TOGETHER WITH

Good morning, human brains. Welcome back to your daily munch of AI news.

Here’s what’s on the menu today:

• Databricks is valued at $43 billion 🧱 🤑

  This follows its latest $500 million funding round.

• AI models can be trained with fake data 🤖 🧠 

  Synthetic data makes it possible to expand any training data set for pennies on the dollar.

• Google’s image-to-loop method 🤓 📸

  Generative Image Dynamics creates seamless loops from images.

 Sponsor Bot Eat Brain | New here? Subscribe!

APPETIZER

Databricks is valued at $43 billion 🧱 🤑

Databricks’ latest funding round brought in $500 million. This put its valuation at $43 billion.

Wowza.

Nickels and dimes:

1/ The investors include NVIDIA, Andreessen Horowitz, T. Rowe Price, Fidelity, Capital One Ventures, and more.

2/ The funding is viewed as a “refresh“ rather than a “recharge.“ It will enable Databricks to make strategic moves in the competitive AI market.

3/ It’s not rushing towards an IPO, opting for more growth before going public.

Our take: Databricks has been busy lately. It acquired MosaicML in July, formed a new partnership with IBM on August 31, and expanded its partnership with Salesforce on September 11.

More on the IBM partnership, here.

More on the Databricks expansion, here.

BUZZWORD OF THE DAY

FROM OUR PARTNERS

Train powerful AI models using synthetic data.

They say you are what you eat, and that’s even more true for an AI than it is for you and me.

To make a powerful and useful AI model, you need to start with great data, and lots of it.

Unfortunately, good data doesn’t come easily, or does it?

Enter synthetic data.

You can save lots of money by training your AI using synthetic data. All you need is a sample of 5,000+ subjects to get started, then use MOSTLY AI to expand your data set as required.

It's privacy-safe. There is no risk of data leaks or leaking sensitive personal information because your data is synthetic.

Easily adjust your training set. You can make your real data bigger, or smaller, or give it a different "flavor" by changing the distribution of your data.

Ready to get started?

Try MOSTLY AI’s free tier and generate up to 100k rows of data per day.

The best part? It’s free.

MAIN COURSE

Google’s image-to-loop method 🤓 📸

Google Research introduced Generative Image Dynamics. It’s a method that transforms still images into interactive, seamless looping videos.

Let’s get loopy.

The deets:

1/ The model learns from motion trajectories extracted from real videos with oscillating motions, like trees swaying or clothes in the wind.

2/ It employs a frequency-coordinated diffusion sampling process to predict per-pixel long-term motion.

huh? Pictures are made of pixels and so are videos. This AI figures out how to move the pixels in an image around to make them into a video.

3/ This tech simulates objects in real pictures and allows you to interact with them, simulating their dynamic response.

4/ The amplitude of motion textures can be minimized or magnified, which offers different animated effects.

5/ The model can also generate slow-motion videos by interpolating the predicted motion textures.

Our take: the interactive element of this could have applications in alternate reality (AR) filters, marketing, and more.

Will they expand to more complex motions besides oscillations? We’ll keep a close eye on this.

MEMES FOR DESSERT

YOUR DAILY MUNCH

Think Pieces

GPT-4 is not getting worse. In June, a notable paper claimed GPT-4’s quality declined, but the author now retracted and clarified this statement.

Is Google’s Gemini model coming sooner than we thought? Some companies received early access to it and it allegedly competes with GPT-4.

Is Japan building its own version of ChatGPT? Yes, it is, here’s why.

Startup News

Microsoft open-sources EvoDiff. It’s a new framework that creates “high-fidelity,“ proteins and cuts out several previously labor-intensive steps.

Pixis raises $85 million in a Series C1 funding round. It’s an AI platform for creating and monitoring marketing campaigns.

Patronus AI launches with $3 million in funding. It’s a startup that has an evaluation/security layer for LLMs with scoring, test generation, and more.

Research

MagiCapture — a method of integrating subject and style concepts to generate high-res images.

PagedAttention — a system that tackles the problems of inefficient memory management in LLMs.

DreamStyler — a framework to generate artistic images via text-to-image and style transfer.

Tools

Kick — an AI-powered bookkeeping tool for startups.

Wave — an IOS app that takes voice recordings and creates brief, concise lists/summaries of meetings, lectures, and more.

Astrocyte — create, share, and chat with AI-powered, 3D avatars.

Portfolio Magic — Contra’s AI-powered portfolio creation tool for freelancers with analytics, payments, and more.

TWEET OF THE DAY

Who coined the term “Hallucination“ in AI? The debate rages on. Jim Fan, NVIDIA AI research scientist, pulls up a 2000 paper about facial recognition AI called “Hallucinating Faces.“

Tag us on Twitter @BotEatBrain for a chance to be featured here tomorrow.

RECOMMENDED READING

If you like Bot Eat Brain, you might like these other newsletters too:

🚢 Semafor Flagship — The daily global news briefing you can trust. Get the biggest stories across the globe summarized for you each morning in 100 words or less. Read by 100k+ intellectually curious subscribers.

AI ART-SHOW

Until next time 🤖😋🧠