Search the archive

ADOBE PREVIEWS ITS UPCOMING FIREFLY VIDEO MODEL

Adobe's Firefly Video Model brings AI-powered features to video editing software like Premiere Pro. The new model, available in beta later this year, offers editors enhanced workflows to explore creative ideas, fill timeline gaps, and add new elements to footage.

MISTRAL RELEASES PIXTRAL 12B, ITS FIRST MULTIMODAL MODEL

French AI startup Mistral has launched Pixtral 12B, a 12-billion-parameter multimodal model capable of processing both images and text. Available via GitHub and Hugging Face, the model can be fine-tuned and used under an Apache 2.0 license. Its release follows Mistral's $645 million funding round and positions the company as a significant player in Europe's AI landscape.

FASHION DATASET FOR PERSONALIZATION

This dataset, generated using large language models, tailors outfits to different occasions, styles, and body types, offering high-quality and relevant suggestions.

3D SCENE RECONSTRUCTION

Researchers are enhancing 3D scene reconstruction methods like Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (GS) by introducing uncertainty estimation techniques. These methods, while delivering high-quality renders, struggle with handling uncertainties from noise, occlusions, and camera inaccuracies.

EXTENDING LLMS' CONTEXT LIMIT

This paper introduces Hierarchical cOntext MERging (HOMER), a new approach designed to extend the context limit of large language models without requiring additional training.

September 25.

LLAMA OMNI

Llama Omni is a full speech in-out model based on Llama 3.1 8B that can run at extremely low latency and still deliver high quality responses.

AWS AI STACK

A ready-to-use, full-stack boilerplate project for building serverless AI applications on AWS. It is a great fit for those seeking a trusted AWS foundation for AI apps and access to powerful LLM models via Bedrock that keeps your app's data separate from model providers.

INTERNET OF AGENTS FOR MULTI-AGENT COLLABORATION

Internet of Agents (IoA) is a new framework designed to improve multi-agent collaboration by integrating diverse third-party agents more effectively.

BOOSTING MODELS WITH DIVERSE GENERATIVE DATA

DiverGen is a new strategy for creating generative datasets to improve instance segmentation models. Unlike costly manual annotations, it uses generative models to produce diverse data, addressing overfitting and enhancing model performance.

ELON MUSK SAYS TESLA HAS ‘NO NEED' TO LICENSE XAI MODELS

Elon Musk has denied reports that Tesla will share revenue with his AI startup xAI to use its AI models. He clarified that Tesla has benefited from xAI engineers but doesn't need to license xAI's models. Musk emphasized that xAI's large models cannot run on Tesla's vehicle computers.

APPLE IS THINKING ABOUT A RIVAL TO META RAY-BAN GLASSES

Apple may develop non-AR smart glasses, potentially competing with Meta's $299 Ray-Ban glasses, which don't have AR capabilities. Meta's glasses include features like a camera and an AI chatbot. With less complexity, Apple's non-AR glasses could be cheaper, lighter, and offer better battery life.

fine-tune AI anywhere. See the specs

OPENAI IN TALKS TO RAISE FUNDS AT $150B VALUATION

OpenAI is in talks to raise $6.5B from investors at a valuation of $150B.

ELL

Ell is a new package from an ex-OpenAI scientist for managing prompts as code.

EMOTIVE PIANO MUSIC GENERATION

This work uses a two stage model to disentangle emotive performances in piano music generation.

get you directly in front of your target audience. Learn more

about running your first campaign with us.

GPT-FABRICATED SCIENTIFIC PAPERS ON GOOGLE SCHOLAR

Generative AI tools like ChatGPT are increasingly producing fraudulent research papers, infiltrating databases like Google Scholar alongside legitimate studies. These papers, often on controversial topics like health and the environment, pose significant risks to scientific integrity and public trust. Enhanced vigilance and more robust filtering in academic search engines are essential to curb this growing issue.

APPLE ANNOUNCES ITS NEW A18 AND A18 PRO IPHONE CHIPS

Apple unveiled the A18 and A18 Pro chips at its “Glowtime” event, promising significant CPU and GPU improvements over the A16 Bionic. The A18 Pro boasts enhanced memory bandwidth and upgraded image-processing capabilities. Both chips support advanced AI features, with the A18 Pro notably improving on-device model performance and thermal design for better gaming experiences.

AMD ANNOUNCES UNIFIED UDNA GPU ARCHITECTURE

AMD announced at IFA 2024 that it will unify its RDNA and CDNA architectures into a combined UDNA microarchitecture, aiming to better compete with Nvidia's CUDA ecosystem. This strategic move seeks to streamline development and bolster AMD's position in AI and HPC markets. The transition to UDNA is a pivotal step, with full-scale implementation expected beyond the upcoming RDNA 4 generation.

CAN LLMS GENERATE NOVEL RESEARCH IDEAS?

This longitudinal study assessed 100 NLP researchers' abilities to write novel ideas and then review those ideas. It found that LLMs can generate more novel ideas, but these ideas were slightly less feasible than the ones produced by human research scientists.

