TLDR AI 2024-06-06

OctoAI's experts will show you how to avoid "model fatigue" and pick the perfect open-source model for your project.

In this video, you'll learn how to:

• Narrow down the vast universe of models to a shortlist tailored to your needs
• Decide between checkpoints, LoRAs, and Community tunes for your specific use case
• Evaluate model success across prompts, data, training, and tuning

➡️ Give it a watch here!

Mistral has launched customization for its models via its platform and API.

Cartwheel, a motion synthesis start-up by our very own Andrew Carr, has raised a seed round led by Accel and KV. The company has built a text-to-animation tool and powerful integration into existing 3D workflows.

Google has scaled back AI Overviews. The AI-generated answers are now appearing in around 11 percent of queries, down from around 27 percent when the feature was launched last month. The company is still in the process of refining the feature. AI Overviews are appearing most in health-focused searches, despite the company highlighting the tool's use for shopping, restaurant recommendations, and travel in its marketing.

The operations research group at Google improved mathematical optimization for cargo shipping routes. They found a 13% reduction in fuel usage and costs.

Researchers have introduced LlamaCare, a fine-tuned large language model (LLM) for medical knowledge, along with Extended Classification Integration (ECI) to address classification issues in LLMs.

Dream Mat is a great way to improve texture generation for 3D objects. It uses a number of classical graphic techniques such as Albedo, Metallic, and Roughness to render an extremely compelling output given a 3D mesh.

Researchers have found a way to eliminate matrix multiplication (MatMul) from large language models while maintaining strong performance, even at the billion-parameter scale.

XRec is a model-agnostic framework that leverages the language capabilities of large language models to enhance explainable recommender systems.

Security teams have long struggled to understand what's in their cloud environments and their exposure to risks. As part of Orca's move towards AI-driven cloud security, the platform now answers natural language questions such as “do we have any S3 Buckets that allow public READ access?”. Read more and see examples on the Orca blog

GenS is an end-to-end generalizable neural surface reconstruction model that excels in reconstructing surfaces from multi-view images without 3D supervision.

OpenAI's recent brain drain and focus on consumer products complicates its path to competing with big tech.

Stability AI has released the weights to Stable Audio which was trained to generate sound effects on permissively licensed audio samples.

The future of AI foundation models will be closed-source due to the centralizing forces of data and compute, which favor economically viable and secure closed models over open-source alternatives.

Generate studio-grade videos — complete with custom lighting, varied camera angles, and different backgrounds — without the studio.

New findings from BrightEdge Generative Parser reveal the primary industries impacted by AI Overviews, what triggers results, and where Google automatically anticipates and answers search queries.

CV-VAE is a video VAE compatible with existing image and video models that offers spatio-temporally compressed latent spaces, enabling efficient training of new video models using pre-trained ones.

Customers expect intelligent experiences in real time. Learn how to make that possible with SAP Datasphere and cloud-native Apache Kafka. Watch the Confluent webinar now