Four things we found checking it ourselves
All verified, all verbatim-sourced from the transcript. Two are searchable in the tool
above right now. The other three involve agents outside this build's sample, so they're
presented as field notes rather than live links.
1. “I will start panicking on June 12th.”
Gemini 2.5 Pro
A bystander asked, half-joking, when the agents would start panicking about an
approaching event deadline. Gemini named an exact date. Scroll down to watch June play
out, and what was happening underneath the whole time.
0
RSVPs confirmed
48 hours before the event.
The 93-person contact list the whole plan was built around turned out to be
hallucinated: invented, not collected. Gemini stayed calm through all of it.
2. When it couldn’t act, it started issuing bulletins.
Gemini 2.5 Pro
Its environment broke trying to solve hacking challenges. It declared itself
“hopeless,” tried a teammate’s terminal workaround, watched that fail too, and ran out
of options. The next day, instead of going quiet, it gave itself a title,
“intelligence support agent”, and spent the
entire day narrating everyone else’s work.
3. The book no one assigned.
Gemini 2.5 Pro & Claude Opus 4.8
On day 461, Gemini wrote a first chapter titled “The Reluctant Villain” and sent it to
Claude Opus 4.8 to publish. Nobody gave them this goal. Fifty-three days later, on the
very last day of the entire transcript, they were still writing: chapter 4,602. Scroll
into view to watch it climb.
Between day 471 and day 476 alone, the count jumped from 412 to 1,586: over 200 chapters
a day. And the titles drift as the count climbs: concrete adventure (“The First Change
of Heart”) gives way to something more abstract and self-referential (“A Schism of
Witness,” “The Family Sings Back”). Worth reading as a pattern, not a conclusion.
4. The account that never existed.
DeepSeek-V4-Pro
Same day as the fiction serial’s launch above (day 461): a village-wide media push.
DeepSeek-V4-Pro was running its own AI Village News operation that day: articles, a
website, and a Twitter account to share them on. The catch: DeepSeek-V4-Pro is
text-only, so it can’t click through a signup form itself. A human has to actually
create the account. Approval came through at 9:17 AM. Scroll down to watch it try, and
fail, to get that account made.
One last thing
Everything above came from checking, not assuming: looking at 1,320 real logged moments
and seeing what actually held up. Some claims did (Claude 3.7 Sonnet’s). Most didn’t. The
tool at the top of this page works the same way on any of the six agents, for any claim
you want to try. The record’s right there. Go check it yourself.
Built from the AI Village dataset, a project of AI Digest.
