👉 With this prompt technique, the AI hallucinates less

Hello AInauts,

It's the middle of the night; and we've been experimenting with a new AI tool all day, but did not succeed in our endeavor. But at least we’ve got some learnings to share. More hands-on today, the news will be back on Saturday.

Here's what we have in store for you today:

• 👉 With this prompt technique, the AI hallucinates less

• 👩‍💻 New AI coding app goes viral

• 💻 Build simple AI apps (without programming)

Let's get started!

👉 AI hallucinates less with this prompt

At the beginning of September, we wrote about a very exciting prompt technique that significantly improves the responses of language models (ChatGPT & Co.).

The so-called Reflection Tuning.

Here again the prompt from September, which you can add before any of your requests:

You are a world-class AI system, capable of complex reasoning and reflection. Reason through the request 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.

Even if the model presented at the time didn't quite live up to all the hype, this technique still makes the AI's answers much better. The reason: a chain of thought is reflected on before you receive the answer. As a result, AI hallucinates much less.

The Internet would not be the Internet if this technique would not constantly evolve.

Here is an extended prompt which produces extremely good results, especially with Claude 3.5 Sonnet:

Start by enclosing all thoughts in <thinking> tags and exploring different perspectives and approaches.

Divide the solution into clear steps within <step> tags. Start with a budget of 20 steps and request more for complex problems if necessary.

Use <count> tags after each step to display the remaining budget. Stop when you reach 0.

Continuously adjust your thinking based on interim results and reflections and change your strategy as you progress.

Evaluate progress regularly with <reflection> tags. Be critical and honest about your thought process.

After each reflection, assign a quality score between 0.0 and 1.0 with <reward> tags. Use this to control your approach:

0,8+: Continue current approach
0,5-0,7: Consider minor adjustments
Unter 0,5: Seriously consider going back and trying a different approach

If you are unsure or the reward score is low, go back and try a different approach, explain your decision within <thinking> tags.

If possible, explore several solutions individually and compare the approaches in reflections.

Use thoughts as a notepad, write down all calculations and considerations explicitly.

Summarize the final answer within <answer> tags and give a clear, concise summary.

Conclude with a final reflection on the overall solution, discussing effectiveness, challenges and solutions. Award a final reward score.

Yes, the prompt is quite long. But it can be worthwhile, especially if you have more complex requests and questions for the AI.

You simply start with your question or topic and then add this prompt below it.

The response is located between the <answer> tags.

Have fun trying it out - and as always, please share your experiences with us!

👩‍💻 New AI coding app goes viral

Now let's get into the topic that led us down a rabbit hole and cost us hours - with unsatisfactory results...

AI can provide great support when it comes to coding. We are particularly interested in coding apps that allow non-technical people (like us) to build program apps.

The Cursor app has been the leader of the pack for a few weeks now. We've already covered it, and the guys behind it were invited on Lex Fridman's podcast recently!

But … 15 minutes in we were lost and didn't understand a thing anymore. Very technical.

And that's exactly the issue. Cursor is brilliant for people who can code already and know how to bring a web app live.

People like us like it evene a little bit easier - like with the Replit AI Assistant, which we have already covered (and we still use it often).

And now a new app is going viral: Bolt.new

Bolt allows you to program entire apps based on simple prompts and publish them directly. It builds all the stuff you need for this, and you can always see and test your current version.

In our case, we wanted to build a small page to create images from a text prompt using the Replicate API.

You simply chat with Bolt, and it builds, adapts and corrects itself.

It also tells you exactly what to do - e.g. how to connect the Replicate API.

It's really cool, and after a few minutes we had a simple interface connected to Replicate.

But then it got tricky, and we lost hours ...

Errors kept popping up, which we patiently fixed together with the AI.

And then it was time to publish. That's one of the big advantages of Bolt, because you can publish your projects directly via Netlify. At least in theory.

But unfortunately, we failed and have not yet been able to solve it.

Of course, the app is brand new and still in beta. We'll keep working on it and as soon as we've got it right, we'll post another update.

But there are others who have managed it, and that's why we're confident.

Nevertheless, it's worth to play around with it. It's really impressive how the system already works (and currently, it is also free to use).

💻 Building simple AI apps (without programming)

The topic from above shows once again why we love working with other tools like Zapier.com.

It's a bit more limited, but we built our little AI image generator app here in about 10 minutes. And it really works.

A super simple app:

• Enter a prompt

• The image is generated via Replicate

• The image is uploaded to Google Drive

• And the data table in the app is being updated

The whole thing is really simple to set up with Zapier, and you can use it to build all kinds of mini-apps. On the Zapier website you can find a whole library of ready-made templates.

Just like this one for the AI image generator. This template automatically creates the interface for you, which you can edit as you wish.

There is also a table into which all your data is being saved.

And an automation that already contains the necessary steps:

We simply connected the new Flux 1.1 Pro image model from Replicate via a webhook. And that's it!

Btw, Zapier now also offers the option of accepting payments via Stripe. This makes it even easier to build small apps and even take money for them.

For example, you could create a very cool LoRA, address it as an image model and people have to pay a few euros if they want to create images with it.

Sure, Zapier is far from perfect and a bit costly. But it's really very easy to build great automations with it.

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

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.