🧠 Smart prompting: how to master reasoning AIs

Hello AInauts,

Welcome to the latest issue of your favorite newsletter! After an update went live over the weekend that makes ChatGPT-4o smarter and more personalized, Grok-3 - the "smartest AI on earth" according to Elon - will arrive early on tonight. We'll be surprised...

This is what we have in store for today:

• 🧠 Smart prompting: how to master reasoning AIs

• ❤️ It's getting really weird - AI boyfriends & 3D girlfriends

• 🍉 AI-Fun: How to make watermelon robots!

Here we go, should be another exciting week.

🧠 Smart prompting: how to master reasoning AIs

OpenAI o3, DeepSeek R1, Gemini Thinking ... everyone is talking about the new reasoning models. But to exploit their full potential, you also need to know how to control them optimally.

After all, it would be a shame if we were to use the latest AIs with the handbrake on.

So let's take a look at how you can optimize your prompts, what the differences are between the old transformers and new reasoners, what best practices there are and what pitfalls you should avoid.

Classic transformers vs. new reasoning models

Classic Transformer GPT models are versatile generalists - fast, knowledgeable, but sometimes a little superficial. The new reasoners are the specialists who are great at logical analysis, but sometimes overshoot the mark.

The most important differences briefly explained

• The new models are designed to automatically think through complex tasks step by step. The GPT models, on the other hand, shoot an answer straight from the hip without having to think about it for long.

• Reasoners are calibrated for the complex and have a tendency to "overthink", even when it comes to banalities. It's a bit like Einstein being asked what color the sky is. He says blue, but not without first explaining the entire light spectrum.

• The reasoning models score points for context length and input size: o3-mini has a large memory and can accept up to 200k tokens of context (for comparison: GPT-4 was at 8k or 32k tokens for a long time).

• Another difference is the general knowledge. Reasoners are more specialized and excel in their disciplines (e.g. STEM: Science, Technology, Engineering, Math), but also have gaps in their knowledge outside their area of expertise.

We have to adapt our prompting style accordingly to get the best results. Let's take a look at the best way to do this.

Best practices: How to prompt correctly (and what to avoid)!

OpenAI has published a "Best Practices Guide" on how to best prompt the o-model family - and Microsoft's new research also shows that less really is more when it comes to reasoning models. Instead of the previous prompt hacks like "think step by step", these models need clear, minimalist prompts with precise context.

Important takeaways for better prompts with reasoning models:

• ✔ Succinct, precise prompts: Say exactly what you want, including constraints.

• ✔ Facts & context: They don't know everything - so give them what they need*.

• ✔ Defined format: Specify what the answer should look like.

• ✔ Examples only if required: Generally write prompts without examples.

• ✔ Clear prompt structure: Use Markdown, XML tags and headings.

This means: no overloaded prompts, no unnecessary examples, no step-by-step thought process - instead, simply clear inputs for optimal outputs. Sounds almost like "Zen and the art of prompting". 😄

*You can now also upload images & documents for the o1 and o3 models.

Quick guide for prompting

Here is a prompt template that you can adapt for your purposes. If you want to learn more about prompting with reasoning models, you can find a free training course here.

You are [role/expert].  
- [How the model responds in the right context]

Context/Background:  
- [List relevant facts, data, etc. in bullet points]  
- [Only the essentials, no superfluous information]  

Question/task:  
- [Clear question or precise task]  
- [Formulate clearly, e.g. "Assess risk X"]  

Desired response/output format:  
- [What should the answer look like? List, JSON, bullets etc.]  
- [Specify the desired length/depth, e.g. "short & concise"]  

Optional framework conditions/restrictions:  
- [Answer only on the basis of the above facts, do not invent content]  
- [In case of ambiguity: "Name your assumptions"]  
- [Other rules or specifications, e.g. style, language, etc.]

OpenAI's o3-mini becomes more transparent - but remains a gray box

Until now, OpenAI's models were known as a "black box" and were notorious for their lack of transparency. Recently, however, OpenAI has updated its new o3-mini model and now allows a better insight into its thought process.

Why has OpenAI suddenly become more transparent? Quite simply: competitive pressure! DeepSeek, Perplexity and other AI players directly disclose how their models think - and users love to watch the intelligence think.

That's why OpenAI has tried to get out of the affair with a partial opening. While we now get a more detailed "chain of thought", part of the decision-making process still remains under wraps.

