TLDR AI 2024-10-31

DeepMind has talked a little bit more about the audio generation models used to power NotebookLM.

OpenAI has released the SimpleQA benchmark, which measures models' abilities around simple factual questions.

This study explores using feature steering in AI models to interpretably modify outputs. It reveals a "steering sweet spot", where changes do not degrade capabilities. The study results show steering can alter social bias in targeted domains but also brings unexpected off-target effects. Further research is required to refine feature steering for safer, more reliable outcomes in AI models.

Thunder Kittens is a framework for writing extremely performant GPU Kernels. It is built on the idea that GPUs actually want to operate on small 16x16 tiles of data. In turn, the useability is quite high, and 40% faster kernels only take a few hundred lines of code.

MeshRet has introduced a novel approach for improving motion retargeting for 3D characters that focuses on preserving body geometry interactions from the start.

Researchers have enhanced Masked Generative Models (MGMs) with a new self-guidance sampling method, improving their image generation quality while maintaining diversity.

This project introduces PiToMe, an algorithm that compresses Vision Transformers by progressively merging tokens after each layer. This method reduces the number of tokens processed.

PF3plat tackles the challenge of pose-free 3D reconstruction and novel view synthesis from RGB images, eliminating the need for extra data.

SELECT is the first large-scale benchmark for comparing data curation strategies in image classification. ImageNet++ is a new dataset that extends ImageNet-1K with five new training-data shifts, each assembled using different curation techniques.

Assembled uses LLMs to accelerate and improve software testing, enabling test generation in minutes instead of hours. This approach increases engineering productivity, saving time and shifting focus to feature development. LLMs generate comprehensive and accurate tests that maintain code quality and development velocity.

A CNET survey revealed that only 18% of smartphone users are motivated by AI features to upgrade their devices, with privacy and cost being significant concerns. Major manufacturers like Apple, Google, and Samsung are integrating more AI capabilities in their phones, yet many users prioritize battery life and storage over AI functions. AI subscriptions are set to become common, but nearly half of users are unwilling to pay for these features.

BitNet, developed by Microsoft Research, introduces a transformer architecture that reduces LLM computational and memory requirements by using ternary precision (-1, 0, 1) equating to 1.58 bits per parameter. Models are required to be trained from scratch. BitNet can also fine-tune existing models to this low-precision format, maintaining strong performance on downstream tasks. This approach significantly reduces energy consumption and improves inference speed using specialized kernels for efficient matrix multiplication.

Rime AI, who trains custom speech synthesis models with over 200 distinct voices, achieved <300 milliseconds p99 API latency with perfect uptime after switching to Baseten for model inference. Read Rime's story.

Google's Project Jarvis, powered by Gemini 2.0, aims to automate web-based tasks in Chrome by using AI agents capable of reasoning and planning.

Concerns have risen over OpenAI's Whisper introducing hallucinations in transcriptions, even in medical contexts.

Kepler has unveiled the Forerunner K2 humanoid robot, which has advanced AI, improved hardware, and enhanced vision and navigation systems for better real-time interaction.