Parse notes
Generative AI exploded onto the scene in 2022, bringing with it buzzy startups and funding rounds. Insider asked investors to name the most promising generative-AI startups of 2023. Notable companies include foundational models like Open AI and Stability AI, next layer tooling like LangChain and Fixie, and early stage startups like Cursor and DeepScribe.
Sam Altman, co-founder and chief executive officer of OpenAI, said the organization is looking at opening a Japan office and expanding Japanese language services after meeting with Prime Minister Fumio Kishida.
Context length is a hot topic in AI. We're trying to scale text models to longer and longer sequences. This work scales an RNN to a million context tokens. It's not directly applicable, but there are some lovely nuggets here.
Believable proxies of human behavior can empower interactive applications ranging from immersive environments to rehearsal spaces for interpersonal communication to prototyping tools. In this paper, the authors introduce generative agents--computational software agents that simulate believable human behavior. In an evaluation, these generative agents produce believable individual and emergent social behaviors: for example, starting with only a single user-specified notion that one agent wants to throw a Valentine's Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time.
Researchers from Google propose Video-kMaX, a unified approach for online and near-online Video Panoptic Segmentation, comprising a within-clip segmenter and a cross-clip associater. By leveraging clip-kMaX and HiLA-MB for these components, this method establishes a new state-of-the-art on KITTI-STEP, VIPSeg, and VSPW datasets, bridging the gap between online and near-online segmentation approaches.
A challenge in training large text models is fitting them on your hardware accelerators. This is often easier on TPUs than GPUs due to Pjit capabilities of Jax. This library (which was used for the imminently popular FlanT5 models) is easy to use and seems to scale quite well.
A faux pilot based local coding assistant for vs code that runs local with 4GB ram. It uses the Salesforce 6B codegen model.
Sponsored
It seems like a new LLM API launches every other day. But how do you know which is the best fit for your use case? Is bigger always better? What are the benefits of specialized models? How does model size affect latency? This blog post by Inworld [https://bit.ly/3nXnURp] identifies 8 key factors for comparing large language models. Read the article [https://bit.ly/3nXnURp]
OpenAI will start paying people as much as $20,000 to help the company find bugs in its artificial intelligence systems.