One week with Claude

Whether you like Large Language Models, AI and the whole hype around them or not, I think we can all agree that they are shaping our present and future in one way or another. My view of them was always kind of two-folded. On one hand they use a ton of energy, could concentrate power even more into the hands of a few and open the doors to forms of abuse we don’t seem to fully comprehend today. On the other hand, I have a strong interest in and generally feel drawn towards new technology and the possibilities they provide and it is astonishing how fast the development in this space has been in the last few years.

Up until recently, I played around with the free version of ChatGPT for text, a little with Suno for music generation and MidJourney, Adobe FireFly and even some locally installed StableDiffusion for image generation a few months ago. I never used any of those productively, though, and it was all more of playful experimentation so far.

Watching this interview by Lex Fridman with Dario Amodei (the CEO of Anthropic, creators of Claude.ai) and listening to him talking about his views on AI, their different models and also their views on ethics, safety, and regulation surrounding these topics, made me think I should give Claude another go.

Because my free limit was reached pretty fast, I went ahead and ordered the smallest $20/month subscription which I’ve been using for about a week now. Here’s my impression so far.

I am still experimenting and finding new things to try every day, but here’s a list of the things (in no particular order) I found it being useful for, so far.

• Explaining Programming Concepts: There are things you hear over and over again, read through the documentation and trying to grasp, but it just never clicks. Claude’s ability to explain something in various levels of detail, patiently and without any feeling of remorse when you ask again and again, made me understand a few things better than I did before.
• Easy Digital Download Custom Dashboard: I always wanted to build a custom dashboard for EDD, showing exactly the numbers we want to know, and nothing else. I went from 0 to a working proof-of-concept in just 15 minutes.
• Easy Digital Download Simple Tracking: I thought about building a simple way of tracking which sites people landed on initially, who purchased later on. Here again, it took me only 15 minutes and I had a working prototype. Not perfect, but enough to get started.
• How about Social Science: To see how Claude will do in a completely different field, I talked to my fiancée and she gave me some questions about her field, Social Science. Particularly about “Systemic Solution-Focused Counseling in Coercive Contexts”, which is the topic she chose for her thesis and knows a thing or two about. To be honest, I didn’t understand a lot from the answers, but seeing her reading and nodding, it looked like it struck a chord. She also did ask for my login afterwards.
• Creating a “Security Checklist” for our maintenance clients: we provide site reviews for our clients and we always wanted to extend this to security checks as well. It took me 2 hours to outline a new checklist from scratch, and that included reading some blogposts, gathering information elsewhere and even testing some of the tools suggested.
• Marketing Strategy for picu: I asked what it would do with a budget of $500 a month to promote a plugin like ours. It did get some numbers wrong, and made a few assumptions out of thin air, but overall it gave me some good ideas, and I’m intrigued to follow up on some of them.
• Explaining how LLMs work (how meta, I know): I had asked a friend about this before, who studied computer science and today works as a teacher in the field. The answer I got from Claude was impressively close to what my friend gave me a few months ago. Both pointed me to N-grams and Claude even gave me some example code to play around with. I felt like I learned something with very little effort.
• Arduino / Hardware: A few years ago, I toyed around with building some fun things with microcontrollers. Clearly an LLM can’t help with hardware, right? Wrong. It was even able to give me schematics for soldering the parts together, which I haven’t tried yet but looked legitimate and would have been a huge time saver for such a project.
• Tax laws and business expenses in Switzerland: Our accountant recently explained a pretty specific thing about taxes and business expenses to me. I thought this could be a good test because it’s so niche and specific to Switzerland. Claude explained it correctly, and about as good as our accountant did, and while I wouldn’t want to ditch our accountant for an AI assistant (at least not yet), it was pretty impressive (and much cheaper, if I may say so).

This is just a small portion of the random things I tried. Claude was also able to understand and “speak” in Swiss german, even in different dialects, and I used it to upload images to analyze or create ALT texts for them. I even had Claude explain Kantian philosophy to me, which might have been unfair to begin with, but it did a decent job as far as I can tell.

In the end, what I found it most useful for, was this feeling of “getting things off the ground” quickly, validating and testing assumptions or generating and analyzing different variations of something I already had. It made playing around with such ideas feel much more direct and natural.

Would I use any of that code or directions it gave me directly in production? Not really! But was it productive and helpful so far? Absolutely!

Also, I was impressed how rarely it produced true “hallucinations”. I expected much more invented stuff. There were some numbers it got wrong and I generally wouldn’t want to depend my life on the results, but overall it did help me and felt incredibly natural doing it. Much more like talking to an assistant than a computer.

I certainly will keep using and testing it for a while and it will be interesting to see if I still find it useful, once the initial excitement fades away.

What are your experiences with “AI” so far?