Welcome, humans.

Sakana AI dropped a new image model that lets you generate images in the traditional Japanese style of ukiyo-e (try it here). Why’s this special? You can now general ukiyo-e style images of cats!! 

Here’s a little cat joke for ya: 

Why did the cats decide to have a tea ceremony? 

They wanted to discuss the latest mews. 

Here’s what you need to know about AI today:

• Elon's xAI fired up the world's largest AI supercluster

• Condé Nast sent a cease-and-desist to Perplexity over content infringement.

• Meta’s open-source Llama 3.1 405B leaked yesterday, and it might be better than GPT 4o.

• A study found that websites are increasingly blocking AI bots from scraping data.

Elon Musk’s team at xAI just powered on the “World's Most Powerful AI Training Cluster.”

And of course it happened at ~4:20am. It’s Elon. 

Elon’s AI company, xAI, began training on its massive new Memphis Supercluster Monday morning. The goal? Create the “world's most powerful AI” by December 2024. 

If you don’t know what a supercluster is, it’s basically a massive network of Nvidia GPUs (computer chips) working together as a single unit to solve “super” complex calculations at unprecedented speeds. 

And this Memphis Supercluster is the most “super” supercluster we’ve ever seen. The new facility, dubbed the “Gigafactory of Compute”, is a beast:

• 100,000 liquid-cooled Nvidia H100 GPUs on a single RDMA fabric (for context, Google snagged only 50,000 H100 GPUs last year).

• Up to 150 megawatts of electricity usage per hour—enough for 100K homes.

• At least one million gallons of water per day to keep cool! 

As some shared on X, this likely exceeds the capabilities of current top supercomputers like Frontier and Aurora by more than two times. The news also inspired some debates over who in tech had the higher GPU counts. What is this, grade school?  

So is bigger always better? Big tech thinks so. Their logic is: better AI models = more compute = bigger supercomputers = bigger superclusters. 

And with bigger superclusters, you also create better models faster. Case in point: a system like xAI’s could theoretically train GPT-4 in a week, while some estimate the real model took three months to train

That’s why, with the Memphis Supercluster now online, Elon Musk thinks Grok 3 will become “the most powerful AI by every metric” by EOY, despite the fact that Grok 2 hasn’t even been released yet (set to launch next month). 

What to expect: Better models, more frequently. That’s been the trend, at least—look at how the last few model releases have become more squished together. 

Meta is planning to release its new Llama 3.1 model today (more below), and OpenAI is rumored to counter with a new model of its own soon.

FROM OUR PARTNERS

Improve your AI models with more accurate, representative human feedback.

At The Neuron, we often discuss how AI researchers are running low on high-quality, real human data for training models.

That’s why organizations like Google and Stanford are turning to Prolific.

Prolific has over 200,00 active participants and domain experts who provide real, reliable data for RLHF and model evaluation.

Simply use their 300+ filters to select the specific demographic you need data from. Diversify and enrich your AI models with data that reflects the breadth of humanity—get started on Prolific for free here.

Around the Horn.

• New stats for Meta’s new Llama 3.1 405B model were potentially leaked, and if they’re correct, it might be better than ChatGPT-4o and Claude 3.5 Sonnet…and it’s open-source for anyone to use.

• Condé Nast sent a cease-and-desist to Perplexity telling the AI search engine to stop including content from its publications in its results, 1 month after Forbes did the same.

• A new study discovered that AI models are increasingly being blocked from accessing training data on major websites, which could impact future model quality. 

Treats To Try.

1. *Incogni removes your personal data from the open internet so scammers and identity thieves can’t access it. Protect yourself online with Incogni—get 55% off with code NEURON. 

2. 54 AI tools to 10x your productivity across research, copywriting, SEO, marketing, presentation, audio, and image + logo generation.

3. Luma Dream Machine now lets you infinitely loop videos for easy AI GIF creation (some examples, try it here). 

4. Spark helps you schedule meetings like an executive assistant—readers of The Neuron get private early access, so sign up today!

5. DigitalFirst.AI recommends tailored marketing strategies and tactics to businesses based on the company’s specific needs and goals (raised 3.5M Euros). 

6. Cohere builds generative AI models specifically for enterprises that are really good at retrieval (raised $500M at a $5.5B valuation).  

7. Airy is an AI reflection journal that analyzes your mood while Gobi offers personalized, real-time wellness exercises and practices.  

8. Vozo re-dubs, rewrites, and lip-syncs viral videos with your own voice via a few simple prompts. 

*This is sponsored content. Advertise in The Neuron here.

Tuesday Ticker

Here are the results from last week’s poll:

• D.G. chose Replace custom footage: “The videos I currently make are made of screenshots, graphics, and occasionally B-roll. I don't have actual footage because it requires stuff I don't have like budget, actors, lighting, cameras & talent! An AI could solve this.”

• J.S. chose Replace generic stock: “Quality Stock footage can be expensive and boring. The last thing I want to see is my stock content show up in someone else's content.” 

• P.G. chose Hollywood level production: “Personalized movies! You can design the plot, characters, and settings, and get it available almost instantly!” 

A Cat's Commentary.