TLDR AI 2023-11-10

Samsung has mostly been quiet in the AI race, but it recently showed off a live AI translation feature for phone calls. The feature is expected to launch in 2024.

Adept is building a useful agent system for enterprises. It is now opening up small mini demos for people to interact with like delegating knowledge working tasks, using new tools, or working with unstructured data.

OpenAI introduces customizable ChatGPT (GPTs), enabling users to tailor the AI for specific tasks without coding. The GPT Store, launching soon, will allow the sharing and monetizing of these custom GPTs, with a focus on privacy, safety, and expanding utility in the real world. Enterprises can deploy internal-only GPTs, enhancing business operations.

This study introduces a better way to clean up the data used to train LLMs, ensuring they are tested fairly and work more reliably.

Google ran a MaxText + Jax training job on an unbelievable amount of chips concurrently while maintaining 66+ utilization, creating a strong contender for the best possible MFU. The project was mostly built on proprietary technology, but it showcases some interesting tidbits for other large-scale distributed computing groups.

MLPerf is a standard machine learning performance benchmark that accurately shows how platforms and models perform in real-world settings. MLPerf Training v3.1 just landed and it showcases amazing H100 performance in absolute terms, but when comparing wall clock performance per dollar spent, Intel still holds its own.

QA Wolf gets web apps to 80% automated test coverage in just 4 months. And working with them is more cost-effective than growing your QA team.

After giving QA Wolf a product tour and access to an environment, their QA engineers will build you a comprehensive end-to-end test suite in Playwright. There's no vendor lock-in and it's easy to export your tests. QA Wolf also provides 24-hour maintenance and triage. You receive human-verified bug reports— no false positives. Plus, you get unlimited parallel test runs on their infra at no additional cost. Schedule a demo to learn more.

Researchers developed a smarter method for self-driving cars to recognize roads by using lidar technology, which requires less manual work but still keeps the system accurate.

LongQLoRA is a new tool that lets LLMs understand longer sequences.

Researchers discovered that large language models can be powerful tools for visual tasks even though they were designed for text. The researchers achieved improved results in various visual tasks by using parts of these text-trained models to process images and videos directly.

Nvidia became the AI kingmaker in 2023. Thousands of companies and countries begged to purchase its GPUs. Will the insatiable demand persist? This piece explores the Nvidia run-up, the resulting shortage, and the future of GPU demand.

OpenAI wants to build partnerships for open and private dataset creation for AI training. This could signal it is running out of tokens for its next model training runs.

Current software is still limited in its ability to understand and assist with diverse tasks. Bill Gates predicts that AI agents will revolutionize software interaction, offering personalized assistance across various tasks without needing specific apps, marking the biggest shift in computing since the move from command-line to graphical interfaces.

Browse the web with GPT-4V and Vimium.

Supercharge your growth with AI-powered cold emails.

Lindy introduces the first platform for building teams of AI employees to perform any task without requiring coding skills. Users can create these 'Lindies' in 30 seconds and set them up in AI teams that can support advanced workflows. Lindy enlarges on-scale productivity with these 24/7 AI employees.