Emerging AI-driven search tools challenge Google by offering direct and explorative answers, but struggle to match its speed, diverse functionalities, and efficient data presentation, underscoring the complexity of replacing traditional search with AI.
Databrix and Mosaic have trained a 132B parameter MoE model with impressive performance. They trained the model on 3,000 H100s and have released the weights. The model is also available on the Databricks API.
AniPortrait is a framework designed to create lifelike animated portraits from a single reference image and audio input. By translating audio into 3D representations and then mapping these onto 2D facial landmarks, this method produces animations that excel in natural facial expressions, varied poses, and high visual quality.
Searching over embedding vectors is a key to RAG pipelines. If you replace the fp32 numbers with a single 0 or 1, then use a KNN clusterer and reranker, you can maintain performance while shrinking memory requirements 30x.
This comprehensive survey delves into the advancements and challenges of deepfake technology and its detection, highlighting the arms race between deepfake creators and those developing technologies to spot them.
LLMs are useful in as much as they are fast, accurate, and follow directions. This combination makes a street fighter emulator with text input an excellent way to figure out which models are good at these three criteria.
The OPTIN framework introduces a novel way to enhance the efficiency of transformer-based AI models across various domains without the need for re-training. By using a technique called intermediate feature distillation, OPTIN can compress networks under specific constraints while barely affecting accuracy.
AID and its variant PAID are two techniques designed to improve image interpolation by incorporating conditions like text and poses. These methods ensure the production of images with enhanced consistency, smoothness, and fidelity without requiring additional training.
The world is grappling with the challenge of regulating AI. A series of high-profile meetings and conferences involving global leaders, tech executives, and policymakers revealed divisions and a lack of consensus on how to control this transformative technology.
New models are regularly released that claim to be the state of the art on standard benchmarks. It is important to measure these models on your own tasks and data. Superpipe is a tool that helps build these evaluation pipelines on your data.
Researchers discovered a side-channel attack that can decipher encrypted AI assistant chats with high accuracy on specific topics by exploiting token transmission within the encryption. The attack utilizes large language models to reconstruct token sequences into readable text, potentially exposing sensitive user conversations. Major AI assistants, except for Google Gemini, are vulnerable to this method, prompting providers to seek mitigation strategies.