Ironically enough, the huge review of criminology research appears to misunderstand p-values in the section where it's talking about common flaws in scientific research.
> With standard hypothesis testing methods, there will be one false claim of a relationship between a cause and purported effect for every nineteen times that it fails to find support
Congrats on the new job! Should I hold out hope you might still write a book on how to stay skeptical without spiraling into conspiracism? I’d sure love to teach with that book. I’m around if you ever want draft comments.
The criminology one is quite useful for me, thank you so much for including it! Even in the short-term it's not clear some of those interventions do much. From my understanding of singular programs you can get some modest reductions in violence via ramping up your supervision (and as a result arrests) of previously arrested individuals, but beyond that it's murky. One problem is that progressive programs - your jobs and therapy programs - often INCLUDE this increase in policing, which means it's hard to discern why the program may have decreased violence. I mean the more people you have in jail the fewer potential murderers are in public life.
Congrats on the new job at anthropic! I wonder though if the culture there is going to be a good fit for you. Anthropic is notorious in the AI space for splitting from OpenAI because, basically, OpenAI weren't woke enough. Their models have been falling DOWN the rankings with each new version largely due to increasing rates of refusals on ethical grounds, even when it makes no sense.
The sort of people who make such decisions have a high overlap with the people who have a blind faith in the truth of anything said by academic specialists. I get they hired you for your record in communication, but a flinty eyed science realist risks conflict with the sort of people who think they have a moral obligation to ensure their models never give out "misinformation".
Anyway I'm sure this thought has already occurred, and it may never come up anyway. Hopefully you can even drag them a bit more towards the center!
Ironically enough, the huge review of criminology research appears to misunderstand p-values in the section where it's talking about common flaws in scientific research.
> With standard hypothesis testing methods, there will be one false claim of a relationship between a cause and purported effect for every nineteen times that it fails to find support
Congrats on the new job! Should I hold out hope you might still write a book on how to stay skeptical without spiraling into conspiracism? I’d sure love to teach with that book. I’m around if you ever want draft comments.
The criminology one is quite useful for me, thank you so much for including it! Even in the short-term it's not clear some of those interventions do much. From my understanding of singular programs you can get some modest reductions in violence via ramping up your supervision (and as a result arrests) of previously arrested individuals, but beyond that it's murky. One problem is that progressive programs - your jobs and therapy programs - often INCLUDE this increase in policing, which means it's hard to discern why the program may have decreased violence. I mean the more people you have in jail the fewer potential murderers are in public life.
Congrats on the new job at anthropic! I wonder though if the culture there is going to be a good fit for you. Anthropic is notorious in the AI space for splitting from OpenAI because, basically, OpenAI weren't woke enough. Their models have been falling DOWN the rankings with each new version largely due to increasing rates of refusals on ethical grounds, even when it makes no sense.
The sort of people who make such decisions have a high overlap with the people who have a blind faith in the truth of anything said by academic specialists. I get they hired you for your record in communication, but a flinty eyed science realist risks conflict with the sort of people who think they have a moral obligation to ensure their models never give out "misinformation".
Anyway I'm sure this thought has already occurred, and it may never come up anyway. Hopefully you can even drag them a bit more towards the center!