The general idea of training LLMs on forum archives seems promising though!
It all went great, Claude even gave references, until it hallucinated an innocuous but significantly game-breaking rule ("rest lets you heal one health for two stamina").
This just reinforced my lack of trust in LLMs, and I would sure as hell not ask an LLM what drugs were safe or not.
What would be interesting would be training using the data from PsychonautWiki[0] which has all sorts of good data about drug information & safety (and experiences)
I can assure you that my lengthy description of waves of orgasmic bliss that descended into a hospitalised bad trip† while on 2C-I and cannabis is written as I experienced it, nothing fanciful about that. But I also don't understand how do you claim stories about people literally on drugs to be fanciful. What do you expect people to see and experience if they've taken strong psychedelics? I have no basis to doubt one has really spoken to the Alien God while on salvia.
† thankfully it was only a panic attack, but they're a little bit scarier to have when the walls are melting.
I wish we were studying them with competency, in environments that were fully untethered from absolute profitism, LEO fear, politics and other non-medical agendas. We could use some better tools to address the negative effects of our increasingly complex (and decreasingly prepared) living.
Also, after I got hospitalised, the ER paramedics returned the small bag of white powder to me. I don't know if that's common practice , but I appreciated the gesture.
I still haven't listened to Maggot Brain since that day.
IIRC they have moderators vet them and only a relatively small percentage get published. The founders and staff actually take their mission statement seriously.
https://erowid.org/experiences/exp.php?ID=97541
The style of writing and the amount of stereotypes makes me convinced this is a rather hilarious piece of writing, not even false, but likely never happened.
I chuckled. Reddit data is THE gold standard for training something such as an RHCF model, properly up voted/downvoted data through an API. FWIW I'm sure a lot of ChatGPT data is trained there.
"tweakers.net" does something very interesting. They allow you to vote 'freely', but if you don't adhere to their voting rules repeatedly, you will have your voting rights locked away.
Their current rules are:
-1 = flamebait, troll, etc.
+0 = irrelevant, inaccurate
+1 = relevant
+2 = informative, helpful
+3 = must read comment
If you constantly give +3's to mediocre comments because they agree with your bias, you will get a warning. Similar with -1 to dissenting opinions. After one warning, your vote rights are locked away and you will have to write a good justification on getting them back.
To give an example on why voting sucks here for example (yes, I know) and it will sound a bit like sour grapes:
I made a comment on how Sonos is great, because their voice assistant processes your voice completely locally. I get berated and told it is not private because you need an account (the EU would fine Sonos severely if they claim local processing but do not do so). I get downvoted to -2 (!), whereas the completely incorrect comment gets sent to +3.
Similarly, another commentor says Sonos has horrible software lifetimes. However, they often update their products for ~10 years, which is a better support timeline than even Apple. I'm sent to 0 twice, incorrect commenter goes to +2.
With the Tweakers system, Dang would swoop in and annihilate the voting rights of anyone who upvoted these people. Alas, HN loves being a bandwagony echo chamber, so the pain continues :)
From the methods section:
> The database of drug narratives from Erowid’s “experience vault” is entirely anonymous. Rigorous quality control is conducted by trained experts with careful consideration given to “quality, credibility, and focus on effects or outcomes” (www.erowid.org). By means of that mechanism, all reports that provided the basis for the present investigation have successfully exceeded a standard quality filter.
Did something similar on the same data at school many years back, but I used the category labels to try and predict good and bad trips. The model worked abysmally, but the features made for great word clouds.