TLDR AI 2023-09-29

Google PaLM + Chroma 🤝, AI for SaaS 💻, Meta's text to image model 🖼️

Billionaires wanted it, but 54,578 everyday investors got it first... and profited (Sponsor)

When incredibly valuable assets come up for sale, it's typically the wealthiest that end up taking home an amazing investment. But not always...

One platform is taking on the billionaires at their own game, fractionalizing prized blue-chip artworks for its investors. In just the last few years, its investors have realized annualized net returns of 17.8%, 21.5%, 35% and more.

It's called Masterworks. When Masterworks sells a painting – like the 16 it's already sold – investors reap their portion of the profits.

Offerings can sell out in minutes, but TLDR AI readers can skip the waitlist to join.

Past performance is not indicative of future returns, Investing involves risk. See disclosures

Headlines & Launches

Chroma x Google for building on PaLM (2 minute read)

Chroma is a leading vector database for storing embeddings for AI applications. They've partnered with PaLM to build applications on top of the flagship Google model.

How Is AI Going to Impact The SaaS Landscape Over The Coming Years with a16z GP Kristina Shen (7 minute read)

We all know SaaS is going to be impacted by AI but the question is what is the next wave going to look like? This in-depth discussion with Kristina Shen goes into why it’s more about insights than information and unlocking the real power of AI for SaaS.

New Forethought tool lets you build workflows with natural language (2 minute read)

Forethought unveils "Autoflows", a new AI-driven system for customer service tasks that goes beyond simple Q&A. Using natural language prompts, Autoflows determines steps and tasks across systems, differing from traditional no-code workflows. Powered by SupportGPT, which is based on OpenAI models, the system blends automation and human oversight.
Research & Innovation

Meta's text to image model (6 minute read)

Pre-trained on 1.1B image pairs and fine-tuned on just a few thousand highly curated images, Emu image generation outperforms SDXL in user preference surveys. It acts as the backbone to much of Meta’s new AI assistant play.

Simplifying VQ-VAE (30 minute read)

VectorQuantized-VAEs are typically the state of the art when trying to learn a specific discrete representation (e.g., tokens or codes). However, they are usually complex and fragile. This new paper proposes a simple quantization scheme that eliminates codebook collapse and complicated machinery such as commitment losses, codebook reseeding, code splitting, entropy penalties, etc.

Jointly training large multimodal models (23 minute read)

Models are usually trained separately on specific tasks (e.g., language vs image generation). The newly proposed joint autoregressive mixture (JAM) algorithm combines disparate models with a clever interleaved cross attention and gentle fine-tuning achieves dominant performance in many multimodal tasks.
Engineering & Resources

SigLIP checkpoints released by Google (GitHub Repo)

Joint embedding models combine two data types into a single space. CLIP is a popular one for images and text. Recently, Google researchers proposed a Sigmoid CLIP model which performed very well and was Vision Transformer based. They've now released more information about the models and updated their code and paper.

The Future of Image-Text Understanding and Creation (GitHub Repo)

InternLM-XComposer is a cutting-edge tool that can create articles with both text and images that fit perfectly together.

LAST CHANCE: Save your spot at the most in-depth virtual conference on LLMs in production (Sponsor)

The MLOps Community’s LLMs in Production conference is happening on Tuesday, October 3rd. This is your last chance to hear from industry leaders who are building production AI apps, and enjoy priceless education and networking (for free!)

On the agenda: 40+ expert speakers, musical interludes, guided meditation, prompt injection games, and plenty of surprises. Sign up now to avoid FOMO later


A Leader's Guide to Generative AI in the Workplace (20 minute read)

For technical leaders, embracing generative AI in the workplace brings a wealth of opportunity to drive productivity and transformation. It also means leaders need to consider how to design meaningful career pathways for team members. Workera’s guide shows leaders how they can do it.

White House could force cloud companies to disclose AI customers (4 minute read)

The White House may mandate cloud companies to report large-scale computing purchases, targeting potential foreign AI threats. Critics cite surveillance concerns and changing compute costs. Entities like Microsoft and OpenAI support similar oversight measures.

Anthropic partners with BCG (2 minute read)

Anthropic has partnered with Boston Consulting Group (BCG) to offer their AI assistant, Claude, to BCG customers globally. This collaboration focuses on responsible AI deployment, with use cases including market research, fraud detection, and business analysis. Both entities aim to establish a new standard for ethically deployed enterprise AI.
Quick Links

Google Adds A Switch For Publishers To Opt Out Of Becoming AI Training Data (1 minute read)

Google introduced a tool, Google-Extended, which enables website publishers to prevent their data from being used in Google's AI training while still appearing in Google Search results.

Docera (Product)

An AI-powered tool that simplifies document creation for personal and professional use.

Metaphor (Product)

Metaphor’s API lets you connect your LLMs to the internet, enabling powerful search and research capabilities. In a few lines of code, you can get high-quality search results as well as instant HTML content.
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