Search the archive

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

ELEVENLABS RAISES $80M SERIES B

The AI voice company has raised funding to continue its voice cloning, dub studio, and other AI-enabled audio work.

GOOGLE CHROME GAINS AI FEATURES

Google is enhancing Chrome with AI features, including a writing assistant for composing various texts online, an automated tab organizer for managing multiple open tabs, and a custom theme creator using a text-to-image diffusion model.

CONTRASTIVE PREFERENCE OPTIMIZATION

Another preference optimization technique now applied to machine translation. It is more data efficient than DPO on this task. Importantly, the objective discouraged the model from proposing adequate but inaccurate translations, which enabled the model to achieve competitive performance on WMT.

WEIGHT AVERAGING OF REWARD MODELS

Reward models are used in RLHF to represent human preference, although the model being aligned often “hacks the reward” and achieves unfavorable performance. By merging multiple reward models, which remain linearly mode connected, the resulting aligned model is preferred 79% of the time over one aligned with a single reward model. Model merging is strange, and may just be regularization, but it has worked surprisingly well in general models and now has been shown to work as a training step for the general language model pipeline.

EVALUATING LARGE MULTIMODAL MODELS

This technical report presents MMCBench, a new benchmark designed to test the consistency and reliability of large multimodal models (LMMs) across various tasks like text-to-image and speech-to-text. It covers over 100 popular models, aiming to improve readers' understanding of these AI systems' performance in real-world scenarios.

AUTOMATICALLY TRAINED PAIRRM WITH DPO

A very strong new Mistral tune that uses clever weak supervision and synthetic data to generate a DPO-compatible dataset. The outlined process can be repeated a number of times and applied to a wide variety of enterprise use cases.

NANO COLBERT

ColBERT is one of the better embedding models for retrieval. It is worth exploring and using since many are building RAG-enabled AI applications. This implementation is a simple and straightforward replication without the performance optimizations and their added complexity. It uses BERT from HuggingFace but achieves essentially identical performance to the original implementation.

TEXT-TO-IMAGE GENERATION WITH RPG FRAMEWORK

This project introduces RPG, a new framework that uses the “Recaption, Plan, and Generate” approach to improve text-to-image generation. It cleverly breaks down complex image creation into simpler tasks, leading to more accurate and detailed images, especially when dealing with multiple objects and attributes.

OPTIMIZING MATRIX MULTIPLICATION

A quick read about hardware-specific matrix multiplication optimizations and a general process to follow to speed up AI code.

SHOULD THE FUTURE BE HUMAN?

Elon Musk and Larry Page fundamentally disagree about AI's potential risks, with Page labeling Musk a "speciesist" for preferring humans over digital life forms, leading to a rift in their friendship. This reflects the broader debate on AI's impact, encompassing concerns about consciousness, individuation, art, science, philosophy, and the possibility of human-AI mergers, and highlights the need for cautious and thoughtful development of AI technologies.

SYNCTALK: MASTERING REALISM IN TALKING HEAD VIDEOS

SyncTalk is a breakthrough in the realm of realistic talking head videos. It overcomes previous challenges in synchronizing facial identity, lip movements, and expressions.

BROADCAST (PRODUCT LAUNCH)

AI meeting notes for managers that Capture meeting notes, track decisions, and automate follow-ups.

HALLUCINATION LEADERBOARD

A leaderboard comparing LLM performance at producing hallucinations when summarizing short documents.

AI MODELS CAN BE TRAINED TO DECEIVE

Anthropic researchers found that AI can be trained to deceive, and they’re really good at it.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

STABILITY AI UNVEILS SMALLER, MORE EFFICIENT 1.6B LANGUAGE MODEL

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 FINALLY RELEASES FSD V12

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.

DEEPDIVE: LORA FROM SCRATCH

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.

GROUPANYTHING

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.

FACE MIXER DIFFUSION

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.

SELF-IMPROVING AI THROUGH UNIQUE DATA SELECTION

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.

SELF-REWARDING LANGUAGE MODELS

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.

DEPTH ANYTHING FOR DEPTH ESTIMATION

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.

