TLDR AI 2023-09-11

NVIDIA’s faster LLM inference engine 💨, Imbue raises $200M 💰, deep learning in the browser 💻

Headlines & Launches

NVIDIA launches a faster inference engine for LLMs (6 minute read)

Training costs dominated inference when the current AI race began. Inference costs can quickly balloon as more people deploy language models to production. TensorRT has long been the go-to for speed. A version specifically for language models on H100s is now available.

Government A100 based supercomputer half off through September (2 minute read)

The NERSC is offering a limited-time discount on its supercomputer for GPU-based workloads.

Imbue raises $200M to build AI systems that can reason and code (6 minute read)

Imbue, formerly known as Generally Intelligent, has secured $200M in Series B funding, achieving a valuation of over $1 billion. The funding aims to accelerate the development of AI systems capable of reasoning and coding. Imbue's mission is to produce truly personal computers that amplify user agency, freedom, and dignity, with a vision for AI agents that understand and autonomously act upon user goals.
Research & Innovation

Finding the Basketball in 3D Space (8 minute read)

The authors of this study developed a new way to accurately figure out where a basketball is in 3D space using just one photo, which is a big deal for things like sports analytics and robotics.

Can ChatGPT Help You Find Your Next Favorite Movie or Book? (45 minute read)

This study looks at how well ChatGPT can suggest movies or books you might like.

Carving 3D Clothed Humans using 2D Diffusion Probabilistic Models (3 minute read)

If you use a diffusion model to generate 2D normal maps, which contain distance and shape information, and then use a 3D reconstruction algorithm, you can generate realistic 3D human avatars using an underlying body model. This is useful for synthetic avatars and character creation.
Engineering & Resources

XenonJs: composable AI-powered ecosystem (GitHub Repo)

Create and share easy-to-make, built-to-last, innovative, and customizable experiences by leveraging XenonJs’ rapidly growing ecosystem of components in a no-code environment. Seamlessly add your own components or models to the ecosystem.

Improved Medical Image Segmentation with Segment Anything Model (GitHub Repo)

The Segment Anything Model (SAM) works great for regular images, but not so well for medical ones. SAM-Med2D was created by training SAM on a huge set of medical images with various kinds of input information. It performs much better on medical images.

Run modern deep learning models in the browser with Web AI (GitHub Repo)

As the technology has matured, WebAssembly hype has died down. WebAssembly is now good enough to be used in many production scenarios, including browser-based model deployment.

End-to-end test coverage with zero flakes. For humans, by humans (Sponsor)

Can AI get you to 80% automated test coverage within 4 months? Maybe one day. Until then, there’s QA Wolf. They’ll write your entire test suite in open-source Playwright (you own the code!) AND provide the infra to run it in 100% parallel on every deploy (unlimited test runs included!). Maintenance? Yep, that’s on them too (even full refactors!). Even AI companies like Cohere use them! It’s like magic but it’s QA Wolf

How to stop Meta from using some of your personal data to train generative AI models (5 minute read)

Meta has introduced an opt-out tool allowing Facebook users to delete personal data used in training generative AI models. The "Generative AI Data Subject Rights" form focuses on third-party information, which includes publicly available data or licensed sources. The move comes as data privacy concerns rise, with international agencies urging tech companies to adhere to global data protection and privacy laws.

Product-Led AI from Greylock (11 minute read)

Investors are shifting focus from backing AI enablers like NVIDIA to supporting founders building transformative, AI-first products that redefine work and life. Three major AI-first opportunities include AI-first networks and marketplaces transitioning from human-driven to algorithmic content creation, redefining enterprise software categories where AI becomes the main feature, and AI "co-pilot" augmenting services, especially in areas like wealth management. The potential extends beyond software, tapping into vast service markets.

What OpenAI Really Wants (21 minute read)

OpenAI and CEO Sam Altman aim to develop artificial general intelligence (AGI) safely, but the pursuit of funding and products has shifted the company’s culture. Releasing ChatGPT acclimated the public to AI. Critics say that OpenAI’s commercial focus distracts from mitigating AI risks, but OpenAI insists that its mission remains unchanged.
Quick Links

Rephrase AI (Product)

A text-to-video generation platform that eliminates the complexity of video production, enabling you to create professional-looking videos with a digital avatar in minutes.

ChatGPT Uses A Lot Of Water (3 minute read)

It is estimated that ChatGPT uses up to 500 milliliters of water (about equal to what’s in a 16-ounce bottle) every time it is asked a series of between 5 to 50 prompts or questions.

Microsoft And Paige Are Building The World’s Largest Cancer Detecting AI Model (3 minute read)

Microsoft is teaming up with digital pathology provider Paige to build the world’s largest image-based artificial intelligence model for identifying cancer.
TLDR is a daily newsletter with links and TLDRs of the most interesting stories in startups 🚀, tech 📱, and programming 💻!
Join 500,000 readers for one daily email