Parse notes
Sponsored
Join the AI infrastructure revolution NEXT MONTH in Salt Lake City! KubeCon + CloudNativeCon North America [4] brings together leaders in AI, open source, and cloud native technologies for four intensive days of knowledge sharing.
(individual or corporate), granting entry to KubeCon + CloudNativeCon North America + 15 CNCF-hosted co-located events. Claim your offer
Google has introduced voice recognition for 15 additional African languages across its platforms, including Voice Search, Gboard talk-to-type, and Translate dictation. This update empowers approximately 300 million more Africans to interact with digital content in their native languages.
MIT researchers have introduced Heterogeneous Pretrained Transformers (HPT), a new model architecture inspired by large language models, to train versatile robots using data from various domains and modalities.
Cohere announced that its Embed 3 AI model is now multimodal, enabling fast and accurate search across important enterprise image data sources like graphs, charts, product catalogs, and design files. With this upgrade, Embed 3 is the most generally capable multimodal embedding model on the market.
Novel view synthesis has long been thought to require strong 3D inductive biases. This work shows that with scale and some weak inductive bias, you can dramatically outperform previously thought limits.
Auto regressive models still dominate in many cases. However recent work on diffusion heads for images has inspired continuous AR diffusion. This work extends the per token diffusion idea to variable length outputs.
This paper introduces a change description instruction dataset to finetune LMMs for better remote sensing change detection.
Sponsored
AI costs money. Enforce user limits and prevent abusive bots using Arcjet security as code.
IC Light is the current best way to relate an image with a pretrained text-to-image backbone. This discussion is the beginning of extending that functionality to the strong Flux models.
Generating 3D scenes from scratch is challenging for a number of reasons, such as data constraints. This work introduces a programming-like language to describe 3D scenes and shows that Claude Sonnet can generate extremely compelling scenes despite not having been explicitly trained on this task.
FtD++ is an approach for cross-modal learning that improves unsupervised domain adaptation for 3D semantic segmentation.
Anthropic recently released two works that highlight its new interpretability method. This post is an open replication of the crosscoder on Gemma 2B.
This repository lists papers on graph out-of-distribution learning, covering three primary scenarios: graph OOD generalization, training-time graph OOD adaptation, and test-time graph OOD adaptation.
LLMs function as sophisticated autocomplete systems, predicting the next token based on training data and current input. Tiny changes in input affect predictions, leading to different outputs even for the same question. Understanding token prediction, training data context, and memory limitations can help refine AI use.
Sponsored
Mintlify is the modern standard for public-facing documentation: beautiful out of the box, easy to maintain, and designed to convert
OpenWebVoyager provides tools, data, and models for developing multimodal web agents capable of navigating and learning through real-world web interactions.
Researchers have unveiled a novel inclusion-matching approach that addresses the challenges of automated colorization especially in animations where occlusions and wrinkles can disrupt traditional segment matching.
A dataset of music clips and detailed text descriptions generated by a music generative model.