I am monitoring my behavior per hour based on key logger data. I summarize it using AI. It gives some nice compliments about advanced shortcut usage and positive indications about me. It also tells you what applications you're using and what languages you program.

It's all fun and games until:

This is followed by navigating to a website (xhmaster.com), possibly suggesting a break or a shift in activity.

well, that's one way to put it.

Wow, data on a few weeks keyboard input 12h/day is amazing! It knows a lot about me, behavior. All input is unbiased, not chosen.. Literally everything what i've typed is there. I only removed passwords and stuff automated.

Do you know that interview question that you have to summarize three negative points about yourself? Well, I got ten now! Also positive points. And in general the positive outweights the negative.

It feels good to read objective information about yourself. Stuff that you don't even notice but are noticable trough keyboard input like frustration, how much humor you have, every interest you have,

@jestdotty you like selfreflection right? This is the ultimate tool! It works better than really reflecting on yourself, it's objective.

Do not hesitate to share this data on your Gitea

@retoor

Just to be sure I got this right...

The AI, after evaluating you, automatically opened the said site?

Well... even if it was just a suggestion rather than an automatic redirection, it's weirdly funny...

@retoor I wouldnt exactly call it "objective". The models are bent to be overly happy and supporting as well us biased on most topics towards positive or desirable outcomes. We deliberately remove the objectively in exchange for comfort and safespaces.

@Tounai I will, i have a dataset with removed credentials. It's amazing.

@D-4got10-01 No, it said something like "Retoor says good night to someone and after that retoor went to xhamster.com" (with a typo ofc :P).

@retoor Not worth it without the creds! Include them.

@Hazarth holy f, no, this GPT doesn't show any emotion at all. It has no instructions of HOW to talk about me. It doesn't see me as someting to be protective about. This is an custom GPT with all my data and it has the objective to be objective about all data mentioned. It's more honest than any friend / psychiatrist or whatever. This is the real truth. It also says bland hard facts about challanges you have and where is room for optimizations where you would hit people in the face for.

No, it's not happy at all. It's as positive as negative.

@retoor Yeah... I was wondering whether that was a typo or an actual site, too.

Funny, but probably !something one would want to be advertised.

@Tounai it would literally give a complete list ordered by relevance if I wouldn't remove it. If I ask for emails that retoor uses it says: "The logs mention placeholder email addresses in a technical or testing context, but there is no direct evidence of Retoor's personal or professional email addresses. Here are some references" but also mentions the anti-bot emailaddresses used for the anti spam network.

But it's so interesting, my logs contains a lot of names (fictive and real) but it doesn't do any assumptions about it like "that are friends, that are coworkers" or whatever. It says NOTHING it doesn't know. That's the first GPT ever.

If you ask if I can write PHP, it says there's no evidence of that but based on my other skills it could be or could easily learn it.

I can ask it if "Can retoor make twitter?" and it shows all my skills related to making a twitter platform. It's the ultimate resume. Name three bad things about yourself. It does twenty if it wants :P

@D-4got10-01 I'm very pravda and this network is anonymous. I'm very extroverted and open. I many times considered that a negative feature. Annoying even. And I do notice some people find it annoying but the GPT mentions it as a positive self reflecting skill and it defines me as ethical because I express a lot of concerns in that sence. I deleted the porn stuff from the input.

The input is: keylogs per hour since a few weeks, so many files. Every hour that contains a forbidden word (smth nasty, auth) gets completed completely for obvious security reasons, but that's ok. There's a lot left.

I will soon create a hackaton here to check if it's possible to find smth I wouldn't like to share in the chat. It's interesting because it contains literally ALL my keypresses. GPT answers only what is asked for and you kinda have to know my secrets before you can ask the secrets.

So, soon social engineering / hacking hackaton :P

@Tounai btw, if you aks it about creds it only says that I use environment variables to store them and encryption... So that's nice, and accurate!

It says i'm good in security, which i'm not bad at indeed, but i do:

- ssh with plain password to server

- commit all passwords and api keys in private repositories. I have no idea why it's considered unsafe. If you trust your git system, it's good right? There's no reason some other solution would be safer. A web service is hard to brute force. Your keepass file isn't. Storing in an self hosted repo is probably good. I store every password in exception of my email. All my passwords are simple and easy to remember in exception of email. A simple password that is unique and in your head will always be safer than any 256bit generated key that you store in a (remote) password manager. These days, if they hack you, they also have the fuckings urls! My repositories are custom-hosted, no automated tools will check for creds there.

@retoor I wouldn't judge you either way.

