In the world of AI, a year's worth of news can happen in a week.

New tools, better models, and plenty of breakthroughs.

Anticipating what's next isn't easy.

Let's take a look at what the future holds. At least for this week. 🙂

Today's Ai5:

• 🖼️ Brand New Midjourney Competitor
• 🎁 Introducing the Ai5 Referral Program
• 📈 Google Makes a Comeback
• 🔧 Reverse Prompt Engineering
• 🤖 Everyone's Building Robots

Prompt of the Day 🎨

Made with FLUX.1.

Brand New Midjourney Competitor 🖼️

It's not often I get excited about a new AI image gen release these days. But this one's different.

The creators of Stable Diffusion have just released a brand new open source image generator, FLUX.1. The team from Black Forest Labs have already secured $31m in funding based on the premiere model, and for good reason.

After putting Flux through a workout, here's what I noticed:

• Single image generations are super fast - about 5 seconds each
• The general model aesthetic is cartoony for anything other than human portraits. It's not bad, just something I noticed.
• Strong prompt adherence
• Model sometimes generates 6 fingers and phantom limbs

Despite what people are saying, I don't think Flux compares to Midjourney or Phoenix yet. BUT, it's definitely up there. And it's definitely the best open sourced model yet.

You can use Flux for free here (you'll need a GitHub account to login).

Follow Black Forest Labs on X here.

Ai5 Referral Program 🎁

Over the last few weeks I've been working hard to put together a referral program for Ai5!

If you enjoy Ai5, refer friends and get rewarded. I think the prizes are pretty awesome too. Here's a last minute infographic I slapped together in Canva. 😁

The prompt pages are pretty awesome. I formatted all the prompts so you can easily copy/paste them. I've been using them myself quite a bit.

How does it work?

From now on, at the bottom of every newsletter you'll see a block where you can get your unique referral link. Copy, paste, and share it to get rewards.

Where should I share it?

Wherever! Post it on socials, send it to colleagues, or tell your nan. 🙂

Google Makes a Comeback 📈

Google's journey in the generative AI race so far has been rocky at best. But after a series of blunders it looks like Google might have gotten it right this time.

Gemini 1.5 Pro quickly rose to the top of the Chatbot Arena Leaderboard over the weekend. Where the outputs of 2 random (unknown) LLMs are voted on by users to see which give the best results by human opinion.

From a technical perspective though, it still lags GPT4o, Claude Sonnet, and Llama 3 for technical tasks like coding and reasoning.

On another note, if you're as confused as me by Google's AI ecosystem, click the image below to get a run down.

Funny but true.

In other Google news, they've just signed a deal to gain access to the non-exclusive rights to Character.ai's language models. At the same time, Google snagged back Character's CEO, who left Google to work on the startup back in 2021.

Reverse Prompt Engineering 🔧

Let's be honest. Most AI tools like Jasper, Copy.ai, and Writesonic are just glorified ChatGPT wrappers. What you're really paying for here is a user-friendly front end, and some well thought out system prompts.

The good news is if you don't want to pay for these tools, you can reverse engineer the prompts quite easily.

Depending how crazy you want to get, you can use the method outlined in this paper. The method involves generating 64 unique output samples then building a custom inversion model that can reverse engineer the system prompt.

It's a bit much for what we're after!

Instead, try out this prompt. 👇

You can include 1 or more examples, the more the better!

Using this method, I reverse engineered this Writesonic system prompt for writing simple blog posts:

It works really well too!

Everyone's Building Robots 🤖

You've probably heard me say this before: The first half of this decade belongs to software based AI while the second half will see major breakthroughs in hardware.

Here's some of the crazy advancements we're seeing lately.

Below shows how a new Reinforcement Learning technique allows for significantly faster learning and better performance in complex control tasks. Check out the paper for heaps of cool examples.

Until recently most robots have controlled movement at their joints with small electric motors or in Boston Dynamics case, hydraulics. These designs are moving toward tendon/muscle style actuation that mimic human anatomy.

A newer example... 👇

Another trend is human controlled robotics via VR. This concept might actually beat fully autonomous deigns to market. Think: robots controlled by humans in a cheap labor country.

Nathie 🔜 Gamescom @NathieVR NVIDIA demonstrated how an Apple Vision Pro can be used to remotely control a robot. Pretty cool! 1:37 AM • Aug 4, 2024 208 Retweets 1491 Likes Read 34 replies

​

Things are starting to get really interesting in the world of robotics!

Snack Sized 5 🍪

1️⃣ Runway Gen-3 Alpha now generates videos 7x faster than before.

2️⃣ The robotics company Figure just released this teaser for their brand new robot Figure 02.

3️⃣ Warren Buffett sold $84.2 billion of Apple stock last Friday, and markets are crashing, except for... Lockheed Martin. (bit concerning)

4️⃣ Lex Fridman and Elon Musk did an 8 hour podcast last week. If you don't have time to watch it, this post summarizes everything they talked about.

5️⃣ I put together a bunch more examples from the new image generator Flux1 here.

Ps. I've started taking this course on developing LLM powered apps with LangChain. Feel free to join me and send a DM if you do!

Join the conversation on X.

Connect on LinkedIn.

Read the Blog on Rareconnections.

Feedback? Yes please! matt@myai5.com