🏠 Llama 3

Meta is catching up

Sponsored by

GM! Brett. If you have a startup and are fundraising (or thinking about it) I'd love to try to put you in front of some investors and give you early access to this sweet fundraising tool we've been building. Please fill out this short survey.

🏠 AI

Meta: almost there!

Meta dropped their Llama 3 models a couple of days ago and they're really good.

I found some interesting stuff under the hood.

On the model itself, they've made solid gains with the 8B and 70B parameter versions.

Key things they've done:

  • Way more training data (7x vs Llama 2)

  • Tweaks to model architecture like expanded tokenizer vocab and grouped query attention

  • A blend of training approaches: supervised learning, rejection sampling, PPO, preference optimization

But the bits that really caught my eye were:

  1. Using Llama 2 to filter the training data for Llama 3. Kinda meta.

  2. Detailed scaling laws let them predict the performance of their biggest models before training them.

  3. On 16K GPUs, they're pushing 400 TFLOPS per GPU. That is a LOT of compute.

  4. Instruction tuning / RLHF on reasoning tasks teaches the model how to actually use its own reasoning to pick the best answers. Interesting.

Responsibility-wise, good to see things like Llama Guard 2 for filtering, CyberSecEval 2 for catching vulnerabilities, and Code Shield for blocking sketchy generated code.

The new Meta AI assistant rolling out on Facebook, Insta, WhatsApp etc is powered by Llama 3. It can help with all sorts of tasks like planning trips, generating images, explaining complex topics. Tighter search integration means you don't have to jump between apps.

The image gen in particular sounds slick - faster, higher quality, animations and remixing.

Zooming out, the revelation that Meta's largest Llama 3 models in the works are 400B+ parameters and posting impressive numbers mid-training really underscores the stakes. The technical moat around the "closed-source first" crowd looks to be thinning faster than anticipated.

The Big Picture: Open source models are slowly but steadily catching up to OpenAI.

Fun fact: Today is Sam Altman's birthday and I'm wondering if the below tweet might turn out correct...

What’s the secret to staying ahead of the curve in the world of AI? Information. Luckily, you can join early adopters reading The Rundown– the free newsletter that makes you smarter on AI with just a 5-minute read per day.

🤝 THE LATEST IN…

TECH

AI

SCIENCE

CRYPTO

🏃‍♀️ QUICKIES

Raise: Uniuni, an e-commerce last-mile logistics company, raised $50M in Series C funding to expand delivery coverage across the United States

Stat: $370M. The money, bitcoin evangelist, Michael Saylor has made from MicroStrategy stock sales this year.

Rabbit Hole: The Acceleration of Addictiveness (Paul Graham)

🤩 MONDAY MOTIVATION

“The person who is willing to suffer the longest wins.”

🛠️ FOUNDERS CORNER

The best resources we recently came across that will help you become a better founder, builder, or investor.

📿 Limitless’ wearable AI lets you preserve all your conversations

💵 Narrative helps you standardize invoices and catch errors

🔧 Creo helps you build your internal tools 10x faster

❓️ AI GENERATED OR NOT

Is this AI generated?

Login or Subscribe to participate in polls.

Last week’s answer: If you guessed AI-generated… you were wrong.

💉DOPAMINE HIT

HOW WAS TODAY'S NEWSLETTER?

Login or Subscribe to participate in polls.

REACH 40K+ FOUNDERS, INVESTORS & OPERATORS

If you’re interested in advertising with us, send an email over to [email protected] with the subject “Homescreen Ads”.