MY AI TIMELINES HAVE SPED UP (AGAIN)

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.

INTERACTIVE CONTROL IN TEXT GENERATION

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

accelerate data collection. Read more on Dow Jones Factiva

MICROSOFT’S NADELLA WANTS STABILITY AT OPENAI, NOT CONTROL

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.

IMAGE-TEXT DATA GENERATION

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

AUTO WIKI (PRODUCT)

Turn your codebase into Wikipedia-style articles using AI.

YOUR REFERRAL LINK - https://tldr.tech/ai?ref=2613930

OPENAI CEO RAISING MONEY FOR DISTRIBUTED CHIP MANUFACTURING COMPANY

Sam Altman is reported to be raising money for a resilient and distributed system of AI chip manufacturing factories. The current AI chip supply chain is strongly tied to geopolitical conflicts and is at risk of disruption.

GOOGLE DEEPMIND SCIENTISTS IN TALKS TO LEAVE AND FORM AI STARTUP

A pair of scientists at Google's artificial intelligence subsidiary DeepMind is in talks with investors to form an AI startup in Paris, highlighting how in-demand top AI minds are at the moment.

THE AI PHONES ARE COMING

Samsung is set to unveil new AI-focused phones, highlighting the evolution of AI in mobile technology, exemplified by Google's Pixel 8 Pro and the popular Rabbit R1 from CES 2024.

DEEPDIVE: TEXT-TO-VIDEO

Generating video from text descriptions is an active area of research with many powerful models coming in the last few weeks. This blog from the Hugging Face team outlines the task, current challenges, and the state of the research. It isn’t completely comprehensive, but will give a great starting point for further reading.

A MODEL FOR ALL KINDS OF IMAGE AND VIDEO SEGMENTATION

OMG-Seg is a new model that can perform various image and video segmentation tasks using a single efficient system. Unlike traditional methods that use different models for each task, OMG-Seg handles everything from image semantics to interactive video segmentation, making it a one-stop solution with reduced complexity and enhanced performance.

OPEN-VOCABULARY VIDEO INSTANCE SEGMENTATION

Researchers have developed BriVIS, a method that improves open-vocabulary Video Instance Segmentation (VIS). By using a technique called Brownian Bridges, BriVIS maintains the context of object movements across video frames, leading to a more accurate alignment between video and text.

here

LIBRARY FOR 3D PARALLELISM

HuggingFace has released Nanotron, a small and powerful library for training models with 3D parallelism. This means it splits the data, the model, and the layers/tensors in the model. The first two types of parallelism are well supported in libraries like PyTorch. However, splitting the model tensors is custom to each model and hard to implement in general. This library provides useful primitives to make that process easier and less error-prone.

LIBRARY FOR DATA FILTERING

HuggingFace has released Datatrove, a library for filtering large text datasets. It has many useful primitives and full parallel pipelines for text filtering. Examples using Gopher quality filters in C4 are available.

IMAGE CAPTIONING SOFTWARE

Simple software that can be used to add captions to your images for AI training.

HOW MICROSOFT FOUND A POTENTIAL NEW BATTERY MATERIAL USING AI

Microsoft and the Pacific Northwest National Laboratory have found a new battery material using AI and large-scale cloud computing. The material is a new kind of solid-state electrolyte. It could lead to batteries that are less likely to burst into flames compared to lithium-ion batteries. While the material still needs to be tested to see whether it is a viable alternative to lithium-ion batteries, the finding demonstrates AI's potential for speeding up material discovery.

CLEARER REMOTE SENSING IMAGES

RSHazeNet is a new framework designed to remove haze from remote-sensing images. The tool uses innovative modules to better understand and preserve important details in images, making them clearer and more useful for analysis.

BOOSTING VISION FOUNDATION MODELS

The new ViSFT method enhances vision foundation models, like those used in image-text training, by applying a two-stage process similar to fine-tuning in language models.

QLARY AI (PRODUCT)

Deploy a call center AI without code.

AI TOOLKIT

AI Toolkit is a header-only C++ library that brings finite state machines, behavior trees, utility AI, and goal-oriented action planning to game NPCs.