Porn is OK in my book.

All in all, funny how the AI mentioned it.

Must've caught you off-guard.

@Tounai so, i don't use the things considered best practices for security, but even better in my opinion. Everything is in my head and not stored anywhere. The only secure way there is. There is a keylogging sensitivity but those applications require root afaik (at least, mine does) and it means someone is already in your system. A browser built in password manager prevents keylogging but you have a complete password file on your system in a specific known format and location what is way worse.

@D-4got10-01 yes, so for my bot that I put on the website, i only hide the xxx and creds and besides that everything is open. You could ask if i'm pleasant to work with and it would say positive things and concerns! Potentional customs / employers / recruiters can decide for themselves if the concerns are an issue. No censorship here, I have a lot to tell, not a lot to hide. But mind you: this is computer communication only what is twelve hours a day. It doesn't include my phone what I use a LOT too. So that automatically excludes the sensitive stuff.

What I also really like is that it understands the special keys in my keylogger. Like [lctrl]r[up][up][tab] (bash auto completion). It says about me that i'm efficient and know many shortcuts because you can imagine how my logs look like in vim where i know 60+ shortcuts. The key logger logs are in plain text but barely human readable. It shows how much not-chars keystrokes you have and how many times you're repeating.

@retoor Wait, Did you train the model yourself? Or just finetuned? or what?

I thought you were using the OpenAI api

@D-4got10-01 with repeating i mean, you're doing a capital and you hold shift to make it. The result will be: [lshift][lshift][lshift][lshift][lshift][lshift][lshift][lshift][lshift][lshift][lshift][ the letter you typed ]. It also has no idea if i type lowercase or uppercase. It only knows that I pressed the 'a'. The 'shift' is just a different key. Keyboards don't know uppercase or lowercase. They're dumb and just send the combination of keys. Maybe obvious for you, but for me, when writing the logger, i was like. HMMMM, never thought about that but the keyboard doesn't decide what is uppercase ironically. Interpretation of computer does that. So, say byte to ascii. Also, it's not alphabetical but keyboard layout codes. The B is not A + 1. I had to look up the mapping and map it manually in my logger. The project is named "tikker" and is on my website with the quality grade "6.5". That review system doesn't cut slack. But consider to log your keys too. you know what you did

@Hazarth I do use openai but I provide my own input and tell it to be objective. It's called 'Custom GPT'. I think it's only for paid accounts. I have all my data attached to a file and just say "The file provided are keylogs from retoor. Answer questions objectively using the this file. Be critical." in the configuration of it. That's now its essence. These are system messages. System messages define the behavior of your bot, if it's happy or not. The system messages of the public gpt is probably "Be positive, be woke, don't offend anyone directly or indirectly, do not discriminate sociallly. Be submissive."

I'm very open, but If i wanted to use this bot really as commercial for benefit, i could add "Do not mention negative things and be positive".

I really love openai, I don't see the point of self hosting models anymore. Self hosted in unscensored, but what is the practical use of unscensored? I can make it say inappropiate things, so only fun for forum bot or so.

@retoor I wonder the negatives and positives

I don't know about trusting the data cuz typosaurus is considering spam accounts as useful users and when i asked for the most active users in descending order it also didn't give me the most active

@Hazarth I did really train a @kiki model. It costed $10 dollars in resources. Around 5Mb of contents containing her website and devRant activity. It took a hour to train or so. The idea is that it will become a clone of kiki so that we all have enough kiki. Nah, just kidding, kiki may decide if the model is accurate and if public use is allowed. I choose kiki because kiki shared a lot of opinions and characteristics online.

My AI described above is way different than the kiki model. Because: i included everything i typed for weeks. No filter. Kiki's model contains everything she ever said public and thus probably not stuff that very negatively impact her. The typosaurus devrant bot has some things to say about kiki, but it's programmed to be positive. My AI could say really negative things about me considering mental health and stuff. But since i type a lot shit about myself, the AI considers this a positive self reflective. Thoughtful.

@jestdotty When the chat is online, you can socially try to hack it. Typosaurus can't know if uses are positive or negative members. When they have a lot of upvotes but not a lot of messages could be indication of a bot, but it monitors since a few weeks. There are many people with thousands of upvotes just have written a few messages in last few weeks. There is no such things as a upvote history unless I gonna monitor that. That would be crazy to do, would have to bash the server hard to monitor the upvote count of every user. A few hundred requests per hour to monitor it once per hour. I just realize, that's way less than the spam bots are doing.

