2018s Crazy Ideas

Posted on February 22, 2019 by Noon van der Silk

I can’t believe I forgot to do this at the start of the year!

Let’s look back over 2018’s ideas:

The Ideas

  1. using ai to reduce medical dosages: i.e. imagine you need such-and-such amount of radiation in order for certain whatever to be seen in some scan can you lower the dosage and then use some ai technique to enhance the quality of the image?
  2. something about shape-based reasoning: i.e. the fact that some grammar-parsing problem teems like a good fit for tensor-network works because they both have the “shape” of a tree another example would be thinking of solving a problem of identifying different types of plants as it fits in the “shape” of a classifier but these things could potentially have other shapes? and therefore fit into other kinds of problems?
  3. a q&a bot that can answer tests on the citizenship test: Q: What is our home girt by? A: ..
  4. buddhist neural network: it features significant self-attention and self-awareness it performs classification, but predicts the same class for every input as it has a nondual mind it’s loss function features a single term for each of the 6 paramitas: - Generosity: to cultivate the attitude of generosity. - Discipline: refraining from harm. - Patience: the ability not to be perturbed by anything. - Diligence: to find joy in what is virtuous, positive or wholesome. - Meditative concentration: not to be distracted. - Wisdom: the perfect discrimination of phenomena, all knowable things.
  5. music-to-image: say you’re listening to greek music you want to generate an image of a singer singing on some clifftop on greece under an olive tree near a vineyard surely a simple matter of deep learning
  6. code story by word frequency: take all the words in a code repo, order them by frequency, then match that up to some standard book, then remap the code according to the frequency
  7. generalisation of deutch-jozsa’s problem: here - https://en.wikipedia.org/wiki/Deutsch%E2%80%93Jozsa_algorithm generalise it so that we have multiple f’s that have different promises; i.e. i’m constant “50% of the time”. what now?
  8. analogy-lang: reading https://maetl.net/membrane-oriented-programming by @maetl i had an idea for a programming language consider two types, “person” and “frog”, let’s say there are at least two problems; one is the one in the article - it’s hard to come up with a complete list of methods/properties that should exactly define one of these things. in surfaces and essences they argue that in reality no categorisation has perfectly defined boundaries; so how to define the types? another problem is what happens if i want to build a thing that is both “person-like” and “frog-like”? you can’t. especially not if the differences are far away (i.e. is frog hopping like “walking”? should it be the implementation of a “walk” method? probably not; but a “move” method? probably? but what about less-clear things? and doesn’t it depend on what your merging with) in this way it seems like it’s impossible to come up with strict types for anything so here’s an idea: analogy-lang: let’s you define types in a much more relaxed way; “this thing is like that thing in these ways, but different in these ways”
  9. multi-layered network: train some network f to produce y_1 from x_1 f(x_1) = y_1 then, wrap that in a new network, g, that produces y_1 and y_2 from x_1 and x_2 g(x_1, x_2) = ( f(x_1,), g’(x_1, x_2) ) and so on. interesting
  10. psychadelics for ai: ideas: - locate some kind of “default-mode network” in your model and inhibit it - after training; allow many more connections - have two modes of network; one this “high-entropy” learning one, which prunes down to a more efficient one that can’t learn but can decide quickly
  11. ultimate notification page: it’s just a page with a bunch of iframes to all your different websites where you get the little notification things and then it just tiles them; showing in-which places you have notifications.
  12. use deep learning to replace a movie character with yourself and your own acting: 1. semantic segmentation to remove an actor 2. film yourself saying their lines 3. plug yourself back into the missing spot 4. dress yourself in the appropriate clothes 5. adjust the lighting 6. ??? 7. profit!