SUPERHUMAN AUTOMATED FORECASTING

The Safe AI Institute has released research on a system that can outperform human experts at forecasting.

ADEMAMIX OPTIMIZER

This paper from Apple proposes an alternative to traditional exponential moving average optimization that uses contributions from older gradients to dramatically improve learning convergence.

data. Save your spot!

OPTIMIZING INFERENCE PROXIES FOR LLMS

Optillm helps to optimize prompts based on a number of proven research algorithms like Monte Carlo Tree Search, Z3 solvers, and Self Consistency.

3D SEMANTIC SEGMENTATION

Researchers addressed the challenge of source-free unsupervised domain adaptation for 3D semantic segmentation by introducing regularization techniques and a novel criterion.

MEMORY-EFFICIENT OPTICAL FLOW

HCVFlow is a new memory-efficient optical flow method that tackles the high computational cost of all-pairs cost volumes in high-resolution images.

CONCEPT SLIDERS

Concept Sliders are a powerful way to control the output of diffusion models. Work has been done to make them work with the new Flux suite of models.

MINIFYING HTML FOR GPT-4O: REMOVE ALL THE HTML TAGS

Stripping HTML to plain text reduces costs with minimal performance loss in GPT-4o for data extraction tasks. Tests on the Mercury Prize dataset showed that GPT-4o performs well without HTML structure and that GPT-4o mini is a cost-effective choice for unstructured questions. For structured extraction tasks, consider testing both versions to balance cost and accuracy.

WAYMO GIVING 100,000 ROBOTAXI RIDES PER WEEK BUT NOT MAKING ANY MONEY

Waymo is providing over 100,000 paid autonomous rides weekly in San Francisco, Phoenix, and Los Angeles, doubling its May numbers. Despite its growth, Waymo remains unprofitable, with Google's experimental division incurring a $2 billion operating loss. High costs for vehicles and city mapping, combined with public hesitancy, continue to challenge its path to profitability.

IOS 18.1 WITH APPLE INTELLIGENCE LAUNCHES IN OCTOBER, MORE LANGUAGES ROLLING OUT OVER TIME

Apple announced that Apple Intelligence will enter beta with iOS 18.1 in October, initially only for US English users.

SEGMENTATION OF INFECTED AREAS IN CHEST X-RAYS

SGSeg is a segmentation framework for chest X-rays that uses language guidance during training but enables text-free inference.

HOOP (PRODUCT LAUNCH)

Hoop is an AI task management app for professionals that connects to tools like Google Meet, Zoom, and Slack.

See what you can build with email data →

WHAT IS APPLE INTELLIGENCE, WHEN IS IT COMING AND WHO WILL GET IT?

Apple introduced Apple Intelligence at WWDC 2024, focusing on integrating AI features into existing apps like Mail, Messages, and Siri. The platform uses large language models for tasks such as text summarization and image generation, aiming for a seamless user experience. A beta version will roll out in the U.S. this October, with broader international support planned for 2025.

HERE'S HOW 7 NEWS AUDIENCE DIRECTORS ARE THINKING ABOUT GOOGLE'S AI OVERVIEWS

Google's AI Overviews, powered by the Gemini language model, faced heavy criticism for inaccuracies and dangerous suggestions after its U.S. launch. Despite the backlash, Google expanded the feature to six more countries, raising concerns among publishers about reduced traffic and misrepresented content. AI strategists and SEO experts emphasize the need for transparency and better citation practices to maintain trust and traffic.

NEW OPEN SOURCE AI LEADER REFLECTION 70B'S PERFORMANCE QUESTIONED

HyperWrite's Reflection 70B, a variant of Meta's Llama 3.1 LLM, is facing scrutiny after third-party evaluators couldn't replicate its claimed performance. Issues were attributed to model weights being corrupted during the upload to Hugging Face, which led to discrepancies. The AI community awaits further responses and updates to clarify the model's actual capabilities.

IMITATING LANGUAGE VIA SCALABLE INVERSE REINFORCEMENT LEARNING

Much of modern language modeling is a specific case of imitation learning. Fortunately, there is a depth of study in the field broadly. This paper explores how to use inverse reinforcement learning to imitate entire sequences instead of just tokens. The results are promising and point in the direction of continued RL in language model training pipelines.

VISUAL GENERATION WITH PROXY TOKEN DIFFUSION TRANSFORMER

Proxy Token Diffusion Transformer reduces redundant computations in diffusion transformers by using fewer representative tokens for attention.

3D OBJECT DETECTION FOR INDOOR ENVIRONMENTS

UniDet3D is a powerful 3D object detection model trained on multiple indoor datasets.

Swahili, and many more. Watch the Oxford Dictionaries API tour

FAST 3D RECONSTRUCTION FROM MULTIPLE IMAGES

A great tool that uses Mast3r and some clever optimizations to reconstruct 3D scenes from a few 2D images.