TLDR AI 2024-01-23

Stability AI's latest release, Stable LM 2 1.6B, is a compact yet powerful language model that supports seven languages. It is designed to outperform other models under 2 billion parameters, including its own previous 3B model. With its smaller size, it aims to lower barriers for developers, offering various versions including a unique "half-cooked" model for more customization flexibility.

Tesla has started rolling out its Full Self-Driving Beta v12 update, which shifts vehicle control from explicit C++ code to an AI-powered single neural network. The release marks a significant move towards fulfilling the company's self-driving ambitions, although the software is still labeled as beta. Skepticism will remain until tangible improvements in autonomous driving capabilities are observed as the software is rolled out to cautious select beta testers.

This work shows how you can use diffusion to clone faces in images. There are many ways to do this with deep fakes, but diffusion is interesting due to its ability to inpaint other pieces of the image as needed.

LoRAs are Low Rank Adapters that allow you to fine-tune only a small amount of parameters in a language model. They can dramatically improve and alter the performance of these models.

Grouping in 3D is a challenging and ambiguous task because you don’t know what granularity you’ll need for the grouping operation (e.g., keys on a keyboard vs the entire keyboard itself). This work uses multi-level masks and makes great progress on the problem of semantic 3D grouping.

Retrieval augmented generation (RAG) might be catching all the AI buzz, but how do you know that your RAG system is actually working?

Join Gretel's CPO and AI research team and Microsoft's AI Platform Architect for this free virtual workshop to learn how you can use Gretel’s Tabular LLM to create diverse synthetic datasets for RAG evaluation.

Sign up (free) to submit your questions for the live Q&A

This project explores Self-Rewarding Language Models, where the model acts as its own judge to improve training. By using this method, researchers were able to enhance a model's performance, surpassing other systems like GPT-4.

DiverseEvol is a new method that lets an AI model choose its own training data to get better without needing help from people or other advanced AI systems.

Depth Anything is a new approach to monocular depth estimation that relies on a large-scale dataset of about 62 million images to improve its accuracy. By using data augmentation and auxiliary supervision from pre-trained encoders, this model achieves impressive generalization capabilities and sets new standards in depth estimation.

This author revised their AI timeline predictions based on advancements in scaling up models. They are now estimating a 10% chance of achieving Artificial General Intelligence by 2028 and a 50% chance by 2045. These changes are attributed to the effectiveness of large language models and the realization that many intelligent capabilities may emerge at scale.

Researchers introduce the "Prompt Highlighter," a method that revolutionizes text generation in multi-modal language models by allowing users to highlight parts of prompts.

Acquiring high-quality datasets is critical for AI and ML. Third-party data - and specifically aggregated news articles - can accelerate data collection. Read more on Dow Jones Factiva

MM-Interleaved is a new generative model that excels in processing and creating interleaved image-text data.

Turn your codebase into Wikipedia-style articles using AI.

Microsoft CEO Satya Nadella is content with the current state of Microsoft's partnership with OpenAI, focusing on stability rather than control amid regulatory reviews in the EU and UK. He emphasizes Microsoft's independent AI developments alongside their significant investment in OpenAI.