Google presents LUMIERE A Space-Time Diffusion Model for Video Generation paper page: huggingface.co/papers/2401.12… Demonstrate state-of-the-art text-to-video generation results, and show that our design easily facilitates a wide range of content creation tasks and video editing

Fun story from our internal testing on Claude 3 Opus. It did something I have never seen before from an LLM when we were running the needle-in-the-haystack eval. For background, this tests a model’s recall ability by inserting a target sentence (the "needle") into a corpus of

Fascinating! MAML to make LLMs NOT learn fast. If true/robust, it could really help with AI safety. Also, great name! "Results verify that fine-tuning SOPHON-protected models incurs an overhead comparable to or even greater than training from scratch." arxiv.org/pdf/2404.12699

So why does it matter that glycans are genetically encoded? Glycans can be >50% of protein’s weight, cover >50% of protein surface, and vital in immunology, metastasis, and autoimmune pathogenesis. Yet, many biologists ignored them because they are hard to measure and control

It was lovely talking alife, open endedness and generative modeling with Bert Chan and Yingtao Tian in tokyo, thanks for being such great hosts!

My first PhD paper!🎉We learn *diffusion* models for code generation that learn to directly *edit* syntax trees of programs. The result is a system that can incrementally write code, see the execution output, and debug it. 🧵1/n

continually impressed by the creative, original work coming out of this company

Google presents Open-Endedness is Essential for Artificial Superhuman Intelligence - Argues that the ingredients are now in place to achieve openendedness in AI systems - Claims that such open-endedness is an essential property of any ASI arxiv.org/abs/2406.04268

Reminds me of small world LSTMs from OpenAI in 2017 openai.com/index/block-sp… The brain is modular, our models should be too

“Foundation model” is a poor metaphor for LLMs. A 3-year old is a foundation model. A 3-year old has a foundation of skills and basic concepts upon which almost anything can be built. An LLM is a multiple-personality mega-vac - it literally sucks up every concept and

Quality-Diversity + Lenia is a match made in heaven. ❤️ With Antoine Cully, we are excited to introduce Leniabreeder, a framework designed to automate the discovery of diverse artificial life. 🧬 🌐 Website: leniabreeder.github.io 📄 Paper: arxiv.org/abs/2406.04235 🧵 1/11

New paper, surprising result: We finetune an LLM on just (x,y) pairs from an unknown function f. Remarkably, the LLM can: a) Define f in code b) Invert f c) Compose f —without in-context examples or chain-of-thought. So reasoning occurs non-transparently in weights/activations!

Here's a WIP browser extension which lets you pull apart any website, create views and transformations with LLMs, add new UI, pull in context from knowledgebases and APIs, and mix pieces of multiple websites together.

New Experiment: SerendipityLM - Interactive evolutionary exploration of generative design spaces with large language models. Check out the extensive experiment report, incl. many generative art examples & link to open source repo: samim.io/studio/work/se… Find out more in this

thememorypalace.us/shadowboxing/ short episode of the memory palace about what happens to our heroes when their moment ends

"Illusion knitting produces knit planes that, when viewed head-on, appear to be meaningless and shapeless gray rectangles... walk around the mounted image, however, and magic happens: an image emerges from the noise." uwplse.org/2024/03/18/Ill…

This is Kenneth Stanley and Tim Rocktäschel on the importance of open-endedness in AI research. Edward Hughes, TimR and colleagues at Google DeepMind have just written "Open-Endedness is Essential for Artificial Superhuman Intelligence"

How did ML get fixated on the idea of using the same step size for every example? This is bad if your data has fat tails. For Cauchy-distributed data, SGD fails to converge for every setting of step-size and batch-size colab.research.google.com/drive/1KQQGJvJ…

We released the #Jamba 1.5 open model family: - 256K #contextwindow - Up to 2.5X faster on #longcontext in its size class - Native support for structured JSON output, function calling, digesting doc objects & generating citations twtr.to/giIEE #AI #LLM #AI21Jamba

Enabling LLMs to reason more deeply at inference time via search is one of the most exciting directions in AI right now. We introduce PlanSearch, a novel method for code generation that searches over high-level "plans" in natural language as a means of encouraging diversity.