  13. paramterised papers: “specifically, this model first generates an aligned position p_t for each target word at time t” show me this sentence with t = whatever and p = some dimension
  14. on a slide-deck, have a little small game that plays out in a tiny section of each slide
  15. use autocomplete and random phrases to guess things about people: “i love you …” find out who they love “give me …” find out what they want and so on
  16. visualise the difference between streaming and shuffling sampling algorithms: streaming -> for low numbers, misses items shuffling -> for low numbers, bad distribution across indices that are shuffled. @dseuss @martiningram
  17. symbolic tensorflow: so i can do convs and explain them really easily
  18. the “lentictocular”: uses lenticular technology on a sphere with AI so that it watches your gaze and moves itself accordingly so that it always displays the appropriate time
  19. ai email judgement front: intercepts all your emails for every email, it decides the optimal time that someone will respond it sends it at that time so that they respond
  20. “growth maps” for determining affected areas of projects w.r.t. a pattern language
  21. friend tee: lights up when other friends are nearby
  22. lenticular business cards: this is already done by many people
  23. innovative holiday inventor: thinks up cool holidays
  24. buddhist twitter: there’s only one account, and no password
  25. programming ombudsman: @sordina @kirillrdy
  26. the computational complexity of religion: given various religious abilities, what computational problems can you solve? what are the implications on computational complexity by buddhism? and so on.
  27. spacetime calendar: we have calendars for dates but they don’t often contain space constraints so why not a space-time calendar, defined in some kind of light cone?
  28. app to check consistency of items before you leave: it’s an app you configure it to be aware of your keys, laptop, wallet, glasses then, as you leave your house, it can inform you of the status of those items: - “hey, your computer is at home” - “all g, your glasses are at work” etc @gacafe
  29. gradient tug-of-war demonstration: given a function f(x,y) = x + y then if we have competing loss functions then it would be nice to visualise the gradient flow as a tug of war
  30. a website that is entirely defined in the scroll bars/url link bar, whatever: you can move pages by moving your mouse to different parts of the scroll bars and so on in that fashion
  31. quantum calendar: it’s a calendar where on any given day in the future, items can be scheduled at the same time. but up to some limit (say 1 week) the items get collapsed and locked in @silky @dseuss
  32. streaming terminal graph receiving updates over mqtt: then, can use it to plot tensorboard logs to the terminal instead of tensorboard using blessed + blessed-contrib seems to be the easiest way - https://github.com/yaronn/blessed-contrib/blob/master/examples/line-random-colors.js just need to put in the mqtt part and update the data that way.
  33. ai escape room: this is an idea of dave build an escape room controlled entirely by ai the only way out is by interacting with the machine it can control everything: heating, doors, whatever
  34. programmable themepark: here’s a ride; how you interact with it is defined via your own programs you play minigolf, but instead of a club you use programming @sordina
  35. graph+weights to neuronal net rendering: https://github.com/BlueBrain/NeuroMorphoVis
  36. Arbitrary-Task Programming: given that programming is just arranging symbols, and we can use deep learning to interpret the real world into symbols, then it’s possible to do programming by performing arbitrary tasks i.e. any job can be programming, if we can build the deep-learning system that converts actions in that job into symbols in a programming language
  37. Sitcom-Lang: it’s a pre-processor, or something, for an arbtrariy language whenever a symbol is defined, that symbol is embued with a soul and a “nature”. it starts to have wants and needs; and those must be satisfied in order for it to stay in it’s present form (i.e. as a “for loop”), otherwise it might change (i.e. to be a “while” loop, or maybe even a string instead). all the symbols will interact with each other, and in that way a program will be made @sordina
  38. brain2object2teeshirt: this - https://scirate.com/arxiv/1810.02223 - but once it’s decided on the object, it gets rendered on your LED tee-shirt @geekscape @sordina
  39. pix2pix sourcecode2cat: generate pictures of source code convert to pictures of cat instant machine for generating cat pictures from code what cat does your code look like? @sordina
  40. physical xmonad: use the uarm to be a “physical” xmonad you want to write on some piece of paper? no worries, the uarm will re-arrange your physical desk so that everything is conveniently arranged to do that @sordina
  41. collaborative painting in the style of christopher alexander: it has 3 parts done by 3 artists i draw the left part; you draw the middle, and it has to interact coherently with whatever i’ve drawn; then another person draws the right side, again, it must interact that’s it.
