🪄 New prompt strategy beats them all!

AI-HOI AInauts,

Welcome to the latest issue of your favorite newsletter. Highlights include nothing less than two new and extremely good AI tools (including prompt tricks) that you can use immediately. And a few news items to keep you up to date with everything important.

That's what we've got for you today:

• 🌀 Text-to-video for free? MiniMax shows you how - and it's EPIC!

• 🔥 Reflection 70B thinks for itself

• 📰 AI news quickie: the HAI highlights from the industry

Here we go!

🌀 Text-to-video for free? MiniMax shows how it's done - and it's EPIC!

And yet another new text-to-video model! The Chinese AI company MiniMax is currently making waves. Why? Because their new model can generate damn good videos.

You can even use it for free without registering if you know the direct link. Here it is: https://hailuoai.com/video

Whether futuristic cities, ninja fights in a bamboo forest or beautiful nature shots from a bird's eye view, the tool spits out hyper realistic scenes - and all that without a Hollywood budget!

A few tips for use

If the page is displayed in Chinese, right-click in the browser and select "Translate to English" in the context menu (or simply enter your prompt and click on the 😎button).

English prompts are understood. Take inspiration from the existing prompts on the site, or use our new AI video prompt generator here.

It takes a few minutes to create the video. The output is 6 seconds long and in a solid resolution of 1280x720 pixels at 25 frames per second. With a little effort, you can do some really, really cool things!

Our take: This thing makes you want more!

As with every new AI tool, this is the worst version we will ever have. And the results for short clips are already extremely strong.

Longer video lengths and more creative control options are still missing - but are already in the works (e.g. picture-to-video, as we know it from the competitors Runway, KLING, LumaLabs and Co.).

We'll definitely be hearing more about MiniMax - use it now and take advantage of the free honeymoon period.

🔥 Reflection 70B thinks for itself

In the last few days, the "new" Reflection 70B model on X has also gone viral. It is coined as the most powerful open source model. At least if you believe the makers.

We (naturally) wanted to find out more and did some research:

• It's not a "new" model, but was upcycled based on Meta's Llama 3.1-70B.

• It was developed by prompt magician Matt Shumer, the creator of HyperWrite, and Sahil Chaudhary of Glaive, a synthetic data startup (in which Matt has invested).

• Since it was trained with synthetic data, there's no privacy issue and no copyright drama! It's ethically clean and offers a solution to data scarcity in sensitive areas.

• The exciting thing is that it apparently achieves better results thanks to an innovative reflection tuning technique. More on this in a moment...

• In various benchmarks, it is almost on a par with the larger models from other providers (yes, we know, don't trust any benchmark that you haven't tuned yourself...).

All this has been achieved without a large team or massive computing power. The two of them simply conjured up one of the world's best models as a side project!

Clever prompt engineering reduces hallucinations

There is one important point that stands out in this project: the importance of the system prompt.

The model only works optimally with the proposed system prompt. Without this special configuration, it seems to fall back to the level of the basic version of Llama 3.1.

The prompt that leads to the extraordinary results is super simple:

You are a world-class AI system, capable of complex reasoning and reflection. Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags. If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags.

Reflection 70B is therefore so good because it goes through a chain of thought during the conclusion and reflects!

This technique gives it a new opportunity to pause, check itself and set things right. It integrates special commands such as <thinking>, <reflection>, and <output>, which enable a more structured and precise output.

The method is based on an "older" paper on the reflection tuning technique, and is a further development of the chain-of-thought methods.

This approach reduces one of the biggest flaws of language models - hallucinations. Normally, chatbots just plow on regardless of whether they are wrong or not. This shows once again that you can massively improve the results with the right prompt.

And you can use this to your advantage by simply including the system prompt in your favorite model or adding it to the custom instructions. Or use the specialized Reflection GPT.

Here is an example with the popular 🍓("How many R's does the word Strawberry have?").

Our take: Impressive performance, but …

Reflection 70B seems like a significant step towards democratized AI thanks to open source - thanks, Meta! You can find it on Hugging Face.

Over the past days, there was some controversy happening. We are curious to see how the model will really prove itself in practice.

In any case, the example shows that remarkable results can be achieved with the right techniques. And you now have another tool at hand without having to rely on big tech.

📰 AI news quickie: The HAI highlights from the industry

Finally, a few tidbits - here are the most exciting news. But first: check out this really well-made AI clip to get you in the mood!

Industry

• After celebrating the AI development environment Cursor in last week's issue, Replit has now introduced Replit Agents. Not only can you use it to create an app, you can also deploy it directly! Or as one user puts it: "I had the idea for 15 years and realized it in 15 minutes!" Game changer, more to come.

• TIME magazine has once again selected the 100 most important people in the AI industry this year - here is the complete list. However, the online community sees it a little differently, namely as follows...

• Elon Musk didn't make the list (for whatever reason) - even though he has just brought the world's largest AI supercluster online with Colossus...

• Coinbase enables AI-to-AI crypto transactions for the first time - a milestone for autonomous AI agents in the decentralized financial world.

• Musician arrested: He swindled 10 million dollars from streaming services with AI-generated songs and bots!

• ChatGPT reports over 200 million weekly users (and a total of 1 million paying business users), while Meta AI announces 400 million monthly users and 185 million weekly users.

Anthropic

• Anthropic has launched Claude for Enterprise - with 500k context windows (almost 4x more than ChatGPT Enterprise!), native GitHub integration and full control over your company data.

• And you can now experiment with AI at lightning speed - Anthropic releases starter kits for popular programming languages.

Funding

• OpenAI co-founder Ilya Sutskever has raised one billion dollars (at a 5 billion valuation!) for his new start-up, SSI - Safe Superintelligence. The goal: a safe superintelligent A(G)I. The capital is invested in computing power and talent; the team currently consists of just 10 people.

• Canva drastically increases the prices for Teams subscriptions by 300% and justifies this with new AI functions. This should probably bring some cash into the coffers before the IPO, after a lot has been invested (acquisition of Leonardo AI, Affinity, ...).

• Sakana AI, the Tokyo-based startup behind the autonomous AI researcher, has received a 100 million dollar cash injection from NVIDIA, among others.

Research

• Loopy shows how you can create a realistic video with lip-sync output using just one image and audio - currently only as a research demo ...

• DeepMind has developed AlphaProteo, an AI that develops novel proteins and thus revolutionizes drug development and disease understanding.

• A robot controlled by a fungus? Yes, that's possible!

• Scientists have developed a Minecraft civilization with 1000 autonomous AI agents that communicate and cooperate with each other to solve complex tasks. Fascinating!

🎧 Podcasts: Some food for thought

• No Priors Ep. 80 | With Andrej Karpathy from OpenAI and Tesla (youtube.com)

• Humanoid Robots, the Job Market & Mass Automation - The Current State of AI w/ Emad Mostaque | EP114 (youtube.com)

• Stanford ECON295/CS323 I 2024 I Business of AI, Reid Hoffman (youtube.com)

• Richard Sutton - Humanity Never Had Control in the First Place (Worthy Successor Series, Episode 2) (youtube.com)

That's it! Thank you for joining us.

Reto & Fabian from the AInauts

P.S.: Follow us on social media - that motivates us to keep going 😁!
Twitter, LinkedIn, Facebook, Insta, YouTube, TikTok

Your feedback is essential for us. We read EVERY comment and feedback, just respond to this email. Tell us what was (not) good and what is interesting for YOU.