- TechTok Newsletter
- Posts
- New Models from NVIDIA 🚀, Video Language Learning 👀, AMD CEO Interview 🎮
New Models from NVIDIA 🚀, Video Language Learning 👀, AMD CEO Interview 🎮
Today's TechToks: NVIDIA's model pipeline for synthetic data generation, Health risk for astronauts on Mars, Uber's personalized marketing at a scale, trending GitHub repositories, product picks, and more!
TechTok by João Martins
June 18, 2024
Today’s Summaries
See all as stories in techtok.today
Today's Picks 🍒
🟢 NVIDIA unveils a family of open models to generate synthetic data for LLM training, and alternate language models to transformers.
👩🚀 Mars missions might cause irreversible kidney damage to astronauts.
🐳 The DenseAV AI that learns language by watching videos without any text input or pre-trained models.
💼 From Uber Engineering: Personalized marketing at scale with an out-of-app system.
💡 Blog post: An in-depth interview with AMD CEO Lisa Su.
GitHub Trending
📄 pocketbase: Open Source realtime backend in 1 file.
🎞️ yt-dlp: a command-line tool for downloading audio/video from various websites, is trending on GitHub.
🐝 StableSwarmUI: A Modular Stable Diffusion Web-User-Interface for image generation.
📺 iptv: publicly available IPTV channels from all over the world.
Product picks
🦄 Unicorns Club: revolutionizes fundraising for founders.
👾 DynaUI: UI library for developers.
🔑 Passkeys: Quickly integrate passkeys into any app.
🗣️Teameet: Video calls with interpretation in the speaker‘s own voice.
🏰 histories: Explore history of places through audio stories & fun facts.
🤯 Game Changers
The most impactful articles of the day
NVIDIA released Nemotron-4 340B: a family of open models to generate synthetic data for training large language models (LLMs). They provide a 3 model pipeline, optimized for NVIDIA NeMo and TensorRT-LLM: base, instruct and reward.
Base is a foundation model for custom LLMs, Instruct creates diverse synthetic data mimicking the characteristics of real-world data, Reward filters for high-quality responses (graded on helpfulness, correctness, coherence, complexity and verbosity). Currently on HuggingFace, and soon to be available as a NVIDIA NIM microservice.
Mamba-2 and Mamba-2-Hybrid were also released — selective state-space models, alternate to transformers — which exceeded a transformer on 12 standard tasks, and closely matched it on 23 long-context tasks (Hybrid).
✨ "Too long don't README"
GitHub trending repos of today
pocketbase/pocketbase
PocketBase is an open-source backend written in Go that provides an embedded database with real-time subscriptions, file and user management, an admin dashboard UI, and a REST-ish API. SDK clients are available in JavaScript and Dart. It can be used as a standalone app or as a Go framework/toolkit for building custom business logic, and is currently under active development.
What you can do with it: Build custom backend solutions, and extend functionalities using Javascript or Dart via the SDK clients.
yt-dlp is a feature-rich command-line tool for downloading audio/video from various websites, being a fork of 'youtube-dl'.
What you can do with it: custom download media files from thousands of supported websites, and embed it in your own project.
A collection of publicly available IPTV (Internet Protocol television) channels from around the world grouped by category and language.
What you can do with it: Use the provided links in any video player that supports live streaming to watch TV around the globe.
StableSwarmUI is a web-user-interface for stable diffusion image generation in beta-version, designed for easy access to powerful tools with excellent performance and extensibility capabilities. It allows users to generate various things using an easy-to-use Generate tab and a Comfy Workflow tab for more complex tasks.
What you can do with it: try it on Google Colab or Runpod, generate images with a variety of features like image editor and auto-workflow-generation.
📱 Product Picks
Curated products from Product Hunt
Develop animated websites in hours, not weeks. DynaUi contains 50+ modern and beautiful components, sections, and templates, all fully customizable, that you can copy and paste into any of your projects!
Quickly integrate passkeys into any app. Comes with a Javascript SDK, a passkey provider for Next-Auth, and guides for other frameworks. Self-host or use Hanko Cloud hosted Passkey API.
Teameet is an AI-powered video meeting platform, now offering the new feature called Speech Translation, which interprets meeting participants' speeches in real-time, preserving their tone, pitch, and emotion for seamless cross-language communication.
Dive into the heart of every place with stories. Histories brings locations to life, narrating their unique past. Explore ancient ruins, bustling cities, and tranquil spots hands-free. Enhance your travel with captivating facts, making each journey memorable.
Revolutionizes the approach to fundraising by moving away from the exhausting hunt for investors. Instead, we empower founders to focus on what really matters: developing their products and sales.
🧐 Daily Picks
Curated picks and most shared articles on techtok.today
A UCL-led study warns that astronauts' kidneys could suffer irreversible damage on a roundtrip to Mars due to radiation. The in-depth study, which included over 40 institutions across five continents, found that space flight 'remodels' kidneys and is a primary cause of kidney stone development in astronauts. Particularly concerning is the permanent kidney damage found in mice exposed to radiation simulating Galactic Cosmic Radiation for the span of a potential Mars mission.
AMD CEO Lisa Su discusses her career path, the turnaround of AMD, and the company's strategy in competing with Nvidia in the AI and GPU markets. Key points include: Su's experience at IBM and lessons learned, how AMD leveraged its CPU and GPU capabilities to diversify beyond PCs, the importance of customization and partnerships in the console and hyperscaler markets, and AMD's approach to the rapidly evolving AI landscape where it sees opportunities to differentiate through open and modular solutions.
DenseAV learns the meaning of words and the location of sounds (visual grounding) without supervision or text. Trained on AudioSet (wich has 2 million YouTube videos) and avoiding pre-trained language models, it searches for pixels in a video based on the words spoken — i.e. When a person said 'dog', it searched for dogs pixels in the video. Or when given a two-sided brain, hearing words 'dog bark', half of it focused on language, like the word “dog,” and the other side focused on sounds like barking. Initially conceived to help decode animal speech, such as whale communication, it will be presented at the IEEE/CVF Computer Vision and Pattern Recognition Conference this month.
⚙️ Engineering Blogs
Articles from engineering blogs of big tech companies
Uber has developed an out-of-app communication system that can target billions of users with personalised messages, despite significant challenges such as lack of user context and system costs. The system employs a candidate retrieval process to identify potential recommendations, a filtering and blending process to weed out less relevant options, and a ranking stage to determine the order of presentation.
