Companion to What Kind of Attention-at-Scale We Build with AI
Aidpage
Aidpage (2005–2014) was the longest-running of the 4 projects I'll be talking about in this piece. A social support network for people in need to talk about deeply personal problems – financial, medical, domestic abuse, depression, etc. – the type of problems most of us don't want to open up about with family or friends… so Aidpage was proposing talking to strangers. A problem shared is a problem halved – was one tagline; People Helping People – was another. Pseudonymous by design. Zero tolerance for abuse. UI/UX built for late web adopters at lower educational and socioeconomic levels. By 2011, ~350,000 monthly unique visitors, mostly low-to-middle income – homemakers, unemployed, disabled, retired. An emerging corps of members who called themselves "Aidpage angels" was basically running the place.
The business model (bootstrapped, zero external funding) was contextual ads via AdSense. I want to state this plainly: Aidpage was effectively funded by attention-capture (aka ad) economy because our users were in moments of need. We reconciled the tension by (1) assuming Google AdSense is maintaining the ad operation's integrity, (2) putting the majority of the ad slots on public facing pages and exposing members of the site to very few ads, (3) offering ad-free private messaging to members.
And (4) – we built a Revenue Sharing program. Active members would receive a portion of the ad revenue generated by their participation. We thought: if Aidpage is an attention exchange platform (which was explicitly stated), the attention generators should share the economics.
It worked too well and then not so well. One active member bought a car with the income. Another – someone who posted a single-sentence question, never returned, and whose post generated 5,000 replies and years of search traffic – would have been collecting what amounted to a monthly salary – for one sentence trigger of Google's search and AdSense algos. So we built a complex algorithm to evaluate contributions more fairly. The hate that came back from the community when we introduced the change was intense. After a year or two of trying to evolve it, we had to discontinue the Revenue Sharing program. It had become a source of conflict that was eroding the trust between members and between members and team.
The Revenue Sharing program crisis was a result (I think) of a widely shared belief about the "ownership" of "my" data generated in a conversation. I always thought the opposite of that belief – you can't own "your" part of a conversation (out of context, it would be meaningless anyway).
Separately, the bad apples problem. A normal range of unhappy Aidpage users plus the occasional real sociopath. Some easy to spot and remove. Some took years to identify. The monitoring was a 24/7 job. Pseudonymity, which the care function required (vulnerable people need it), was exploitable by bad actors. There was no clean resolution. We operated in that tension.
We never came up with good intrinsic viral mechanics (like Facebook's natural family/friends attention graphs). Aidpage was dependent on content-based search traffic from Google. Thus, the site was effectively ended (~2014) by a change of Google's search algo that devalued conversational web content.
GoFundMe launched in 2010 (5 years after Aidpage) as the "Facebook" of "People Helping People"... based on the well-known mental model of fundraising and the primary attention graph (and virality) of family and friends. Still hugely successful.
On a personal level, the Aidpage experience was a mixture between seeing "people helping people"... and the sad moments – longtime members who simply stopped posting, and we'd never know what happened. Some died of the illnesses that brought them to Aidpage, and we would hear from family afterward. Governance includes pain... as it should – if you care.
People Networks
People Networks (2007–2008) was the platform ambition underneath Aidpage. A social platform for non-geeks, explicitly framed as an attention market. Federated identity, user-controlled attention data, cross-network portability of social graphs, user-controlled premium services sold on their own spaces. A framework that roughly anticipated what's now called open trust infrastructure. We didn't manage to fund it and ship it.
JuiceTorrent
JuiceTorrent (2008) was the most conceptually pure of the four. It took the machinery of the ad economy – contextual ads, widgets, click-through monetization – and pumped it in the opposite, gift economy direction.
Here's how it worked. A supported entity (musician, nonprofit, blogger, anyone with fans) opens a JuiceTorrent account and places a "Support me" widget on their site. Fans grab the widget and embed it plus JT-placed ad spots on their own web media spaces (blogs and sites). JuiceTorrent serves contextual ads into those spots. The ad revenue aggregates back to the supported entity. JT takes under 10% to run the automated attention/revenue redistribution operation. All that on top of the Google AdSense platform.
Advertising as gift economy. The attention-capture machine used to serve the attention-care exchange. No payments/subs from fans. Just ad revenue redirected from fan-owned attention spaces. We filed a provisional patent.
JT didn't catch on… perhaps was too complex (indirect) as a mental model and asked too much technically from the fans and the supported entities (they had to open their own AdSense accounts). Kickstarter started in 2009 (a year after JT) and Patreon in 2013, both with simpler direct payment models. The concept of direct crowdfunding from supporters as attention-care infra at scale was proven right.
Bookmentors
Bookmentors (2012–2015) – a nonprofit platform, funded by one of the Benchmark founders – matching teachers, donors, and books for high-poverty schools. I did the architecture and led the dev team.
The architecture was social and event-driven, with more than 50 email templates doing the attention vectors' work – transactions were not just exchanges, they were social events triggering messages, responses, follow-ups, thank-yous, book tributes. A teacher would post a book request… triggering emails to donors who either suggested this book or wrote a tribute to it; a donor would respond with a donation; the books would be automatically delivered to the school; thank you notes, comments, and suggestions for other books would follow; teachers would get emails about suggested books, etc.
You didn't have to be a donor or a teacher. You could be a Bookmentor by simply suggesting and offering tributes to books you deemed worthy of the attention of teachers and donors.
Members called it "gamelike and addictive". We designed it that way. The engagement patterns used by attention capture optimizing social media – feeds, notifications, likes, rankings, status – worked. We made them work for book donations to kids in high-poverty schools.
Bookmentors was a fully automated operation including the physical book deliveries to schools – via multiple external APIs (Library of Congress, Baker & Taylor, GreatSchools, Goodreads, etc.) There was no office, no paid employees. Thousands of teachers and schools signed up. A donor in Australia was giving books to schools in America. Google indexed ~20,000 Bookmentors URLs within the first 2-3 months. Bookmentors had a good chance to grow. Failed, or rather was discontinued, mostly due to internal organizational and people reasons (not relevant to the project's value or viability).
Some of the people now building the AI environments will have to work the attention-care angle somehow... fully aware of the inevitable compromises.