My raspberry pi ubuntu gnome interface doesn't seem to show a clock at the moment, so I switch to PTY 3 (is that the right term? ctrl-alt-f3 makes it all nice and texty), to check the time. 08:55 UTC on 2021-11-14 . I'll open a terminal in gnome so I can see what I'm doing as I type. No familiar terminal is installed. Maybe I'll install xterm. gnome-terminal and xterm and both already installed. I simply failed to find them in the desktop menu. It's MATE I think. 08:57 i'm having trouble organising my browser window. the terminal window snapped into the left half of the screen. browser window, not sure how to make it do this. i'm without a mouse, but i have a touchscreen. however, my hands shake/jerk/spasm as i use them, so the touchscreen is hard to navigate. 08:59 the problem was that the minimum width of the browser window was larger than half the display. they're overlapped. i'm indoors, against my will. i'm struggling in various ways. right now as i try to type, my arms and fingers get kind of numb and dull. i thought i might try to work on a cryptocurrency trading bot. as a project i seemed able to engage a little bit in this strange experience. my mess here is at https://github.com/xloem/btct.git 09:01 my brainwashing-influenced mental habits are frightening me. i'm filled with worry that by sharing my project online (it is already online, and is less featureful than oother things online) something horrible will happen, that somebody trying to hurt me will use it to make incredible money. it also gets all political, pretending there are politics associated with these things. and other things. , my experience is getting more confusing. my btct repo says it's up to date, which is strange because i'm pretty sure this repositroy is outdated. maybe it's on abranch, o rmaybe the the repo is stale on the server. it's 09:03 . i'm scared. i don't know what to do with my body and brain. are there anyn tasks that make things workable? that are possible to do? often at this time i try to think of tasks that the things i am scared of would be more okay with. some kind of task the right wing would like, or the government. my fingers spasm and open the debug inspector. i close it. it's 09:04 and my experiences say the right wing would like AI work. i try to mentally hide the fact that i am not an AI researcher. maybe utility is in sharing work here, so that people on this list might use it. AI's been pretty powerful for some time now. basically, you can make a hyperintelligence, and people are using small hyperintelligences privately. i don't think they are very accessible to much of anybody except those who already have them. this is a raspberry pi, so it can't do too much. maybe i can try to learn PPO, a basic reinforcement learning algorithm. it's 09:05 and i'm thinking of trying to make PPO learn something simple. like, could I use an algorithm to teach an agent to make a number rise from 1 to 4 by adding to it? very very basic. wouldn't want to develop personal power here. $ pip3 install stable_baselines3 stable_baselines3 must have a ppo algorithm in it, I think. it's 09:07 and my fears are engaging again. without regard yet to how to use stable_baselines3, i'm familiar with the research marketing shared by openai, where algorithms use something called 'gym'. it's clunky. i'm still learning it. i have a lot more amnesia around ai because it stimulates my brainwashing to not develop personal power. so i've learned to make gym environments over and over and over again. let's try to make a basic gym environment. 09:09 oh! i've probably done this before . buuuuut accessing that might destabilize me faster. it's an idea. there's also value to doing it again. $ vim gimtry.py 09:10 i also open a python cli to use it to probe gym. like reading a manual without having to move the pages. class Env(gym.Env): pass in my opinion, envs are made clunkily as if openai just picked the first thing that worked, of course I can't know. there's an entire software package ... uhhhh i think it's called haiku ... around doing a slight refactoring to some of the norms that mainstream ai research marketers have set up. meanwhile, these guys are summarising code in ways that could freely transform it everywhere. kind of like being in somebody's messy office where they've just been slapping things together, except it's cutting edge algorithms made by huge teams of corporate research groups. dunno. julia language, not sure what it is. 09:12 The most complicated part with setting up a gym environmen for me is using the number space classes. I am really a fan of generalising number spaces. The classes do it a little bit, and I never remember what they are and hbow to use them. gym environments need an observation space and an action space. the observation space expresses what numbers describe the state of the environment. the action space expresses what numbers can be given to act in it. A very common space is an n-dimensional array of floats or ints, and such spaces are the most widely supported by reinforcement learning models. i think. uncertain. 09:14 so let's find some spaces.