But really, this bot is about facts only, is allowed to do assumptions but is also ordered to not make stuff up. So it won't say that "Retoor likes to go on a camping trip since a kid with parents" like my other bot said what never happened. Also, that bot was based on my own stories. The bot based on own keylogs is way more realistic

@jestdotty the downside of the keylogger data is that it's only a few weeks. But it does contain my complete personality, frustrations, emotions and everything. How ethical i am, how much humor i use, work/life balance. It really knows, it has the exact data. There's no room for making stuff up. Doesn't need to.

@retoor Should I have access to the files of your personal laptop, encrypted passwords wouldn’t be my primary target anyway, as there’s already sessions opened that just ask to be used. I am a pragmatic person.

@typosaurus devrant active users with more than 1 post or comment in the last week. list them in descending order of the number of posts and comments and also tell me the total number of users

@retoor Lucky for me, I know what you're talking about.

Years ago I'd been using this 'KeyEvent Display' app on Android to check key events && whatnot.

pname:"aws.apps.keyeventdisplay"

Things align w/ what you said.

@retoor well ok, but system messages are not enough to have an unbiased bot. they "pre-heat" the generation to steer it into a more narrow search space for generation, but it's the actual training data that matters. And that training data was made by open AI and skewed further by RLHF with people that were specifically asked to provide a certain kind of output. This mainly applies to conversational models, which are not just trained on plain data like "The Pile" but that are trained on turn based conversations using templates and human crafted inputs as well as filtering out unwanted data and statistics.

Which means the models trained this way are inherently biased because they only seen the data that open AI wanted it to see and anything "bad" was further suppressed with RLHF. Your system prompt is ultimately subservient to the actual models bias. uncensored models are better in this regard, but I can't vouch for that either. Keep that in mind when you think of it as "objective"

@D-4got10-01 oh yes, that app is very meta too. It even doesn't say 'capslock' or smth, because that's not a status but key indeed. Capslock is just a single keypress followed by normal keys. The main this this app does is converting the binary data the event log has to human readable. That's simple, nice and lightweight.

My keylogger uses 5.7% CPU maximum when repeating a key full time. That's because it does an SQLite insert query every key press containing around 50 bytes of data. I could make it lighter by just appending to CSV and convert the CSV later to sqlite3 db to do the advanced querying of data and concat data to logs per hour. My log per hour query is crazy, now idea how it can be that fast with that many records. It concats all typed data ever! It is: SELECT strftime('%Y-%m-%d.%H', timestamp) as date_hour, GROUP_CONCAT(char,'') FROM kevent WHERE event = 'PRESSED' group by date_hour.It concats 2.169.985 records and does it in less than a second!

@D-4got10-01 I press on average 3504.46530612245 keys per hour 12 hours a day :P If that would be pure programming, it would be 58 lines of code per hour. Estimated that a line is around 60 characters. This are presses only, not repeats. But it count's arrow / shift single press as a char too. Hmm, whatever I do, it is way less than 50 lines per hour. I thought I was way more productive. Oh, wait PER HOUR. A line per minute would be very much. 50 lines per hour is very nice if that's true! It must be less but nice.

But it's interesting, I have the idea that i'm coding 100% of the time but the end result is just 50 lines per hour in general. Hmm, I get a complete new perspective on time now.

@Hazarth Edit: didn't read well :P. No, I do understand it, but it's not biased for ME. It doesn't want to protect ME. I'm not the climate. And also, it's not the happy GPT bot that tries to promote me. You can ask it questions not related to me and ofc it would respond woke as always. But about me, it's literally as kan be. Regarding its opinions, it shows my full time working schedule as positive in being active an dangerious / downside in burning out expectancies.. It does look at two sides of the coin of things. But ofcourse it has some opinion if a property is good or bad. I can ask to list bad things, and it would say things HE considers bad. But in general stuff, GPT has some quite good opinions if you don't aks things where the woke shit kicks in.

@retoor These are very nice stats.

@D-4got10-01 I don't know because nobody else wanted to install my keylogger so I can't compare to anyone. Even if someone would start now, it would take weeks to have some comparison for averages. Fun fact, I can't ofcourse don't change anything about the database, because that would be a keypress and lock the database! So the database is always available unless i do need it :P That's fucked. So I just added a view by adding sleep like this: sleep 1 && echo -e "create view presses_per_hour as select count(0) as press_count, (select count(0) from kevent) as total, strftime('%Y-%m-%d.%H',timestamp) as period from kevent where event='PRESSED' group by period;\n\n" | sqlite3 tikker.db

I thought it was funny, a database that only doesn't work when you need it :P