  42. lego laptops: laptops that plug together in a lego-like way
  43. business version of 30 kids vs 2 professional soccer players: 30 grads vs 2 ceos 30 ceos vs 2 grads etc.
  44. shops in parks: would make the parks safer/nicer, because people would be in them more could limit the type of shops, and their size, but would be a nice way to build a bit more of a community feeling in them
  45. an icon next to your public email address that indicates how many unread emails you have: then people can gauge what will happen if they email you
  46. ml for configuring linux: “what file should i look at to change default font sizes?” “how can i set up my new gpu? what settings should i set?”
  47. water-t-shirt: the essence of a water-bed, in t-shirt form!
  48. deep antiques roadshow: the idea explains itself
  49. a being that is by-default inherently abstract, instead of inherently practical like us: for them, being practical would be really hard by default they live at the other end of the abstraction spectrum
  50. business card bowl: through all the business cards into a bowl; each day, call them, if they don’t want to do business, throw the card out
  51. “live blog”: whenever someone visits your blog, instead of reading articles, they get to open a chat window with you in your terminal then you tell them what you’ve been up to; and they can ask you questions
  52. software art: take all the source code; stack them up as if the line count is one slice, that’s the structure
  53. t-shirt whiteboard: in essence i.e. you can draw on the t-shirt, and the writing just washes out the next time then you can design whatever you want would this just work?
  54. physics simulation + diagrams: would be great to define things such as “two pieces of rope inter-twined”, and then “drop” them, but then let that resulting expression become a haskell diagrams Diagram, so that you can then do diagram stuff with it
  55. git version flattener: clone a git repo at every revision, into some folder.
  56. see-through jacket that is also warm: optionally also magnetic @sordina
  57. magnetic glass: @sordina
  58. heated keyboard: keeps your fingers warm
  59. run an experiment where monkeys/dogs/whatever are encouraged to learn some kind of programming to solve a task: i.e. a monkey gets 1 food package per day but if learns to program, using the tools provided to it, ( something like a giant physical version of scratch ) then it gets 3 food packages in some sense people have tried this, with them solving problems, but has anyone tried it where the tool they use to solve the problem is general, and can be applied to other areas of their life?
  60. tree to code: physical trees 1. order trees by the number of leaves 2. order code by the number of statements train a deeplearning network to map between these things then, trees can write computer programs @sordina
  61. ethical algorithms testing ground: related to the last two #409 #408 basically, people can sign up to be ethical tester algorithms can join to provide games for people to play how would it work?
  62. ethical testers: beta testers game testers ethics testers
  63. simulation for ethical machine learning problems: consider the situation: “how do i know if this algorithm X is unethical?” well, instead of waiting for the salespeople to tell you, you could just have it run in a simulated environment and see if it’s unethical by the way that it acts.
  64. minecraft file browser: walk around your filesystem in 3d
  65. ocr clipboard copy and paste: select an image region, send it to some text api thing, get the text back in the clipboard
  66. low-powered de-colourisation network: learns to convert colour -> black and white if it doesn’t do a good job, it’ll look awesome
  67. physical quine: a robot that can type on the computer and write code that writes the program that writes itself
  68. deep learning “do” networks: can you include the “do calculus” into neural networks somehow? to make it do some causal things?
  69. plot websites on cesium map, browse the internet that way.: web-world
  70. animated colour schemes for vim: the colour scheme rotates as you code
  71. a tale of three dickens: the movie: it’s an auto-generated movie from locally-coherent slices instead of the book, we make a movie, where all the scenes in the movie are interspersed based on “local coherence” i.e. from two movies select two people having a conversation with someone named bill or, flick between scenes at the beach @sordina
  72. revolutionary walls: the floor is fixed; but the walls are a tube you pull on some part of the wall to rotate it @tall-josh
  73. activation function composer: or more generally, a function composer 1. what does the graph of relu look like? 2. what about the graph of relu . tanh ? and so on, indefinitely and arbitrarily. some features: - what points should be push through? maybe could add certain kinds of initialisations and ranges - add things like drop-out and whatnot.
