Gm. It’s my first day in Nice, and dang is it hot. I don’t know what we have to do to fix this heat wave. Carbon capture? White paint? Block out the sun? I don’t care, I need it done expeditiously.

— Stephen Flanders

BROUGHT TO YOU BY:

Just a week after Anthropic released their newest chatbot Claude 2, we have another addition to the AI family.

This member comes courtesy of Meta: Llama 2, the second generation of their open-source Llama model. As you’d expect, it’s a significant improvement over Llama 1. Makes sense, considering that Llama 2 was trained on 40% more data than Llama 1 (2 trillion tokens for Llama 2) and has double the context length.

The result is performance slightly below the premier closed-source models GPT-4 (OpenAI) and PaLM 2 (Google). That’s pretty impressive for an open-source model.

Couple that with Meta’s AI partnerships with Microsoft (OpenAI’s biggest backer) and Qualcomm (who will bring Llama to phones), and you have something big brewing here.

But the really interesting thing about this whole launch is that Llama 2 is completely open-source.

We’ve spilled much ink on the threat of open-source AI to the big boys. And open-sourcing Llama 2 is the biggest blow Meta could have dealt to the OpenAI’s and Google’s of the world:

Now we’ll just have to wait and see if people can create appealing open-source alternatives to ChatGPT, DALL-E, and Bard.

The Bottom Line: Nobody knows whether open-source or closed-source AI will win this war, but we do know one thing for sure: Microsoft will profit either way.

OctoAI is the fastest place for developers to run, tune, and scale generative AI models so your apps win over more users.

We have the fastest/cheapest Stable Diffusion 2.1 on the market: 3X faster and 5X less expensive to run in production than alternative endpoint solutions while leveraging widely available hardware so you can infinitely scale your app.

Check out our report card and see the data yourself.

Now, it’s your turn to put our price/performance to the test with Stable Diffusion, Whisper, or any model you bring: We’re handing out 25 free GPU hours for everyone who signs up before the end of the month.

Sign up for OctoAI (now publicly available)!

When GPT-4 was first announced earlier this year, we were promised multimodal capabilities. In English, we were told it would be able to accept text and image inputs.

Image inputs would be absolutely amazing. Imagine uploading a picture of any problem you have and instantly receiving advice on how to fix it. Or an app that describes to blind people their surroundings in real-time.

Both are possible with image inputs.

But, unfortunately, that has not come to fruition yet, and now we know why. OpenAI is worried about what ChatGPT will say about people’s faces.

In other words, there are serious privacy concerns related to GPT-4 being trained on people’s faces. In many regions of the world, you need explicit consent to access biometric information.

Not to mention the fact that people’s faces are a sensitive topic, and it wouldn’t be very fun to hear a chatbot call you ugly or the wrong gender.

It’s clear OpenAI has a lot to work through before unveiling image inputs. Hopefully they figure it out soon, if for no other reason than not having to use Google’s awful image search anymore.

The Big Picture: Multimodal models are the future, but it might be a while before they are legally viable.

What else is going on?

Stat: 76%: How much crypto VC funding has declined year over year. It’s an AI world now.

Rabbit hole: Generative AI Space And The Mental Imagery Of Alien Minds (Stephen Wolfram)

Hope that broom isn’t over 25.

News is so weird these days. Three of these headlines are fake stories from The Onion while one is actually true. Can you spot the truth amongst the fakes?

A. Hundreds Of Golden Retrievers Gather In Highlands

B. Majority Of Bodies Donated To Science Dressed Up Like Clowns

C. Jonah Hill Debuts Inspiring Documentary About His Lawyer

D. Heat Safety Experts Advise Americans To Seek Privilege

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

🥶 Bifrost uses AI to automatically turn your Figma designs into clean React code.

🥛 Milk Road is the best crypto newsletter in the game.

📜 Reforge Artifacts are a library of documents capturing lessons and insights from established founders and operators.

Maybe it is a real sport after all.

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

A.