TLDR Index
TLDR AI

TLDR AI 2024-02-02

5 sections17 entrieslinks blockcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01HEADLINES & LAUNCHES

NOUS RESEARCH RELEASES HERMES DATASET

OpenHermes-2.5 data has been used to train some of the best open models out there. The dataset contains 1M+ examples of high-quality data points. It has now been released.

huggingface.co3 min read

BARD CAN GENERATE IMAGES

Bard has now expanded its service to more languages and launched an image generation service.

blog.google4 min read

THE AMERICAN DYNAMISM 50: AI

a16z put together this list of 50 startups tackling some of America’s most pressing problems around defense, energy, transportation, and manufacturing. All of them are using AI in some form to accelerate their efforts. If you’re curious about real-world applications of AI, this is a great insight.

a16z.com30 min read

02RESEARCH & INNOVATION

OLMO: OPEN LANGUAGE MODEL

AI2 released a fully open language model training framework. The models come with comprehensive resources including full training data, model weights, training and evaluation code, over 500 checkpoints for each model, and future plans for fine-tuning code, all released under the Apache 2.0 License. These resources are provided by AI2’s Dolma and WIMBD. They include evaluation frameworks Catwalk and Paloma and support for analysis and fine-tuning through Weights & Biases logs and HuggingFace revisions. The models are competitive with the best in class.

blog.allenai.org12 min read

MOBILEDIFFUSION

Google has showcased a latent consistency diffusion model with some architectural changes that it trained to run on mobile devices with sub-second generation times.

blog.research.google21 min read

ENHANCED VIDEO-TEXT LEARNING

Shared Network Pre-training (SNP) improves learning of video and text together. This method is more efficient and versatile than previous models and includes a unique strategy called Significant Semantic Strengthening (S3) for better understanding of key words in sentences.

arxiv.org14 min read

03ENGINEERING & RESOURCES

Sponsored

WORKSHOP: BETTER RAG WITH SYNTHETIC DATA

RAG is all the rage. Get insider tips on how to build more robust RAG in this upcoming virtual workshop [11]! Learn how to use Gretel’s Tabular LLM [11] to create diverse synthetic datasets for RAG evaluation. The session will be led by Sam Kemp, Principal AI Platform Architect at Microsoft, and the Gretel

info.gretel.aiSponsor

NEW STATE OF THE ART IN OPEN EMBEDDING MODEL (HUGGINGFACE HUB)

The BGE-M3 project introduces a versatile embedding model that excels in Multi-Functionality (dense, multi-vector, and sparse retrieval), Multi-Linguality (supporting over 100 languages), and Multi-Granularity (handling inputs from short sentences to documents up to 8192 tokens). It uses a hybrid retrieval pipeline that combines different methods and re-ranking for higher accuracy and generalization, leveraging its ability to perform embedding and sparse retrieval simultaneously.

huggingface.co

HIERARCHICAL TEXT SEGMENTATION

The Segment Anything Model (SAM) has been enhanced to create Hi-SAM, a model specializing in hierarchical text segmentation. Hi-SAM excels in segmenting text across multiple levels - from strokes to paragraphs - and even performs layout analysis.

github.comGitHub repository

RAGS

RAGs is a Streamlit app that lets users create RAG pipelines from data sources using natural language. Users just have to describe the tasks and parameters they want from their RAG systems. The RAG can be queried and it will answer questions over the data.

github.comGitHub repository

04MISCELLANEOUS

BLOCKCHAIN-POWERED SOLUTION TO HELP IMPROVE TRANSPARENCY AND AUDITABILITY FOR GENERATIVE AI SYSTEMS

IBM Consulting and Casper Labs are working together to build a pioneering AI governance tool powered by blockchain. Built leveraging IBM’s watsonx.governance platform, it will introduce AI version control, advanced monitoring, permissioning, and access controls to AI models for enterprise use. This could be essential for businesses looking to de-risk AI and harness its true potential while adhering to internal or external guidelines and regulations.

casperlabs.io5 min read

YOU ARE NOT LATE TO ARTIFICIAL INTELLIGENCE

Just as the early days of the internet offered immense opportunities, today's AI landscape holds similar potential. The key is to shift perception to see the possibilities in existing technologies like AI and follow passion and joy to explore various paths.

thealgorithmicbridge.substack.com4 min read