  74. record videos of people doing interviews but have their voice replaced by obama and their image replaced by obama
  75. hair cut & deep learning deal with the hair-dresser across the road: sit down for a hair cut, get an hour of deep learning consulting as well
  76. live action star wars playing out across many websites in the background of cesium js windows: on my website, a death-star is driving around on it’s way somewhere eventually it reachers your website, and destroy’s it’s logo, or something
  77. deep learning tcp or upd: find something inbetween
  78. meta-search in google: “i want to see all the alternatives to cloud-ranger” it’s impossible to do this search.
  79. umbrella-scarf / fresh-scarf: it’s a scarf but also, it has a hood that you can pull up, maybe even a clear hood, that let’s you see out front of it, but keeps you under cover could also keep smoke out of your face
  80. meme-net: watch video, extract meme i.e. and the rollsafe guy
  81. ultimate computer setup person: someone who just has the worlds best computer set up everything works no data is duplicated whole operating system exists in 1.5 gig; they’ve got 510 gig free no conda/ruby/stack issues
  82. codebase -> readme: looks at an entire codebase; learns to predict the readme
  83. divangulation theorem for websites: @sordina surfaces can be triangulated websites can be divangulated what are the associated theorems?
  84. tasting plates for saas, *aas: instead of just saying “sign up now for 6 months free”; just auto-sign people up for x free things, then let them use it up. easy way to get a billion more dollars for your saas business. @sordina
  85. self-skating skateboard: it drives down to the skate park; skates around on the pipes, does flips, 180s, griding, whatever. @sordina @tall-josh
  86. different password entry forms: 1. any password you type logs you in, but they all take you to a different computer. only your password takes to you yours. “honeypassword” 2. you password consists of the actual letters, but also the timing between the letters @sordina 3. any key you press is irrelevant, all that matters is the spacing; everything is done via morse-code (@geekscape)
  87. congratulations!!!
  88. meeting chaos monkey: every time a meeting is scheduled, a random attendee is replaced by some other random staff member
  89. consulto the consulting teddybear: @sordina “that sounds good in theory” “have you tried kan-ban’ing that” “moving forward that sounds good, but right now i think we should be pragmatic”
  90. small magnets in fabric that can attach to other magnets so-as to create customisable clothing: just put a diff design on by switching out the magnets i just need some small magnets. jaycar sells them
  91. “collaboration card”: some way of listing and engaging with people in various projects you’re interested in
  92. nlp self-defending thesis
  93. rent factor charged in the city based on how innovative your store is: hairdresser: f = 0.85 funky clothing store: f = 0.6 some weird shop that only sells whatever: 0.2 cafe: 1 or some kind of scheme like so
  94. e-fabric: like e-ink, but for fabric
  95. clothes that change colour with respect to the magnetic fields that are around it
  96. grand designs: of computer programs: follow the development of some kind of app, over a few years. hahahaha would be terrible.
  97. giant magnet that aligns all the spins of the atoms of objects (people?!) so that they can pass through each other with different polarisations
  98. dance-curve net
  99. shoes that look like little cars: volvo shoes monster-truck shoes lamboghini shoes fi-car shoes etc.
  100. augmentation reality glasses that convert what people are saying into words that float in front of you that you can read: so you can “hear” what people say to you when you’re wearing headphones
  101. “html/css layout searcher”, like visual image search, but for how to lay things out with flex/css/react/whatever: input: some scribble about how you want your content laid out in boxes: output: the css/html that achieves this. there’s some networks that do this already, where they convert the drawings to code. but maybe that can be augmented by thinking of it like a search across already-existing content?
