets while an admin defines when their actions need approval.
3 min readArchive / Research & notes
Elicit measured how often its own summaries invent claims
The vendor's own hallucination study puts a number on a risk most literature-review tools only gesture at.
3 min readArchive / Research & notes
A yes-or-no meter for scientific claims comes with no error rate
Consensus links every claim to a peer-reviewed paper but does not publish a measured hallucination rate for its own meter.
3 min readArchive / Research & notes
A citation checker labels a claim as supported with no audit trail
Scite classifies citations as supporting or contrasting but its own pages do not document how, or by whom, that call is made.
3 min readArchive / Research & notes
Semantic Scholar shipped a summary feature it could not prove worked
The nonprofit's own blog says its A/B tests found no conclusive evidence its one-sentence summaries helped users at all.
3 min readArchive / Research & notes
Glean's search tool inherits whatever permissions were already wrong
Glean's own security page says it enforces existing access rules rather than fixing them, and pricing is not published.
3 min readArchive / Research & notes
Tana split into two products and its notes tool got a new address
The structured-notes app once called Tana is now Tana Outliner at a separate site, with AI features metered by shared credits.
3 min readArchive / Research & notes
Reflect keeps notes encrypted until you ask the assistant something
Reflect's marketing page names one AI provider; its privacy policy discloses three, plus a training-data caveat.
3 min readArchive / Research & notes
Mem rebuilt itself again around a chief-of-staff style AI agent
Mem's chat feature searches the whole workspace by default, but its own FAQ leaves the model choice and empty results undocumented.
3 min readArchive / Research & notes
Wikipedia's enforceable AI rule is not the page people cite
The often-linked Large language models page is explicitly non-binding; the real content guideline bans generated prose outright.
3 min readArchive / Research & notes
ArXiv's 2023 AI rule was about authorship, not banning drafting help
ArXiv required disclosure and barred naming AI as an author; its current policy still reads almost the same way.
3 min readArchive / Coding & data
ChatGPT gained a firewalled Python sandbox for uploaded files
OpenAI's own materials describe Code Interpreter's July 2023 rollout as a sandboxed, firewalled Python environment for uploaded files.
3 min readArchive / Coding & data
GitHub's Copilot Workspace plans a change before writing code
GitHub's own announcement describes a step-by-step, fully editable plan stage that still ends in ordinary human code review.
3 min readArchive / Coding & data
Amazon Q Developer folded CodeWhisperer into a wider assistant
AWS's own GA post describes automatic Java upgrades alongside a plan-then-insert step for building new code.
3 min readArchive / Coding & data
Google stopped serving its free Gemini coding assistant tier
Google's own documentation records the free Gemini Code Assist tier's June 2026 shutdown while Standard and Enterprise pricing continued unchanged.
3 min readArchive / Coding & data
JetBrains still prices AI access apart from the IDE
JetBrains' AI pages describe a subscription or bring-your-own-key layer on the IDE license, with a 2026 post disclosing a tenfold rise in its own AI spend.
3 min readArchive / Coding & data
Hex's AI agent needs an admin to mark data as sensitive
Hex's own pages describe AI context from semantic models and metadata, and an admin control to keep some connections out of agent threads.
3 min readArchive / Coding & data
Julius AI computes in an isolated sandbox with a RAM ceiling
Julius's own FAQ names the file types it accepts, the sandbox's memory limit, and a documented tendency for the assistant to hallucinate.
3 min readArchive / Coding & data
dbt Copilot's 2024 debut became dbt Wizard by 2026
dbt Labs' 2024 launch post and its current product page, now describing dbt Wizard, both keep a human-review step before anything ships.
3 min readArchive / Coding & data
Sourcegraph's Cody docs still stand while its blog moves on
Sourcegraph's docs describe Cody's codebase-wide context and exclusion filters, even as its 2026 blog posts describe newer agent products instead.
3 min readArchive / Coding & data
Stack Overflow's 2023 AI search became a 2026 agent exchange
Stack Overflow's 2023 launch post and its 2026 update trace OverflowAI's shift from grounding chat answers to serving AI agents directly.