P.S. In the future, all models are to be combined in one model and enable a better user experience.

❤️ It's getting really weird - AI boyfriends & 3D girlfriends

Although we had the phenomenon of AI boyfriends and girlfriends on our radar, we hadn't yet looked into it in depth. A few weeks ago, this article in the New York Times, which is well worth reading, led to a WTF moment for us.

Millions of people are already chatting, flirting and dating virtual partners - they're just a prompt away...

Welcome to the era in which swipe right also applies to chatbots! There are plenty of platforms, from completely vanilla to completely uncensored. And apparently it can even get a bit spicier with the ChatGPT Advanced Voice Mode.

Virtual romance at the touch of a button

While AI friendships have long been part of everyday life in Japan and South Korea, the rest of the world is only just beginning to understand the phenomenon. But what happens when an algorithm can respond more intimately to your wishes and needs than a real person? Or when ChatGPT knows how to deal with your heartache better than your best friends?

Take a pinch of loneliness, add a clever chatbot, and voilà: the virtual friend or partner is born. Apps such as Replika, Character.AI, DreamGF, CrushOn.AI and Couple.me are falling on fertile digital ground and enticing users with slogans such as "Your perfect partner, available 24/7!".

There's still a lot to come socially, and the movie Her, which is well worth seeing, gives us a foretaste. Whether we like this development or not is irrelevant - the trend is clear: 📈

User numbers are shooting through the roof, according to Pareto: around 80% of users are male, 20% female, and the average user is 27 years old.

Sex sells. And so does emotional support.

Wherever emotions are involved, money flows - that was already the case with dating agencies and is no different here. Market researchers are outdoing each other with rosy forecasts, and the tech industry senses a billion-dollar business here.

Many of the AI partner apps work with freemium models: Basic functions are free, but you have to pay for "more" (i.e. more explicit conversations, voice messages, individual avatars, etc.).

Influencer Caryn Marjorie has 2.6 million followers on Snapchat. With CarynAI, she has created an AI clone of herself that flirts with her fans as a virtual girlfriend. The highlight: chatting with Caryn via voice output, for the price of 1 US dollar per minute. Absurd? Perhaps, but extremely lucrative - the 24-year-old is now a millionaire.

Users' willingness to pay apparently increases as soon as they feel an emotional connection: what's a few euros a month for a dream partner?

The popular platform Replika offers a subscription to unlock romantic role-playing games and augmented reality features - and hundreds of thousands are willing to dig into their pockets for it.

In close contact - chat, image, audio, virtual reality, ...

The combination of generative image AI and chatbot even makes it possible to create intimate media, which then delivers whatever your heart (or eyes) desires to match the chat experience. In other apps, you can put on VR glasses and sit opposite your virtual sweetheart.

And why not simply 3D print your AI partner? Find out more here.

In Japan, a device called Gatebox was developed back in 2016 that projects anime holograms as "spouses", including a daily greeting in the hallway. In 2019, a Japanese man married a virtual singer to great media effect - via hologram, complete with ceremony and ring.

The holographic wife helped her husband overcome a deep depression and find new courage to face life. And according to Instagram, they are still happy, as this photo from last week shows.

By the way, a study has just announced that ChatGPT performs better in couples therapy than real therapists.

Our take: virtual companions are here to stay!

Let's face it: AI boyfriends and virtual girlfriends, as well as other companions are far more than just hype. They satisfy real needs for closeness, control and understanding, and shift the boundary between man and machine.

And why is the whole thing booming? Quite simply because it works! For many people, this is a game changer. The virtual partner is there for you 24/7 and offers emotional support without any drama.

AI companions can be a valuable addition, whether as a training ground for social interactions, as an emotional buffer in difficult times or simply as an interesting experiment with the possibilities of technology.

The flip side of the coin: if you fixate on your AI partner, you might miss out on real life out there. It's like fast food for the soul - tastes good, fills you up in the short term, but something is missing.

The art lies in finding a balance between the best of both worlds - in other words, combining the reliability and playing field of technology with the irreplaceable depth of real human relationships.

🍉 AI-Fun: How to make watermelon robots!

Let's end this issue with a nice example of the possibilities of the new AI video models!

We made it! But no need to be sad. The AInauts will be back soon, with new stuff for you.

Reto & Fabian from the AInauts

P.S.: Follow us on social media - that motivates us to keep going 😁!
X, 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.