🦾 Better than GPT-4? Claude 3 raises the bar

Good morning, AInauten!

Welcome to the new AI newsletter with the most important news, tools & hacks on the topic of artificial intelligence. Today, we have a CLAUDE special (... but if you're not interested in the best of all chatbots, you can of course skip straight to the news section 😎).

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

• 🦾 Better than GPT-4? Claude 3 raises the bar

• 🧑‍💻 How to use AI to make your prompts perfect

• 📰 AI-News-Quickie: The HAI-lights

Let's go!

🦾 Better than GPT-4? Claude 3 raises the bar, and we are thrilled!

We have to start with the most important news of the week: The Anthropic team (founded by former OpenAI employees) released version 3 of their language model Claude this week!

The eye-catcher: It is supposed to be better than GPT 4 from OpenAI!

Of course, it hit like a bomb, and we have also spent a lot of time testing it over the last few days.

As always, however, we don't just want to parrot the news, but also point out a few things that we have noticed in practice and that could be valuable for you.

First, the most important things in brief:

• According to expert estimates, Claude 3 was trained with around 40 trillion tokens. (That's really, really a lot of content ...)

• There are three models of different sizes. Accordingly, they are suitable for different tasks and also cost more or less.

• Haiku is the most compact (~20B), followed by Sonnet (~70B) and the flagship Opus (~2T).

• Sonnet is available in the free chat version at claude.ai, and the large Opus model costs 20 USD a month.

• Claude 3 can memorize a lot (between 200,000 - 1,000,000 tokens)!

• It was most likely trained primarily based on AI-generated content from the predecessor Claude 2 and GPT-4.

• And it is now also multimodal, i.e. it can understand images - but not generate them!

• Also the API is available, therefore exciting for developers.

• And Anthropic even shows us the system prompt behind it.

• Claude is not yet officially available in Germany.

• But we'll show you how you can get around this and use Claude in all its variants for free!

So, enough of the facts: Let's take a look at a few points that we find exciting.

Is Claude 3 really better than GPT-4?

According to Anthropic, Claude 3 ranks better than GPT-4 in the largest Model Opus in most benchmarks.

No question, it's a great model and we love it.

But you should know this: The comparison is based on the original GPT-4 release from a year ago! In comparison with the most up-to-date model of OpenAI, GPT-4 Turbo, Claude is no longer ahead.

But hey, we're marketers too and can understand the approach to communication. 😉

The downside is that Claude does not have web access and therefore cannot yet read URLs. The fact that image generation is not possible is not ideal, but not a dealbreaker either.

Image recognition, on the other hand, is great because you can simply upload images from websites etc. and then have them translated, paraphrased, summarized etc. - or even have images and scenery described.

Of course, Anthropic's servers are full to bursting at the moment. Therefore, you will probably often get the following message in the free version. We then simply send our prompt as often as necessary until the system takes over. Works well at the moment.

How you can use Claude in Europe

So, now that we've made your mouth water, let's take a look at how you can get access to it - because unfortunately Claude is officially still not available in Europe.

However, as we have already written about Claude 2, this can be easily avoided with a VPN. In general, we tend to have reservations about free VPNs ("if you don't pay anything, you're the product"). Well-known providers such as NordVPN (and others) have a free 30-day trial period, for example, or a restricted free version.

And currently, once you have created an account on claude.ai with a VPN, you can still access it without a VPN.

In other words, you can get a VPN once, teleport yourself virtually to the USA, create an account on claude.ai and then cancel the VPN during the free trial period.

As already mentioned, Claude.ai runs on the mid-range Sonnet model and is currently still free of charge.

You can even use the Pro version (also free of charge!)

For the largest model Opus, however, it is not yet possible to take out a Claude subscription with Anthropic using a European payment method ... what now?

Don't despair, we have two options for you:

1. Use a third-party provider, such as Perplexity.ai or Poe.com. You can register with Poe.com for $20/month - and also have access to many other top models for text and images (GPT-4, Mistral, DALL-E 3, ...) via a central interface. And since Claude 3 Opus has also been included this week, it's a really good offer.

We even have friends who’ve already canceled their ChatGPT Plus subscription and switched to Poe. However, since ChatGPT's GPTs are indispensable for us, we are currently still dancing at several weddings.

Insider tip: How to use ALL models completely free of charge!

2. Alternatively and completely free of charge, you can also use the Chatbot Arena, even without a VPN. That's pretty cool, because in addition to Claude Sonnet AND Opus, many other models such as GPT-4, Gemini Pro, Mistral etc. are also available there free of charge (!) and can be tested in parallel!

What's the catch? Your chats are not saved and you have no history. In addition, your information is not private, meaning user dialog data, including text and images, could be distributed under a Creative Commons Attribution (CC-BY) or similar license. You can, however, play around with it for non-critical inquiries!

Perhaps it won't be the case that we all become dependent on OpenAI after all - and instead use a meta-app that provides the best user experience for the models.

Initial conclusion and practical examples

We have always been enthusiastic about Claude and version 3 is another mega step forward - everyone should play around with it.

Pietro Schirano has thus, for example, resolved (approximately) the blacked-out elements in the legal dispute between Elon and OpenAI. Clever!

Alex Utopia uses it for Dungeons & Dragons. And Twitter is full of cool use cases and enthusiastic users. It is absolutely fascinating how quickly different models become extremely good.

Practical tip: We have pinned ChatGPT Plus, Google Gemini Advanced and Claude 3 as tabs in the browser and then simply pack our tasks/prompts into all three models in parallel.

Three times the charm! So far, Claude shines in many areas, with a few weaknesses.

