iak8god

Per the very paper we are discussing, LLMs when asked to act as therapists reinforce stigmas about mental health, and "respond inappropriately" (e.g. encourage delusional thinking). This is not just lower quality than professional therapy, it is actively harmful, and worse than doing nothing.

> the idea is to give you an AI generalist that might ne able to alert you to things outside of your field that may be related to your work

That might be a good goal. It doesn't seem to be the goal of this project.

The main concerns expressed in Robyn's note, as I read them, seem to be 1) generative AI has polluted the web with text that was not written by humans, and so it is no longer feasible to produce reliable word frequency data that reflects how humans use natural language; and 2) simultaneously, sources of natural language text that were previously accessible to researchers are now less accessible because the owners of that content don't want it used by others to create AI models without their permission. A third concern seems to be that support for and practice of any other NLP approaches is vanishing.

Making resources like wordfreq more visible won't exacerbate any of these concerns.

Word salad is "confused or unintelligible mixture of seemingly random words and phrases." Which part of the article seemed confusing/unintelligible/random to you?

I mean that the full article is paywalled for me at that link.

"Study 3" was September 2021 - May 2022:

> Our data came from a larger project investigating college students’ lifestyle and well-being. In this project, one hundred and seventy-one college students (102 females; Mage = 19.24, SDage = 2.68) participated for partial course credit between September 2021 and May 2022. For our purpose, we focus on the longitudinal component of this study, which consisted of two parts: a baseline lab session (Phase 1, N = 171) and two follow-up weekly surveys (Phase 2, N = 160; Phase 3, N = 160).

Here's what they say in the abstract:

> Providing stronger evidence of the directionality of this effect, Study 3 (N = 171) showed that in college students CRP predicted increased social media use in the subsequent week even after controlling for current week’s use.

The basic methodology for this part is:

* Blood sample (measured for CRP)

* 1 week later: survey on social media use (participants using Screen Time app to objectively measure their activity)

* 1 week after that: survey on social media use again

Then they ran some regression analyses.

I didn't see anything here that rules out the alternative possible explanation: college students who use social media more are more stressed out as a result of their social media use.

Most people won't be able to access the paper via that link.

I was able to get to it by going to Google Scholar and searching "Brain, Behavior and Immunity" david lee. The first result was a sciencedirect link that did allow me to read the whole thing.

They had students use screen tracking software to log what apps they were using. The researchers ignored TikTok because it wasn't included in some pre-existing methodology they were following.

Ignoring TikTok for any reason in 2021-2022 is a huge mistake. Once I saw that, I decided I'm not interested in their findings.