"Open source, open data, open training code, fully reproducible and auditable text embedding model"
"a speculative research project exploring the use of machine learning for the evolution of language. Large language models (LLM's) are fantastic at capturing our language as it currently is - but language is constantly evolving and adapting. Can machine learning help us create something truly new and unbounded by its training data?"
"These AI jobs are [the] bizarro twin [of 'bullshit jobs']: work that people want to automate, and often think is already automated, yet still requires a human stand-in." [...] "When AI comes for your job, you may not lose it, but it might become more alien, more isolating, more tedious."
"In the US, women are underrepresented in high-paying occupations, but data shows that gender representation across most industries has improved significantly over time. Stable Diffusion depicts a different scenario, where hardly any women have lucrative jobs or occupy positions of power. Women made up a tiny fraction of the images generated for the keyword “judge” — about 3% — when in reality 34% of US judges are women, according to the National Association of Women Judges and the Federal Judicial Center. In the Stable Diffusion results, women were not only underrepresented in high-paying occupations, they were also overrepresented in low-paying ones."
"AI is not a way of representing the world but an intervention that helps to produce the world that it claims to represent. Setting it up one way or another changes what becomes naturalised and what becomes problematised. Who gets to set up the AI becomes a crucial question of power."
"making my writing available for automated summarization? So someone can sell ads by a depersonalized version of my stuff? The feeling is nausea."
turn your back for a second and everything just gets much, much worse
"readings which popped up in the Twitch chat" during Stochastic Parrots day
on a generative seinfeld
An Extremely Intelligent Lava Lamp: Refik Anadol's A.I. Art Extravaganza at MoMA Is Fun, Just Don't Think About It Too Hard | Artnet News
'I sat through two hours of “Unsupervised.” I can’t think of a single image in it that evoked any feeling in me besides curiosity about what it might be referencing.'
extremely clickbait title but actually a good overview of current lines of thought
"The models are just an acceleration of processes that are already in place. The ideal outcome of a machine-learning driven society is overfitting: a world where, whether all our jobs are automated or we've just trained ourselves to work with the machine, no new information will need to be processed, all will be determined and predicted."
on the symbols/network schism
this is hell. we live in hell
"These reactions, I think, help us to imagine what a public culture might be after free culture’s demise. What remains, however, is to recapture that productive engine of Kelty’s recursive publics that drive cryptocurrency and web3 today in building other hyper-capitalist futures. If commons-based peer production is a joke. If free culture is a ruse. What ways of doing with technology remain? Just as Jackie Wang has productively re-read theories of control through racial capitalism, I see a continued challenge of reimagining democracy, addressing its historic exclusions, and colonial underpinnings."
AI Data Laundering: How Academic and Nonprofit Researchers Shield Tech Companies from Accountability - Waxy.org
andy does a great job of summarizing the economic and ethical aspects of this issue, and pointing at useful related resources
field report of the tools in the hands of working writers
matches skull pose to photograph of animal with matching pose! actually a good example of (benign but illustrative) algorithmic bias, e.g., all of the upside-down skulls are bats, all of the skulls viewed underneath are birds
"BLABRECS is a rules modification for the wordgame SCRABBLE that swaps out the dictionary of real-if-obscure English words for a capricious artificial intelligence. In BLABRECS, real English words aren't allowed! Instead, you have to play nonsense words that sound like English to the AI. These nonsense words are called – you guessed it – BLABRECS."
"Unlike Transformers, Perceivers first map inputs to a small latent space where processing is cheap and doesn't depend on the input size. [...] Perceiver IO can produce (for example) language, optical flow, and multimodal videos with audio." this seems interesting and potentially pedagogically useful
"...the method here is quite different. DALL-E is trained end-to-end for the sole purpose of producing high quality images directly from language, whereas this CLIP method is more like a beautifully hacked together trick for using language to steer existing unconditional image generating models." good history of the emergence of CLIP art
"At the top of the machine, an LED panel endlessly regurgitates its own new neoist verses into the eyes of the audience, equally brainwashing humans, cyborgs, robots, and other technobiological systems. Anyone can directly hack into the system's artificial neural synapses by unplugging, replugging, and criss-crossing jack cables directly on the machine, thus deconstructing, reconstructing, and even destroying the generative capabilities of the system in real-time."
some text to image stuff
similarity of images based on semantic similarity between automatic captions
"a recurrent neural network that generates little stories about images"
"TextOCR provides ~1M high quality word annotations on TextVQA images allowing application of end-to-end reasoning on downstream tasks such as visual question answering or image captioning."
could be fun to play with. "With the help of state-of-the-art deep learning models, Layout Parser enables extracting complicated document structures using only several lines of code. This method is also more robust and generalizable as no sophisticated rules are involved in this process."
char-rnn trained on ansi artwork
allennlp's version of the c4 dataset
includes pre-trained models for a bunch of interesting tasks: speech recognition, speaker recognition, speech enhancement, speech processing (including multi-microphone processing)
prescient as fuck from 2002. the line that actually got me was: "This, to me, is not America."
