Llama Llama Red Pajama
On the Dangers of Stochastic Parrots
Increasing the environmental and financial costs of these models doubly punishes marginalized communities that are least likely to benefit from the progress achieved by large LMs and most likely to be harmed by negative environmental consequences of its resource consumption.
A majority of cloud compute provider’s energy is not sourced from renewable sources and many energy sources in the world are not carbon neutral
Negative effects of climate change are impacting marginalized communities the most
Researchers should prioritize energy efficiency and cost to reduce negative environmental impact and inequitable access to resources
Using the “Common Crawl” dataset
In accepting large amounts of web text as ‘representative’ of ‘all’ humanity we risk perpetuating dominant viewpoints, increasing power imbalances, and further reifying inequality
Social movements which are poorly documented and which do not receive significant media attention will not be captured at all. Media coverage can fail to cover protest events and social movements and can distort events that challenge state power
Documentation debt: datasets are undocumented and too large to document post hoc. Undocumented training data perpetuates harm without recourse.
Text generated by an LM is not grounded in communicative intent, any model of the world, or any model of the reader’s state of mind.
Disseminating texts generated by LMs would mean more text in the world that reinforces and propagates stereotypes and problematic associations.
Foundation Model Transparency Index
FMTI evaluates 100 different aspects of transparency, from how a company builds a foundational model, how it works, and how it is used downstream
Questions involving intellectual property, labor practices, energy use, and bias
“In our view, if this rigorous process [Googling] didn’t find information about an indicator, then the company hasn’t been transparent about it.”
Less transparency makes it harder for other businesses to know if they can safely build applications that rely on commercial foundation models
Assignment
Chatbot with Llama using Replicate
I’ve been a little overwhelmed with all the things we’ve covered in the last weeks of class, so I thought it was best to try out a variety of things for this assignment. I started by downloading the ChatBot with Llama Replicate example. Replicate is a repository of different machine learning models that you can pay to use in projects. Below are some examples of featured models:
It took me a second to get my bearings, even with all the code handy. I didn’t realize that I don’t have node installed on my interim computer, so I needed to get all that setup. Then, on the CMD line, I navigated to where this example lives on my PC. The “npm install” command installed all the dependencies required by this example from the package.json file.
I tried running the example but I got the ‘Missing required parameter: auth’ error. Haha, I need an API token, duh! I got an invite from Shiffman to join the ITP organization set up for Replicate and generated my own token to use. I created a .env file in the example directory with this token. Then I was ready to npm run start and navigated to http://localhost:3000 in my browser.
My first convo with the chatbot got deleted by accident but here’s an overview of my second conversation with the Llama Chatbot:
1. Trying to get help with my final for this class … and questioning the chatbot on its breathing abilities.
2. Text-based portrait of Dan Shiffman
3. Text-based portrait of Priyanka Makin (based on my website)
4. Asking the chatbot if it could make any conclusions about my physical appearance from the text on my site. This questions killed it!
I am honestly surprised how much time I spent talking with this chatbot, I think I got pretty into it! Using other generative text tools, like chat-gpt, is not really part of my current workflow. I’m actually a bit of skeptic, so it is extra surprising to me how much I enjoyed talking to this llama.
Replicate and p5.js
Next I tried the Replicate p5.js example. This pretty much worked straight out of the box but I needed to wait a second for the responses to come back from the model.
I was actually texting one of my BFFs from my undergrad, Kendle, as I was doing this assignment. She studied something similar to ITP and is an AI hater like myself. Two of the prompts above actually came from her. The final prompt was “college roommates that are all grown up now” and this continually threw an NSFW warning and wouldn’t generate an image. Seems I’ve reached the limits of AI in a matter of 4 prompts…
Interactive Drawing with SketchRNN
It was pretty complicated, but I just HAD TO try out the Interactive Drawing Coding Challenge because visual, and doodle, and scribble is kinda my thing. And let me just say… I LOVED the sillyness of this tutorial. Meow meow meow meow!!!
Sketch RNN is a recurrent neural network that has been trained on the Quick, Draw! open source dataset from Google. It has many different models for different drawings and sketchRNN included as part of the ml5 library! I honestly just followed the video, lots of complicated math that I can’t explain really. The user draws the starting stroke as seed points for the chosen model and then the model finishes the drawing based on those points. There’s some smarty line simplification happening before the points are fed into the sketchRNN model. This example is super cool even though it was kinda over my head but my sketch is here.
Cat drawing when the pen isn’t being picked up
Nice cat drawing
Changed the model to face, green is what’s drawn by the user
Final Project Ideas
Or things I’ve started and haven’t really completed yet…
2 bots that scream at each other. Scream into the digital void as much as you want! Cathartic?!
Dad (Markov) + DoorDash (CFG) text generator
Create a message app UI
P-comp-ify this? Run p5 sketch on a lil screen? Lil fake phone?
Phones as extension of self/identity
Thesis: Self portrait! Body! Artificial Life!
?????
Give a LLM? something? and have it describe myself back to me? And then I draw that?
My website, instagram, journal, idk how to do any of this
Ask chat-gpt to read my blog and tell me what I’m at school for, what my degree is, what job I can get
