There is a new AI interpretability paper by anthropic that more people (especially neuro people) should check out. It's both fascinating and terrifying and I wrote a blog about it.
Scary things:
1) It's not great for LLMs to have some mechanism for thinking that we can't observe
2) The bigger/better models are more capable at this
3) This all just emerges accidentally
Another new paper from our team released this month. We used ML and patient health data to predict whether patients prescribed an opioid would go on to use them chronically within one year. Importantly, models worked great in some orgs but not others
https://t.co/DYGUEyyJmc
A new paper from my team, investigating follow-up after a positive stool-based cancer screening where we found that a huge fraction of patients with positive results never received a follow-up colonoscopy to diagnose CRC. This was very surprising to the providers we interviewed!
I have a new paper out w/ @JMGrohNeuro and John Pearson! We compared the way humans and monkeys use vision and hearing together to locate targets. Our modeling results show that both species use causal inference to either combine or separate the two senses
https://t.co/wIkYhkWKJf
@Jonna_Singh_ Related to the paper @Wolfwoodgamer linked. Each project is a folder containing: data (raw), results (processed data, figures), documents (drafts, notes), source code. Self contained except for shared code from a separate 'utilities' project/git repo.
Reviving my twitter because science twitter is actually pretty neat. To commemorate this occasion here is a paper about a nanoparticle that you can inject into your retina to see infrared light. https://t.co/JrXZ3avEcj