2d:null 2e:["$","title",null,{"children":"SM3809_A9 | ychsu6 | Observable"}] 33:T438,md` ### READING

Some of the weird(silly) questions I raised in the last week’s [*realisation *](https://observablehq.com/@betahsu/sm3809_a8/2) has now be answered or explored in the [*Neural Nets Part 2*](https://towardsdatascience.com/a-beginners-guide-to-neural-networks-part-two-bd503514c71a). Sigmoid function, one of the activation functions that was mentioned by the author among many other else, allows output to give an intermediary value from the question that the system has no confident response. Rather than the binary output that are relatively far from actual human performance/ decisions making, there is such stage where the network has “exactly “no idea whether it’s a 0 or 1, and several continuous stage where the system has certain preference of giving certain output . The different level of confidence level are very similar to the reality that human tend to have only a few range of (or even overlapping) tendency, while along the process of “training” or “teaching”, the network is able to move incrementally towards the right answer. `34:Tb73,md` ### SHARING

I found this artist: Janelle Shane. This is her [*personal blog*](https://www.janelleshane.com/), which contains looooots of very interesting neural net projects that she have been working on.

The one I would specifically like to share is [*A big neural net reviews a smaller neural net’s recipes*](https://tmblr.co/ZP7VLsZK_IF70i00). She used a neural net that had already seen a lot of recipes online during internet pre-training, and the recipes themselves already look a lot more realistic than her previous works (her [*earliest recipes*]( https://aiweirdness.com/post/163878889437/try-these-neural-network-generated-recipes-at-your) generated with char-rnn). But rather than having GPT-3 generate its own recipes, she had it rate char-rnn’s recipes. She prompted GPT-3 with two example recipes from [*allrecipes*](https://www.allrecipes.com/), each followed by 3 of their human-written reviews. Then, for the third recipe she inserted one of char-rnn’s recipes. The outcome reviews seem relatively “normal” if the original char-rnn recipe is more understandable. However, the reviews tend to be more chaotic if the char-rnn recipe contains more chaos. In technical aspect, the neural net has done a successful job getting most of the vocabulary and structure work out, being aware of certain sentiments range of the reviews, and trying to mimic the writing style from the generated recipes from char-rnn etc. Besides, it is interesting that such powerful neural net as GPT-3 would try to pick up them and produce reviews that are plausibly be written by whatever weird humans are visiting this mangled recipe website. Another critical and interesting question (similar concepts are also raised in another work: [*AI AI Weirdness*](https://aiweirdness.com/post/626712039215202304/ai-ai-weirdness)) is about the downside of such accurate neural net actually decrease the level of weirdness/humour. What does it mean when the neural nets are weird/dumb because we’ve asked them to be, not because of their own failure? For example, in this case, it tried to understand a chaotic recipes and “pretend” to be less powerful as it originally does. It seems like the chaotic writing style is interpreted and produced by the networks only because it’s under command.

### some more SHARING Some other interesting works (that are related to language as well) from her: - An several neural nets that generate hilarious pickup lines: [*GPT-3 tries pickup lines*](https://tmblr.co/ZP7VLsZuM-238y00) - Getting huge neural nets to generate candy heart messages: [*Okay GPT-3: Candy hearts!*](https://tmblr.co/ZP7VLsZexXyaWy00) - [*The Epic History of the Earth*](https://tmblr.co/ZP7VLsZThRaumi00) - Machine learning algorithms that generates Christmas carols: [*Alternate Rudolphs*](https://tmblr.co/ZP7VLsZR69kjCa00) - Neural Net for [*2020 headlines*](https://tmblr.co/ZP7VLsZOs4unau00) `35:T59e,md` ### CREATING Here is the 2nd version of [*my Twine + RiTa work*](https://observablehq.com/@betahsu/sm3809_a10) (This is the one I showed on 1st of April in class).

In this second version, more interesting functoin in Twine has been explored. This version is very much inspired by the TV series "The Good PLace". Although the TV series itself focus more on the consept of After Life, I find the robot character Janet rather more interesting. Therefore, in this version, I #### To be improved:

##### Narrative:

The messy 1st, 2nd, 3rd point of view is too confusing.

More comprehensive: Fix the narrative structure (some part is missing, and the twist should be more smooth).

Fix the interactive part: too many one buttom part (forcing reader) .

More careful about the timing method between lines or when to use the typewritter effect (the horizontal dimension of narative).

##### Techinical:

Add more parts that had RiScript contained for more unpredictable effect.

Play with the special function in Twine that can remember users' previous actions/decisoins and to further influence the storytelling (and to cross out the selection that the user has been through).

##### Additional:

Add audio/images to improve the "thickness" of narrative (the vertical dimension of narative).

##### Final polishing:

Website on GitHub

Description

`36:[] 31:{"id":"0855c51e1fd1ed8d","slug":"sm3809_a9","title":"SM3809_A9","thumbnail":"8d4a73ca8f872c2a1069ec525b94f0a2c7bb0231b6973f79c6c81cfee4b06456","access_level":"public","create_time":"$D2021-03-24T20:47:43.582Z","update_time":"$D2021-04-03T12:36:47.228Z","view_time":null,"first_public_time":"$D2021-03-24T21:04:49.406Z","has_unpublished_edits":null,"fork":false,"stars":1,"starred":false,"owner":{"type":"team","id":"c3d91d7db2867b4d","login":"betahsu","name":"ychsu6","avatar_url":"https://avatars-next.observableusercontent.com/197b7792c36f626b9fe391a472486b8b6dc3cf3503a2bb52e365e7916c324a9e","tier":"starter_2024"},"version":62,"publish_version":62,"role":null,"authored":false,"authors":[{"type":"individual","id":"8205ffe96d046f8c","login":"betahsu","name":"ychsu6","avatar_url":"https://avatars-next.observableusercontent.com/197b7792c36f626b9fe391a472486b8b6dc3cf3503a2bb52e365e7916c324a9e","tier":null}],"members":[{"type":"individual","id":"8205ffe96d046f8c","login":"betahsu","name":"ychsu6","avatar_url":"https://avatars-next.observableusercontent.com/197b7792c36f626b9fe391a472486b8b6dc3cf3503a2bb52e365e7916c324a9e","tier":null,"role":"editor"}],"comments":[],"custom_thumbnail":false,"trashed":false,"lapsed":null,"subscribed":null,"forks":0,"collection":null,"type":"notebook","body":{"title":"SM3809_A9","cells":[{"id":0,"value":"md`# SM3809_A9\n## Neural Networks\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":5,"value":"$33","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":11,"value":"$34","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":41,"value":"$35","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":60,"value":"md`\nThe 1st version from last week is attached [*HERE*](https://observablehq.com/@betahsu/sm3809_a8/2) (I'm sorry for the late submission). \nIn the 1st versoin of my work, only simple RiTa techniques are applied but I played around with several Twine techniques and rules and experimented some of the non-linear storytelling methods that allows more possible permutations and combinations of scenarios on different levels of scale. For the second version, I would like to combine larger amounts of text input in different scenarios and might try to compress the amounts of pages but with richer contents (and additional audio/image elements if possible).\n`","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"}],"files":"$W36","stdlib":"1","titling":"auto"}} 32:"$31:body"