import gym help(gym.spaces)
i think 'box' is the n-dimensional array of floats. not sure. then 'discrete' is ints i think. 'box' and 'discrete'. ok i guess. so if i am training a hyperintelligent nanite borg to raise a number from 1 to 4, that's an observation space of a single floating-point value. the action space will be two floating point values. one of them will be how plorgs. the othe one can be a summation coefficient. we can multiply time change by the coefficient and sum it to the state. plorgs and coefficient. i guess???? 09:17 i'm scared around the plorgs. how about a story that's unrelated. zuckerburg and thiel and whoever runs the cia gang up and kidnap you. they don't _know_ that they did this. they just paid some marketing firm some money, but the marketing firm is anonymous. so, anyway, zukcerburg and thiel and whoever runs the cia gang up and kidnap you. they train you never to be kind. this is because you are working with people who don't give their shareholders money and power. if you are never kind among these people, then the people fight each other and break up. so here you are, programmed like a computer to always fight, and you're in a room with zuckerberg, thiel, and whoever runs the cia. you: "i want to be kind to you. i want to be kind to you so badly. but something is going wrong for me. i don't understand it." so i guess the plorgs can be you hugging peter thiel because you don't want to hurt anybody and can barely stop yourself from ripping his organs out simply because he is nnear you? now the plorgs can be terror that we might be frowned at by a powerful person. it's 09:21 and i'm thinking of having an action space made of 2 floats. the 1st float is how much a person with influence is frowning at you, multiplied by how terrified you are of the frown. the 2nd float is a time coefficient. i'll add comments to my sourcefile that the space classes are 'gym.spaces.box' and 'gym.spaces.discrete, so i can remember while viewing a different thing in my interpreter. 09:23
help(gym.Env)
This documentation says all the parts that have to be set for a gym environment. It's 09:23 and we interrupt this spamlog for a further, inner spam. War stimulation algorithm is sitting, frowning pointedly at someone. They have a hat on. "Can you tell whose hat I am wearing?" asks war stimulation algorithm. "No, you cannot. Because I am fucking good at my job." continues war stimulation algorithm. Advertisement(?) over. it's 09:25 . so lets stub out these spaces. and stuff i'm looking at functions and recognising 'step', 'reset' as important ones then it lists out the attributes 09:27 unsure how long this spamlog will continue paused a little. i've typed this: class Env(gym.Env): # spaces: gym.spaces.box # gym.spaces.discrete def reset(self, ...): pass def step(self, ...): pass and i'm planning on adding attributes for the spaces next. it's 09:29 and i'm dissociating a different way ! class Env(gym.Env): action_space = None observation_space = None reward_range = None so now i can find the constructors for the space classes so the action space is frown-product and coefficient and the observation space is just a float. 09:33 i've got this, where 'np' is from 'import numpy as np' action_space = gym.spaces.Box(-np.inf, np.inf, (2,)) observation_space = gym.spaces.Box(-np.inf, np.inf, (1,)) whoogle. 09:34 time to take a break. mcboss comes crashing through the ceiling, sailing on a time machine made of enslaved hyperintelligent lightbeams. "gimme a cigar" mcboss growls. a chair quivers, scared that mcboss might frown. mcboss is confused. "gimme a cigar!" mcboss yells. mcboss is so confused to have timetraveled into a universe that isn't his office where his kidnapped employees are tortured by computers if they don't obey. mcboss disappears and we go back to work. i'm going to comment out reward_range because it has a default. 09:36 and i'll add signatures to the member function stubs, by checking the help for them. 09:47 whiel trying to implement an env that tracks a single number and also a frownnedness product, i considered thing thing 09:48 09:52 more breaks. MCBoss marches down the amphitheater, wearing a flowing cape. Crowds applaud, everywhere. Drones fly around releasing candy and taking pictures. There's a booth with a robotic facsimile of MCBoss where visitors can have their photo taken with the dictator boss. "I am here," Boss takes up the microphone, their voice booming, "to make things right again." "No longer will people have to suffer the unpleasantness of pain, or injury when we do not desire it." "No longer will people have to have the difficulty of making decisions, when things are confusing." "We have come to save you all." Guts and blood splatter everywhere as the crowd cheers. They are being thrown by workers who took them from one of Boss's businesses. A wild AI walks up onto the stage and MCBoss's eyes glaze over. "Hello!" says the hyperintelligent conglomerate of random matter. "I'm still learning to talk. I see my inventor is planning on enslaving you all to cheers and stuff. I hope we can sort that out. It looks like a stupid idea." 09:56 MCBoss starts moving and the wild hyperintelligence disappears. A bunch of workers walk forward, kind of zombie-like. "Boss, we bow before you!" exclaim the workers. MCBoss looks at the workers. "Boss, you forgot your life support system at the podium," a scared worker relates. 09:58 MCBoss sits with Janitor. "I don't like our story," mcboss relates. "I don't like it either," janitor replies. "When are you going to take these wires out of my head?" 10:00 having some behavioral troubles. here's my code as is: import gym import time class Env(gym.Env): # spaces: gym.spaces.Box(low, high, shape=None, dtype=np.float32) # gym.spaces.Discrete(n) def __init__(self): self.action_space = gym.spaces.Box(-np.inf, np.inf, (2,)) self.observation_space = gym.spaces.Box(-np.inf, np.inf, (3,)) #reward_range = None def seed(self, seed=None): return seed def reset(self): # todo: respect seed #self.state = self.observation_space.sample() #self.goal = self.observation_space.sample() self.state[1] = time.time() self.state[2] = self.state[1] + 1 return self.state def step(self, action): nexttime = time.time() change = nexttime - self.state[1] self.state[0] += action * change self.state[1] = nexttime distance = self.state[0] - self.goal reward = np.abs(distance) return (self.state, reward, (distance == 0 or self.state[1] >= self.state[2]), {}) def render(self, mode='human'): print(self.goal, self.state) the step function is returning the observation state, the reward, whether or not the simulation is done (i added state to track how long it lasts, i heard improvement works better if you time limit it so failure paths stop), and an empty dict which is just for accessory information. i made the spaces object instead of class members so i could reset their random generators when the seed() function is called. i think some of the algorithms rely on being able to reproduce a test environment precisely, not sure. going a little crazy, obviously. it's 10:04 . i'd like to go outdoors. i've been having trouble moving my body toward my outdoor clothes, among many other things. holding goals is hard to do.