@niplav what omg!! it's "old norse" but apparently I'm not old enough to've seen it before. awesome!
@rime you were looking into lie detectors and their SOTA, right? Any good resources?
I've become convinced this might be really really important, thanks to you
@niplav But semantic reconstruction requires >10h in an fMRI machine while calibrating a GPT-like predictor based on e.g. your brain's recognition of audiobook, and I'm unsure how much the training-hours can be optimized. Also not sure whether smth-like-this generalizes to learning to neurally differentiate self-believed statements and self-unbelieved statements with sufficient accuracy. But just based on vibes, the impressiveness of the technology makes me think lie-detection is more feasible.
@niplav Perhaps the most impressive examples of brain-reading tech in the vicinity of lie-detection is semantic reconstruction, eg:
BCI Award 2023 #1: https://www.youtube.com/watch?v=Q1rctJd37a8&list=PL_JwSzOwE-dS0u9NNhv8__XktdDZaq_ML
BCI Award 2022 #2: https://twitter.com/guillefix/status/1679178300508504064
@niplav j some link:
"An increase in selfish motivation for Pareto lies was associated with higher mean-level activity in both ventral and rostral mPFC. The former showed an increased pattern similarity to selfish lies, and the latter showed a decreased pattern similarity to altruistic lies. … Our findings demonstrated that hidden selfish motivation in white lies can be revealed by neural representation in the mPFC."
@niplav Oh, I hv an abandoned LW draft here: https://www.lesswrong.com/editPost?postId=gt9A3bNxnZRGewMZF&key=6857a642401e8c373cd799e1959f9e
@niplav
fMRI-machines are currently too inaccessible.
Making it hard-to-hack is hard.
Doing the processing on a server, and providing instant results via app connected to the web, may make this more feasible.
If it's a hat, and it needs to be tailored to each individual via upfront calibration at a clinic, the clinic can record your signature and compare it with whatever their servers receive whenever you put your hat on later.
@niplav But below that, just making it harder to get away w obviously antisocial behaviour (like theft, lying in order to tarnish smbody's reputation / get them fired, etc) seems tremendous. What if being a sociopath makes you unfit for being a politician?
Whew.
For most scenarios that I think are pivotal, I think the tech has to be scalable/cheap, highly accurate, hard-to-hack, and launched by a highly reputable company (preferentially nonprofit, open-source—I'm allowed to dream).
@niplav There are many levels/dimensions of this w varying degrees of technological feasibility. I think most of the value is unlocked when the tech is (directly or indirectly) relevant to more or less ordinary social interactions, and can interfere w stuff lk "deception arms races"/"iterative escalation of social deception"/"deceptive equilibria".
@niplav 🕯️
To say that I've "looked into it" would be a big exaggeration, but I've looked into it.
The main reason I've been interested in it is: mass adoption of "veracity/credibility tech"¹ seems potentially extremely good for culture and maybe pivotal wrt many large-scale longterm stuff I care abt.
¹(idionym for stuff that helps us prove that we're being honest *when we actually are*)
@cosmiccitizen oh I'm much on board w prioritizing head-injuries more. I was mostly joking by using myself as "counterexample". I'm ~5% that it had a more or less "pivotal" impact on my life.
spelled out: IF u can interact w smth, it's real.
sometimes social reality falsely colors unreal things real, and we're better off seeing thru it; but don't forget to also look hard for that which has been made falsely unreal! there are colours there beyond our imagination.
Having soulfwly recited the Litany to myself on numerous (~500) occasions for just about one decade now, this is the first time I'm seeing it in the "one man's modus ponens…"-way.
"Anything untrue isn't there to be with." 🔄
@cosmiccitizen Counterargument: I've hit my head lots, and had at least 2.5 grown-up-labelled concussions.
- Slammed into stone-wall at high speed with toboggan. (Though is contributed to my aura of destiny, since I had an iconic scar on my forehead for >1y.)
- Fell off tree while climbing, landed on back.
- Fell off my bike (xN).
- TV fell on my head.
- (more)
...Not sure why I said "counterargument". I'm decent at abstract thought, but have a mental-illness-load that rivals Jeanne d'Arc.
@niplav Whisper (and others) is adequate for speech-recognition, especially if you're just aiming to use it for dictation and not have too thick of an accent.
It probably gets German accent (or whatever you have) better than Norwegian accent, though!
I haven't tested these:
- https://github.com/KoljaB/RealtimeSTT
- https://github.com/johannesCmayer/system-wide-whisper
^ That last repo made by Johannes. Consider it a person-recommendation!
@niplav For GPT speech, I currently use VoiceWave¹, because it allows keywords like "send" and "clear". I talk to myself a lot, so it's important that I be able to control what's actually sent. OpenAI's app stresses me out.
There are better tools for this, including open source and customizable (e.g. KoljaB's repos²), but I haven't prioritized setting setting up yet.
¹ https://chromewebstore.google.com/detail/voicewave-chatgpt-voice-c/baahncfnjojaofhdmdfkpeadigoemkif
² See RealtimeSTT and Linguflex: https://github.com/KoljaB (owo, so many updates since last time!)
@niplav Another thing I use STT/TTS for is being able to quickly ask fact like questions without leaving my computer-free-zones desk/couch/bed. I get too deep into looking things up if I actually use the PC.
Google Assistant is good for very well-defined questions (e.g. "what's the elimination half-life of 5-HTP?"), and has the advantage of not being able to confabulate plausible-sounding answers unless it has a source.
@niplav tell me! 👀
@niplav Writing with it is much faster than writing with my keyboard. The annoying part of it is that it spells things right. It makes things long.
My Task-Manager / Time-Tracker / Note-Taker / Schedule-Keeper / Context-Shower / Thing-Framework-Setup is very much incomplete, so at the moment uh I can say "new task" to iterate Toggl Track, or "new note" for simple notepad (usefwl eg at night), but far from all that I want with it.
Voice commands are very convenient, but hard to get used to.
☙ what i mean by "serendipity rarely distributes ato causality" is that concentration of VoI is often thickest at the joint observation of causally ~unrelated processes. if u want the most generalizable variables, sampling seemingly-unrelated generators is like selectively filtering for stuff that applies everywhere.
☙ if u tune the sensitivity of u's detectors *down*, but sample nature's generators broadly and rapidly… u avoid the diminishing marginal returns associated w any given generator.
Flowers are selective about what kind of pollinator they attract. Diurnal flowers use diverse colours to stand out in a competition for visual salience against their neighbours. But flowers with nocturnal anthesis are generally white, as they aim only to outshine the night.