TLDR AI 2024-09-03

Developers only spend 25% of their time writing code. To be effective, AI must be incorporated across the entire SDLC. GitLab Duo is the only AI solution that supports developers at every stage of their workflow.

Your AI companion across the SDLC.

Generate tests, explain code and chat in real-time.

Fortify your code with AI-powered security.

Understand and remediate vulnerabilities with explanations and auto-generated merge requests.

AI that prioritizes data privacy.

You control access to AI-powered capabilities. Your code and data are never used to train AI models.

Learn more about how GitLab Duo supports AI-driven workflows across the SDLC.

xAI has started the 100,000 Colossus H100 training cluster, the largest in the world, with plans to double it in size in a few months.

Large language models sometimes fail at tasks like counting letters due to their tokenization methods. This highlights limitations in LLM architecture that affect their understanding of text. Nevertheless, advancements continue, such as OpenAI's Strawberry for improved reasoning and Google DeepMind's AlphaGeometry 2 for formal math.

Nvidia's new Blackwell chip demonstrated top per GPU performance in MLPerf's LLM Q&A benchmark, showcasing significant advancements with its 4-bit floating-point precision. However, competitors like Untether AI and AMD also showed promising results, particularly in energy efficiency. Untether AI's speedAI240 chip, for instance, excelled in the edge-closed category, highlighting diverse strengths across new AI inference hardware.

Researchers have developed an advanced machine learning method for identifying both direct and indirect causal variables of a target within a system.

This study presents a multimodal classification model designed to improve the early diagnosis of Alzheimer's Disease.

Diffusion models and auto regressive models are often thought to be fundamentally different forms of methods. However, they may actually be more similar than we thought where diffusion models are actually taking auto regressive steps in the frequency domain.

Bfloat16 and fp8 support for native numpy arrays.

X-Codec is a new approach to enhancing the performance of audio generation by LLMs.

VisionTS is a new approach that leverages the similarities between natural images and time series data to enhance forecasting by reformulating time series forecasting as an image reconstruction task. It uses a visual masked autoencoder (MAE) pre-trained on ImageNet to achieve impressive zero-shot performance.

Don't wait for the AI future—build it. This September, join the world's largest gathering of open-source AI infrastructure leaders and an all-star lineup of speakers from industry-leading companies. Come grow your network while exploring the cutting edge of AI with Ray and Anyscale. Register to attend in-person @ The San Francisco Marriott Marquis

While commonly monitored, GPU utilization may not accurately reflect true GPU performance in ML workloads, as it doesn't account for whether the GPU's computational power is fully leveraged. Trainy discovered this while only achieving ~20% model FLOPS utilization (MFU) despite 100% GPU utilization during LLM training. It recommends tracking SM efficiency for a better performance indicator and achieving a 4x speedup in training time by optimizing with fused kernels and the right model parallelism level.

AI training is growing at an unprecedented 4x per year, outpacing mobile adoption and genome sequencing rates from past technological expansions. Research indicates scaling AI training could feasibly continue until 2030, constrained primarily by power availability and chip manufacturing capacity. Training runs up to 2e29 FLOP may become possible, marking substantial progress akin to the leap from GPT-2 to GPT-4, provided hundreds of billions are invested. Data scarcity and latency also pose challenges, but these may be surmountable through multimodal and synthetic data generation strategies and advanced network topologies.

Boardwalk Robotics has unveiled Alex, a humanoid upper torso robot designed for tasks in manufacturing, logistics, and maintenance. Developed independently while leveraging the pedigree of IHMC's bipedal robot expertise, Alex is legless, with its creators emphasizing manipulation over mobility to ensure safety and efficiency. Commercial partners are currently being selected for pilots, while researchers can purchase Alex immediately.

Timm is a package of vision models.

CogVideo 5B model will launch next week in Hugging Face Diffusers.

A study from MIT Media Lab found that generative chatbots powered by large language models significantly increased the formation of false memories during simulated crime witness interviews, inducing over three times more immediate false memories than a control group.