3 min readArchive / Setup & friction
Italy's privacy regulator made OpenAI change ChatGPT to return
Three Garante documents show what the regulator found, what it demanded, and what it confirmed OpenAI changed before Italian access resumed.
3 min readArchive / Setup & friction
GitLab's public handbook splits AI rules by team, not one policy
Two public GitLab handbook pages, not the security team's gated one, show how Support and Engineering actually restrict AI use on company data.
3 min readArchive / Setup & friction
Windsurf lost its founders and its owner in the same week
Windsurf's and Cognition's own blog posts document a leadership departure and an acquisition landing within days of each other in July 2025.
3 min readArchive / Setup & friction
A wiretap statute from 1967 is being tested on an AI notetaker
A filed federal complaint and its docket show what is alleged against Otter.ai's meeting bot and that no court has yet ruled on it.
3 min readArchive / Setup & friction
A crowdsourced list tracks who charges extra for basic login security
The SSO Wall of Shame documents vendors gating single sign-on behind top tiers, and a live pricing check shows the gap can be worse than listed.
3 min readArchive / Setup & friction
Microsoft dropped its 300-seat floor for buying Copilot
Microsoft's own January 2024 announcement removed the seat minimum for Copilot for Microsoft 365, without changing its per-seat price.
2 min readArchive / Setup & friction
Okta wants AI agents to be identities, not shared secrets
Okta's own documentation proposes short-lived, revocable credentials for AI agents in place of the hardcoded API keys many teams use today.
3 min readArchive / Setup & friction
Anthropic's commercial terms promise not to train on API data
Anthropic's separate commercial terms bar training on customer content and describe a conditional, not absolute, data-destruction commitment.
3 min readArchive / Setup & friction
1Password treats an AI agent as an identity that needs a vault
1Password's 2025 announcements name standing, hardcoded agent access as the setup risk its credential-delivery features are built to reduce.
3 min readArchive / Setup & friction
Chrome's extension rules do not mention AI by name
Chrome Web Store's permissions and disclosure policies govern AI sidebar extensions without naming AI, machine learning, or LLMs anywhere in their text.
3 min readArchive / Evidence
A model maker's own usage index counts conversations, not the economy
Anthropic's first Economic Index sorts a million Claude.ai conversations by occupation but says it cannot show AI use off its platform.
3 min readArchive / Evidence
A government survey put business AI use at 3.8 percent
The Census Bureau's business survey found 3.8% of firms used AI to produce goods or services in late 2023, with wide gaps by sector.
3 min readArchive / Evidence
Consumer ChatGPT use skews toward personal tasks, a paper finds
An OpenAI and outside-economist working paper classifies consumer ChatGPT conversations and finds personal use now dominates work use.
3 min readArchive / Evidence
A regional Fed survey now asks CFOs what they spend on AI
The Richmond Fed's CFO Survey added an AI spending module, a business-conditions read distinct from a vendor's own customer survey.
3 min readArchive / Evidence
Stanford's AI Index reports numbers it credits to other sources
The annual AI Index compiles labor-market and economy figures from named outside providers rather than measuring them itself.
3 min readArchive / Evidence
McKinsey's AI survey samples opinion, not a random slice of firms
McKinsey's own methodology note describes an opt-in online survey of 1,719 executives, not a probability sample of the economy.
3 min readArchive / Evidence
PwC's Jobs Barometer reads a billion job ads, not opinions
PwC's AI Jobs Barometer scores occupations for AI exposure and tracks wage and revenue data by exposure quartile, not sentiment.
2 min readArchive / Evidence
Gallup tracks what workers say about their own AI use
Gallup's quarterly panel poll put US AI use at 52% of workers in their own role, a self-reported figure, not a logged usage count.
3 min readArchive / Evidence
The Fed's business anecdotes now regularly mention AI
The Federal Reserve's Beige Book compiles business contacts' own words on AI's effect on hiring and demand, district by district.
3 min readArchive / Evidence
Employers told the WEF what they expect AI to do by 2030
The World Economic Forum surveyed over 1,000 employers on AI's expected effect on jobs, a forecast, not a measured outcome.