Here are three specific examples from our everyday lives.

Example 1: Combine CSV files

We wanted to combine different structured CSV files according to a template.

• Google's Gemini was out of the question because it does not allow file uploads (only images).

• ChatGPT has omitted half of the data (fail!)!

• Claude, on the other hand, diligently and correctly plowed his way through, only to throw in the towel after the fifth "Continue!".

Conclusion: goal not achieved, but Claude was close.

And in general, the extremely large context window (=memory) is also super valuable because it processes the information really well and keeps it in your head. Even if you upload a lot of documents. Most of the time anyway. 😄

Example 2: Finding an English analogy for a German term

In another case, the aim was to find an English-language equivalent for a German acronym, the ABC method. Here, too, Claude had the edge:

Example 3: Extract and translate text from a screenshot

We've been big fans of Claude for a long time. Especially for marketing and advertising, we always get better results with Claude than with ChatGPT.

That hasn't changed with Claude 3, on the contrary: it's still our number 1 choice when we need marketing and creative copy. Creating advertorials is a dream (with the right prompt)!

Here in the example we have a screenshot analyzed and translated - and Claude also shines when it comes to analyzing images (and is said to be even better than GPT-4 and Google Gemini).

Last but not least: Is Claude afraid of death?

Last point: A post by an Anthropic Researcher shows just how powerful these models are. In it, he describes how Opus has realized that it is currently being tested by developers ...

And another researcher wonders whether Claude even has a consciousness and fears death - prompting Elon Musk to ponder "whether we might just be a CSV file on an alien computer".

And some thought leaders are even convinced that we have the basis for AGI here. We'll keep an eye on this ...

🧑‍💻 How to use AI to make your prompts perfect

After today's issue has already become right massively extensive, here is a short topic that is extremely helpful in our daily work with AI.

We often use chatbots to help us with writing prompts. It's super fast and as a result we get much better answers on all kinds of topics.

We noticed that Claude writes extremely good prompts.

Is this because it is a newer model and has already been trained with much more information on prompt engineering? Maybe… Here's an example to try out:

I'm writing a newsletter about AI. You are a world-class AI prompt engineer. Please write me an extremely powerful prompt based on example texts from me that I can use to generate new texts on new topics, texts that make one think I had written them myself.

And here is the prompt result with which you achieve your actual goal - to produce texts that read as if you had written them yourself!

(Of course, this approach works for all types of tasks you have).

Instructions: Analyze the provided sample text thoroughly to deeply understand your unique writing style, tone and characteristics.

Then create a new text on the given topic that mimics your style as accurately as possible in all stylistic dimensions, so that people would think you wrote it yourself.

Sample text: [Insert your sample newsletter text here]

New topic: [Enter the desired topic for the new text]

This is how the new text is created:

1. perform a comprehensive analysis of the sample text and identify key features such as: - tone and voice (e.g. formal, conversational, humorous) - sentence structure and length - use of active vs. passive constructions - typical vocabulary and word choice - paragraph structure and transitions - rhetorical techniques (e.g. analogies, repetition) - personal writing tics or idiosyncrasies

2. create a new text of approx. 300-500 words on the given topic.

3. carefully imitate the writing style by: - matching the tone, voice and level of formality - mirroring the sentence length and complexity distribution - using similar grammatical structures and word choice - adopting the paragraph structure and transition patterns - using the same rhetorical techniques where appropriate - reproducing identified personal writing idiosyncrasies

4. the aim is for the new text to be indistinguishable from your originals. Prioritize seamless imitation of style over accuracy of content.

📰 AI-News-Quickie: The HAI-lights

So, let's get to dessert - a few news snacks including videos for the weekend!

• "This is how I faked my life with AI!" Filmmaker/author Kyle Vorbach was living his dream - or at least that's how he made it look on social media. In his TEDx Talk, he shows why we should question everything.

• This also illustrates well the problem of generative AI: advertising and reality are drifting further and further apart, as this Willy Wonka example shows. The police were called, and one visitor even compared the location to a "meth lab" ...

• Brave brings the free Chatbot Leo from desktop to Android Phone (and in a few weeks also on Apple)! It runs in the Brave Browser and uses Mistral's Mixtral 8x7B model by default, but also supports Anthropic's Claude Instant and Meta's Llama 2 13B. The special thing about it: the chats with Leo are private and are neither recorded nor used for model training. Install Brave here!

• The Pi chatbot is also getting a makeover, and now has a desktop app.

• The bad side of GPTs ... "Welcome to the Era of BadGPTs", surprising deepfake tactics and new AI worms ...

• Microsoft brings the "Copilot for Finance" to to revolutionize spreadsheets with AI. We are excited!

• Content is no longer necessarily "king", but can definitely be monetized well in times of generative AI: WordPress, Tumblr, Reddit, Axel Springer, Stackoverflow etc. turn their users' own content and data into good money.

• Microsoft has visually illustrated 10 AI terms that everyone should know. Worth reading!

• And another video tip: The interview with Google Deepmind CEO Demis Hassabis. For him, one thing is clear: AGI is coming this decade! Here you can find the most important notes on the interview.

• Everyone is talking about Claude. Meanwhile, Qualcomm AI has just released 80 (!) open source AI models, for image, audio and voice applications. Nice one!

• A transparent laptop display from Lenovo. Cool! But: who needs that, especially in times when privacy is becoming increasingly important?

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

See you soon! The AInauten

Follow us on Twitter & LinkedIn!

Your feedback is essential for us. We read EVERY comment and tip. Tell us what was (not) good, and what is interesting for YOU.