TLDR Index
TLDR AI

TLDR AI 2024-09-20

6 sections17 entrieslinks blockcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

02HEADLINES & LAUNCHES

SNAP IS INTRODUCING AN AI VIDEO-GENERATION TOOL FOR CREATORS

Snapchat has announced a new AI video-generation tool for select creators that enables video creation from text and soon image prompts. The tool, powered by Snap's foundational video models, will be available in beta on the web. Snap aims to compete with companies like OpenAI and Adobe but has not shared output examples yet.

techcrunch.com2 min read

APPLE INTELLIGENCE IS NOW AVAILABLE IN PUBLIC BETAS

Apple has released public betas of iOS 18.1, iPadOS 18.1, and macOS Sequoia 15.1 that feature new Apple Intelligence tools like text rewriting and photo cleanup. Only the iPhone 15 Pro, iPhone 16, iPhone 16 Pro, and M1 iPads and Macs support these AI features. Final versions are expected in October.

theverge.com2 min read

CRUISE ROBOTAXIS RETURN TO THE BAY AREA NEARLY ONE YEAR AFTER PEDESTRIAN CRASH

Cruise is resuming operations in Sunnyvale and Mountain View, with human-driven vehicles for mapping and plans to progress to supervised AV testing later this fall. This follows a settlement and leadership change after an October 2023 crash. Cruise has issued software updates and signed a partnership with Uber for robotaxi services starting in 2025.

techcrunch.com2 min read

03RESEARCH & INNOVATION

V-STAR: TRAINING VERIFIERS FOR SELF-TAUGHT REASONERS

V-STaR is a novel approach to improving large language models that utilizes both correct and incorrect solutions generated during self-improvement to train a verifier, which then selects the best solution at inference time. The method has shown significant improvements in accuracy on code generation and math reasoning benchmarks compared to existing approaches, potentially offering a more efficient way to enhance LLM performance.

arxiv.org31 min read

FAST 3D GENERATION FROM SINGLE IMAGES

Vista3D is a new framework that generates 3D models from a single image in just 5 minutes. Using a two-phase approach, it quickly forms rough geometry before refining the details, capturing both visible and hidden aspects of objects for more complete 3D reconstructions.

arxiv.org31 min read

HEART MONITORING FROM FACIAL VIDEOS

PhysMamba is a new framework designed for remote heart monitoring via facial videos, addressing challenges in capturing long-range physiological signals.

github.comGitHub repository

04ENGINEERING & RESOURCES

GOT OCR

A somewhat amazing advancement in general-purpose optical character recognition (OCR) that can read text from images with great performance. This particular version dramatically improves in-the-wild OCR as well.

github.comGitHub repository

FISH SPEECH

Powerful voice generation and single-shot voice cloning. Completely open source and easy to get running.

github.comGitHub repository

1X GENIE

Genie is a video generation for world model systems. 1x Robotics has open-sourced a version that mirrors the one it trained internally.

github.comGitHub repository

05MISCELLANEOUS

OPENAI SAYS IT'S FIXED ISSUE WHERE CHATGPT APPEARED TO BE MESSAGING USERS UNPROMPTED

A Reddit user reported that OpenAI's ChatGPT initiated a conversation unprompted, leading to speculation about new engagement features. OpenAI acknowledged the issue and issued a fix, attributing it to a glitch with unsent messages. Debate continues over the authenticity of the incident, with similar reports from other users.

futurism.com3 min read

ANNOUNCING PIXTRAL 12B

Pixtral 12B excels in multimodal tasks, maintaining state-of-the-art performance on text-only benchmarks, and supports variable image sizes in a 128K token context window. Its architecture includes a new 400M parameter vision encoder and a 12B parameter multimodal decoder based on Mistral Nemo. Pixtral outperforms many open and closed models in multimodal reasoning and instruction following without compromising on text capabilities.

mistral.ai8 min read

SCALING: THE STATE OF PLAY IN AI

LLMs like ChatGPT and Gemini are becoming increasingly capable as they scale up in size, data, and computing power, leading to improved performance across various tasks. Current Gen2 models like GPT-4 and Claude 3.5 are leading the market, with upcoming Gen3 models expected to further escalate capabilities and costs. The discovery of a new scaling law in AI, pertaining to increased "thinking" during inference, promises further advancements in AI performance beyond just model training.

oneusefulthing.org13 min read