  102. “relax ai” or “mindful story ai”: it makes up nice stories, like “you are walking on the beach, you see a small turtle; you follow the turtle for a swim in the water …” could also use cool accents of people, and make sure the story is consistent with another NLP after the first generative run
  103. comedy audience that instead of laughing they just say the things people say when they think something is funny: instead of “ahahaha” audience (in unison): “that’s funny” audience (in unison): “good one” audience (in unison): “great joke”
  104. Adversial NLP: a sentence so similar to another sentence as to be humanly-indistinguishable, but makes the AI think it’s something wildly different
  105. Inflatable Whiteboard Room: it’s a large room, inflatable like a balloon or whatever, but you can walk into it and use the internals of it as a whiteboard useful for offices
  106. Collaborative Password Manager: say i want to make a password for a system you will control, but we both need access to maybe i can have my program generate part of it; you’re program generate part, then combine them both on our independent computers, without the entire password leaving either of them could build this on top of the public keys on github somehow; so i just pick the github user i’m going to share a password with could clearly do this immediately by encrypting it with their public key, or something. but maybe something richer can be done
  107. Video Issues: https://vimeo.com/265518095
  108. Faux Project Management Generator Thing: it’s an RNN that generates hundreds of tickets in trello or jira or whatever; with arbitrary due dates makes you feel stressed @sordina could be used for project-management training scenarios
  109. quantum cppn
  110. AIWS: ai for aws you: “hey, i need to computer with whatever to be up at whichever.com and to have some database, blah blah” aiws: “no worries, that’s set up for you!” alt. “talky-form for AWS” @sordina
  111. deep haircut mirror: a mirror infront of hair-dressers that lets you look at potential haircuts on your own head
  112. train a network to learn when to laugh in response to jokes: deep-audience
  113. dance led prompt device: it’s a little led board that sits at the front of a dance thing, like a teleprompter, but for dance it puts out the next dance moves a dance-prompter move-prompter
  114. Easter Egg Evangelist for Enterprises (E^4): A floating employee who embeds on teams to consult on how to best add easter-eggs to the features they build.
  115. self-driving food truck: @martiningram
  116. Stabbucks: Starbucks for knives. * https://i.imgur.com/1ZCIQnh.jpg * Order venti, grande, etc knives
  117. BrainRank: A leaderboard of brain-shaped logos.
  118. DeliveryNet: Reads prose with impeccable timing.
  119. Rant Meetup: Rant about stuff that sucks. * No solutions allowed * Surely james has something to say
  120. submit an AI-entry to every large festival in melbourne in a single year: https://whatson.melbourne.vic.gov.au/Whatson/Festivals/Pages/festival_calendar.aspx
  121. Stochasm: Metal band that plays random notes. * Easy to swap out band members!
  122. Seinfreinds: Have the cast of one sitcom act out an episode of another and see if anyone notices.
  123. hire a comedian to come along to your meetings: they can provide background entertainment me: “hey nice to meet you, this is my associate jerry seinfeld, let’s get started” jerry: “what’s the deal with peanuts?” …
  124. stacked double-coffee-cup holder: it’s just a handle, that holds on to two cups, one above the other useful for carrying multiple cups
  125. Auto-generating face detection camouflage: Aka, auto-generating styles from https://cvdazzle.com/
  126. use the technology of marina (ShapeTex) to make little movable people in jackets: https://www.linkedin.com/in/marina-toeters-a55a685/
  127. “studio gan”: it just makes up every single thing, much like #341 , but in more depth and for everything could use for #343 for example.
  128. the journey of your parcel: imagine you’re waiting for a parcel from auspost you put on your VR headset and you get a real-time view into it’s life; maybe it’s sitting on a boat, on it’s way here, or it’s in an airplane, or it’s driving etc. you’d get a full HD video-style image of the thing moving, that would be completely imagined by a gan or something.
  129. menu democracy: buy a coffee, earn 1 voting right to change the menu in some way buy more coffees, proceed in this fashion other food yields you more votes
  130. dynamic videos built on the fly to answer standard google queries: i.e “use python requests to do post request” a video could be made on the fly using the celeb-generating stuff of stack-gan, then the voice-simulation stuff of lyrebird or whoever, then the lip-moving stuff, the text-to-speech of wavenet or whatever, and some other random backing scene gans and music production networks it would get the content by reading the first answer it finds on google, in some summarised way. @sordina
  131. fully-automated fashion design: 1. Fashion-MNIST CPPN - At random, pick a random item of clothing, figure out what it is, and generate a large version. 2. Pick a random (creative-commons) photo from Flickr, train a style transfer network on it. 3. Apply the style transfer to a bunch of different clothing items? To make a theme? 4. Pick a name from an RNN? 5. Upload to PAOM? Run-time should be several hours for one collection? Not so bad.