3 min readArchive / Coding & data
Tabnine's air-gapped option still needs one outbound connection
Tabnine's own deployment docs show on-premises coding assistance is real, but a container-registry connection and a locked model list come with it.
3 min readArchive / Coding & data
An open coding assistant leaves the model choice to the team
Continue's own docs show model choice, including local models, is real, but agent mode still favors models with confirmed tool support.
3 min readArchive / Coding & data
Snowflake keeps its SQL copilot's answers inside its perimeter
Snowflake's documentation states Cortex Analyst runs on models it hosts itself, so a query never leaves the account's governance boundary.
3 min readArchive / Coding & data
Databricks Genie's table list curates answers, not access
Databricks documentation shows Genie's attached tables shape its answers, while Unity Catalog permissions remain the real access boundary.
3 min readArchive / Coding & data
Colab's free AI coding help comes with an 18-month data window
Google's own FAQ says free-tier Colab AI features feed human review and are stored for up to 18 months, separate from Colab Enterprise.
3 min readArchive / Evidence
METR's benchmark tracks AI task length, not messy real work
METR's paper found the length of software tasks AI agents can finish has doubled roughly every seven months, with real-world caveats attached.
3 min readArchive / Evidence
Epoch AI's compute numbers are mostly estimates, and it says so
Epoch AI's own methodology shows most training-compute figures in its tracker are estimated, not measured, and it labels the difference.
3 min readArchive / Evidence
CEOs now rank AI as their second-biggest business worry
The Conference Board's own Q3 2026 survey shows AI and new technology overtaking geopolitics as CEOs' second-highest named business risk.
3 min readArchive / Evidence
LinkedIn's AI skill data comes from what members type in
LinkedIn's own methodology defines AI talent by self-added skills or job titles, scoped to member profiles in roughly fifty countries.
3 min readArchive / Evidence
GitHub counts an AI project by its code imports, not a survey
GitHub's 2025 Octoverse defines an AI project by dependency signals in the code itself, separate from any developer-sentiment question.
3 min readArchive / Writing & documents
Airtable's AI fields bill by length, not a flat per-field price
Airtable's own help center says AI field credit costs scale with prompt length, and batch runs can exhaust credits before finishing.
3 min readArchive / Writing & documents
AI blocks on monday.com are billed in credits, not seats
monday's own support articles show AI blocks cost eight credits per action, metered apart from the per-seat plan a workspace already pays.
3 min readArchive / Writing & documents
ClickUp prices its AI assistant as a second subscription on top
ClickUp's own pricing page adds Brain AI or Everything AI as a per-seat charge stacked on every plan tier, not gated to the top one.
3 min readArchive / Writing & documents
Craft's assistant can search a whole space, not just one document
Craft's own help center describes AI reach across documents and the space, plus a session quota that editing spends faster than reading.
3 min readArchive / Writing & documents
Dropbox bet Dash's cross-app search on data staying self-hosted
Dropbox's own 2024 launch post says Dash for Business searches connected apps while keeping AI processing inside Dropbox's own infrastructure.
3 min readArchive / Meetings & email
Missive's AI assistant reads a thread's internal comments too
Missive's own docs show the assistant sees internal chat and notes in a thread, billed through an org-level provider a rep does not pick.
3 min readArchive / Meetings & email
HubSpot's AI email writer still needs the audience typed in
HubSpot's own knowledge base shows its assistant drafts from typed prompts, not an automatic pull from the contact's CRM record.
3 min readArchive / Meetings & email
Spike summarizes threads on open, with no visible opt-out
Spike's own site and blog describe automatic thread and attachment summaries but no documented setting to turn the behavior off.
3 min readArchive / Meetings & email
SaneBox calls its decade-old header sorting machine learning
SaneBox's own FAQ credits machine learning for its long-standing folder sort, while its Daily Digest stays an interactive list, not prose.
3 min readArchive / Meetings & email
Gmail's thread summaries are click-triggered, not automatic
Google's own help articles show AI Overview summaries run on request and reach personal accounts more broadly than other Gemini features.
3 min read