Raw LLM Responses

I've spent a lot of time developing statistical methods for identifying spam in email. And I assure you, the same fundamentals apply to any sort of predictive methodology. Sepcifically: 1. Dumb beats smart. Every single time. Whatever clever algorithm you come up with, nothing has ever outperformed brute force statistics on raw data. 2. You're only as good as your source data. If your source data is garbage, then every prediction you make will be as well. If your source data is biased, your results will also be biased. 3. Data changes over time. Spam evolves. It's gotten a lot harder to detect over the last several years. Society also evolves. Predicting outcomes based on data from the 1980s will be unreliable (see #2 above). 4. Every prediction has a margin of error. The best spam filters make mistakes. When applying the same methodology to something as unpredictable as behavior, the margin of error will be higher. None of this, however, gets to the biggest problem here - we have a criminal justice system that's predicated on punishment, not rehabilitation. Beyond that, transparency is essential. If I were a judge or a juror, I would never rely on a black box output, ever. Courts should never accept any piece of software or data that can not be audited. Doesn't matter if it's COMPAS, or a breathalyzer, or anything else.