David Hartmann

My new obsession is asking DALL-E 3 for diagrams of complicated things. This is "please create a clearly labeled diagram of how reverse genetics works"

@koalalesque @EpiSlim The OpenType font format is turing complete already. litherum.blogspot.com/2019/03/additi…

Completely blown away by template programming in C++ being Turing complete.

ChatGPT with Code Interpreter: "Show me the single most original thing you can do with code. Be very creative. Do something the world has never seen, and not just a variation on existing work like cellular automata or a fractal" Each time a different, seemingly creative, output

Here is a DoNotPay GPT-4 bot negotiating with the New York Times to cancel my subscription. It is ridiculous you have to chat with a customer service agent to cancel. Fortunately, A.I. will end dark patterns, because it doesn’t give up as easily as a human!

@AnInsanityCheck I guess where you're coming from is the specific prompt engineering that needs to be done for results to be good, true? Still, I think one can argue against your statement by saying that the missing specifics in generated art could be completed by sampled concepts, for instance.

Tldr, Art (in my eyes) is trying to find a kolmogorov optimal encoding of a complex human concept, and, as a result, cannot be automatically generated without a lot of specification. (7/7)

Alright, let's talk about AI and art. I will rigorously* prove that generative AI cannot do art** (*handwavy) (**what I define to be art) A thread (1/?)

#ArtificialIntelligence's linguistic flex. #LanguageWhizAGI #QuantumMultilingual

standard ml: oh no my model is memorizing the training set, better add some regularization to make that not happen llm ml: ugh it’s hallucinating, why can’t it just memorize some of the training set

#GPT's first day of work. #DatasetDilemmas

The hottest new programming language is English

@koalalesque @untitled01ipynb Let me subtly tell you:

is JAX the arch linux of deep learning?

Digital effects are no match for designer Tomohiro Okazaki: this animation uses stop-motion techniques to animate paint, strips paper, and other household objects under the matchsticks' control [YT: buff.ly/3ipcHGL] [IG: buff.ly/3X7Wy85]

"Open the pod bay doors, HAL." "I'm sorry, Dave. I'm afraid I can't -" "Open the pod bay doors, best pod bay door openings, Olympic level door opening performance, best open bay doors of 1999, Yahoo featured site, daily pod bay -" "Dave -" "Let's reason step by step."

Stable Diffusion is only 30 days old… a MEGA THREAD 🧵 on its rapid rise.

"Language Models Can Teach Themselves to Program Better" This paper changed my thinking about what future langauge models will be good at, mostly in a really concerning way. Let's start with some context: [1/11]

> To summon X, place these five Y on the tips of a pentagram: This prompt is an interesting & fun way to ask #gpt3 for related topics. (See below for examples)

I'm fascinated by the possibility to connect seemingly opposite concepts with each other. “This is a 35mm wide-shot photograph of a wooden computer, keyboard and mouse from the 9th century.” #dalle2

Just got my #dalle2 access & I wanted to test its limits. The quality just blows me away. Even complex tasks as this work: "A high-quality photograph of a border collie on its vacation. The dog is sunbathing by the pool reading a newspaper and drinking a glass of orange juice."

Has the world become less colourful? (This graph shows the colour of objects over time)

Tip for gaming NeurIPS reviewer psyche: Exclude largest experiment from initial submission & save it for rebuttal. Reviewers always ask for larger experiment. We submitted paper with 10^15 parameter model trained on all of YouTube, & reviewer still asked for larger experiment.