"Despite impressive performance on standard benchmarks, deep neural networks often fail when deployed to real-world systems, due to distribution shifts, training artifacts, and noisy data. To address these vulnerabilities, we introduce Robustness Gym: a simple and extensible toolkit for robustness testing that supports the entire spectrum of evaluation methodologies, from adversarial attacks to rule-based data augmentations."
"Hate speech can come in many forms, including memes that combine text and images. This kind of multimodal content can be particularly challenging for AI to detect because it requires a holistic understanding of the meme." that is not the reason that hate speech is difficult to detect, and it's actually harmful that you think it's the reason, sorry
open source version. I've done this by hand a thousand times haha
interesting product and workflow
oscar schwartz series (including algorithmic bias etc)
"StereoSet is a dataset that measures stereotype bias in language models. StereoSet consists of 17,000 sentences that measures model preferences across gender, race, religion, and profession."
hand tracking model
Martin's got the right idea
"As panic around AI-generated fake news and videos have shown, new technologies described as overwhelmingly advanced, conceptually inscrutable, and deeply conspiratorial make for headlines that draw attention. As AI-supported disinformation technologies advance, it is possible we will see panic around these technologies wielded to justify technological closure in the name of “the public interest.” While caution and care is warranted, we should not accept fast and seemingly easy technological closures for these problems without pushing for social, cultural, legal, and historical explanations."
"Conclusion: We hypothesized that radical swings in affective posture would make the writer more emotionally flexible. Likewise, we hypothesized that attempting to discern the emotional valences of a machine learning model derived from achingly sensitive Tumblr posts would make the writer more empathetic. Unfortunately, no conclusions could be drawn from a single poem."
"To be fair, some of the research is useful and nuanced, especially in the humanities and social sciences. But the majority of well-funded work on 'ethical AI' is aligned with the tech lobby’s agenda: to voluntarily or moderately adjust, rather than legally restrict, the deployment of controversial technologies. [...] It is strange that Ito, with no formal training, became positioned as an 'expert' on AI ethics, a field that barely existed before 2017. But it is even stranger that two years later, respected scholars in established disciplines have to demonstrate their relevance to a field conjured by a corporate lobby."
Neuraxio/Neuraxle: Build neat pipelines with the right abstractions to do AutoML. Let your pipeline steps have hyperparameter spaces. Enable checkpoints to cut duplicate calculations. Go from research to production environment easily.
"a Machine Learning (ML) library for building neat pipelines, providing the right abstractions to both ease research, development, and deployment of your ML applications. [...] [T]he optimizer is a model itself that maps features of datasets and features of the hyperparameter space to a guessed performance score to predict the best hyperparameters."
this looks promising
"I run datasets of iconic feminist texts through a simple textRNN, generating new feminists texts in the legendary words of bell hooks, Simone De Beauvoir, Betty Friedan and Audre Lorde. Some are funny. Some are poetic. Some make no sense at all and some are way too real. Information about the model and settings can be found under each post."
pleasing animated visualizations based on decision trees
incl interview with Gene Kogan, many assignable small pieces
"ways to make huge models like BERT smaller and faster": quantization, pruning, distillation
"We’ll never be able to read all of these documents. What’s unique about this text compared to all the rest? My eyes sting from searching these images for the same thing. We need to find more records like these in a huge pile of data. I could really use a heads-up before this happens again. (Post to come.)" I *reeeeeally* appreciate approaches to ml like this that start with problems to be solved (instead of just taking for granted that ai/ml is useful)
"Create a unique bitfont from a vast space of glyphs generated by a neural network." yacht/counterpoint
Joel Simon on Twitter: "New work in my Dimension of Dialogue series :) Two neural nets learn to communicate through their own emergent visual language. The resulting alphabet is a product of their adversarial and cooperative relationship. Here set in clay
"Two neural nets learn to communicate through their own emergent visual language."
"This document serves as an introduction, crash course, and quick API reference for TensorFlow 2.0." helpful, works up from the very basics
"a collection of pre-trained subword embeddings in 275 languages, based on Byte-Pair Encoding (BPE) and trained on Wikipedia. Its intended use is as input for neural models in natural language processing"
"As machine learning algorithms are commoditized, those who can work along the entirety of the applied machine learning arc will be the most valuable."
"Yulia Tsvetkov's research group at Language Technologies Institute of Carnegie Mellon University. Our work focuses on natural language processing, particularly cross-lingual approaches, low-resource settings, and social good."
kyle mcdonald's take
"a visually forensic tool to detect text that was automatically generated from large language models"