  132. remote-controlled magnet: a perfectly spherical magnet that can be rolled around by remote.
  133. use cppn to generate a 3d landscape by determining the height by the colour
  134. lunch formation yaml specification: example: lunch: - sandwhich: - bread - butter - lettuce - cucumber - butter - bread region: cbd elements are ordered by height on the plate. @sordina
  135. a tale of 3 dickens: combine: 1. a christmas carol 2. a tale of two cities 3. great expectations in order page-by-page. @sordina
  136. instead of colouring in the retro-haskell tee with colours, print the source code for the program itself in the previous colour: easy!
  137. reverse twospace - use offices for other purposes out of hours: silverpond -> t-shirt business on the weekends
  138. dance karaoke: like karaoke, but instead of singing you need to dance uses some pose-recognition thing
  139. build a markov chain thing and then run all the words through some “smoothing” operation by way of a word embedding: i.e. somehow pick a few lines within embedding space, and move all of the words closer to those lines maybe something would happen
  140. naive nn visualisation: just reshape all the weights to be in the shape of an image, normalise the values, and output it.
  141. map sentences to “the gist”; just a few words: “an embedding that compresses a piece of text to its core concepts” “like if i can compress an image” “then i should be able to compress a book” “and if i can do that that means that i can also write a compressed book” “and have the neural network write my book” @gacafe
  142. version number which, in ascii, eventually approaches the source code itself
  143. personal world map: it’s one of those scaled world maps, where the scaling is determined by say your gps coordinates over a given year, so that it only enlarges the places you go. @mobeets
  144. water doughnut
  145. Deep-Can-I-Do-Deep-Learning-Here?: it’s a network for which you input a situation and it tells you if you can use deep learning to help.
  146. haskell type journey challenge: get from package x to package y using only the following types once ….
  147. make the 3d wall art that we saw at the house of sonya
  148. novels in binder-form so that you can take out small sections of the pages and read them
  149. multi-agent learning where the agents also watch each other locally and learn from each other
  150. ml for learning the life/centers function from christopher alexander: two pictures which one has more life? alt. something about centers?
  151. bureaucratic-net: instead of a network that is really good at explanationing it’s decisions, this network is really bad at it. nothing it says makes sense, or alternatively it’s really long-winded in it’s responses. or maybe it’s always right, but it never has any idea why.
  152. artistic-arxiv: instead of papers, each day take a random few images from every paper and show that. maybe it’d be cool.
  153. DeepWiki: on normal wikipedia, humans edit pages about concepts in the form of words on deepwiki AIs edit concepts in the form of embeddings by way of adjusting the vectors (or something) they’d need to think about how to manage edits and revisions and so on. but that’s the general idea. @sordina
  154. secret walls: wear the streets (or: graffiti on a wall of clothes; and wear them)
  155. a network that is given the punchline and has to work out the setup: @icp-jesus
  156. reverse website or inverse website: normally, you visit a site and see the website and you can view source to see the source what about if you could visit a site and see the source, then view the source to see the site
  157. endless pasta hat: has a self-pesto’ing tube that pushed out a long piece of spaghetti that you can munch on.
  158. in the gan setting, the discriminator isn’t needed when generating, maybe there’s another setting where the discriminator is still useful at the generative stage?
  159. a jacket that makes amazon’s automated shopping thing think you’re a packet of chips: or something similar
  160. CompromiseApp: two people need to agree on something they both have the app person 1 rates the estimated compromise, on a scale, of person 2 person 2 likewise both people record their own true compromise values then, over time, there’s a bunch of things that can be done, such as comparing predicted compromises, total compromises made, etc.
  161. DerivativeNet: it watches all seinfield episodes and sees if it can generate curb your enthusiasm episodes it reads all smbc comics and sees if it can generate xkcd ones etc.
  162. Cap-Gun mechanical keyboard: You pull back a bunch of hammers then as you type it fires the caps.