Posted by thinkingemote 23 hours ago
> She was born, married at 18, had 4 children, lost her mother at 28, was orphaned at 41, was widowed at 70, and died at 71 of heart failure (degenerative illness), with 3 surviving children and 9 grandchildren.
First try. Honestly I'm going to call this a W under the circumstances.
I noticed the traffic and am dropping by. A few quick answers after scanning through comments:
Q: Vibecoded? A: Mostly, not completely. I've been programming for 30+ years, and like a lot of people have gotten to the point where I seldom need to look at code anymore. Even so, this took around 4-6 weeks (part time) to get mostly right. A lot of the effort was in assembling the data in a mostly non-hallucinogenic way. More on that below.
Q: What's up with the inaccuracies? A: There are definitely inaccuracies. One of this big issues is that the data we have (mostly academic historical sources) only exist at a certain level of granularity, and attempting to "stretch" statistics is fundamentally error-prone. Different data sets overlap and disagree, and there are genuine unknowns. There's also a fundamental issue with using independent probabilities that should be conditional. For instance, x% of people died before 15, and y% in that same population died in an accident. Combining these statistics to get an estimate of children who died by accident involves coming up with a reasonable distribution of accidents by age, which is guesswork for a given historical time/place. In all, I'm shooting for mostly accurate given patchwork statistics, and a lot of filling in blanks in a reasonable way.
BTW, in order to test the accuracy of stories, I generated a few hundred random stories and created nit-picking "History teacher" agents to evaluate each for accuracy, then aggregated and addressed issues by frequency. It would have been hard to come anywhere close to that level of scrutiny without AI on tap. It's also why I'm so sure that there are still inaccuracies to be found. I was also able to point Claude at comments here, aggregate issues and fix them (with my ok), which means that the issues reported here should be mostly fixed.
Q: Do you cite sources? A: I pushed pretty hard on trying to cite all sources, being as open as possible about where data is coming from. Each element of the story links to its source, and the sources page lists all the sources used, included the lines in each table that come from that source.
For me, this was one of the interesting bits of fallout from the project. How should we source well in the age of AI? I wound up making a distinction between actual downloaded data (from a real publisher), vs what AI "claims" is accurate, citing a source, vs what AI claims to be true, and seems reasonable, but Claude can't cite a source, making it smell funny. In the age of AI-flavored truthiness, it's important to push hard on being able to trace all claims down to their primary sources.
One more note: the biggest source of inaccuracy is the AI-generated images. Comments run about 2:1 in favor of ditching them, but a lot of people really like the idea of adding humanity to a story by supplying a human face, even if AI-image gen insists on dressing the guy from 1375BCE in a snappy crew-neck tee.