I've noticed that we haven't yet written an article about our day-to-day work with LLMs. I'd like to make up for this - without having ChatGPT or another LLM write the text, of course. By the way: LLM stands for Large Language Model.
LLM & Podcast
With now over 150 episodes, we publish a new episode of our Deep Dive podcast about open source software every weekday at 12:00. We have automated around 80% of the production process so that - apart from listening and re-listening - each episode only takes around five minutes. This is easy to prepare and we are usually four weeks ahead in the editorial schedule.
The episodes are created by NotebookLM. The transcription is done via Whisper, the summary is provided by Kagi, and one or two concise sentences are generated from the key moments. These appear automatically in the Fediverse on publication - via Castopod.
Have a listen to get an idea of the quality. You will also learn a lot about free software and digital sovereignty.
LLM & images
Automatic image generation is almost old hat by now. Many of the header images we needed for our last website relaunch were created with Stable Diffusion. Back then, we often used the generator from perchance.org - we wouldn't do it that way today. The world has moved on and there are now more powerful alternatives available in the form of tools such as Reve or Flux.
LLM & texts
Of course, we also use ChatGPT - but not exclusively. We also regularly try out other models, especially smaller versions that we can host ourselves. This has several advantages: We retain full control over the content, do not have to involve external service providers and remain data protection compliant within the EU.
LLM & code
When it comes to code, we take a clear stance: what is already public can also be processed with SaaS-based LLMs. Especially for repetitive or rather monotonous tasks, an LLM is a welcome “colleague”. For more sensitive, non-public code, we are currently evaluating locally hosted models that can support us securely and efficiently - an exciting side project.
LLM & Music
Suno recently released a new update - and I was immediately impressed with the quality. I did a few tests, including with music from my previous musical life.
Lost Dogs - Van Thies
Lost Dogs - covered suno.com
As an example, I used the track totallyfreesong by Mark Silverman - an experiment worth listening to. Perhaps you can also do something internally with LLM-generated music? For example as a jingle for your next podcast?
Totally Free Song!!! by Mark Silverman
Totally Free Song!!! covered by suno
Totally Free Song!!! covered by suno
Totally Free Song!!! covered by suno
Totally Free Song!!! covered by suno
LLM & Law
LLM-generated content raises exciting legal questions: Who actually “owns” such a work? Is there even a creative achievement that can be protected by copyright? And is an LLM even a “somebody”?
The common assessment is that content generated by AI is not covered by copyright law - and is therefore not eligible for protection.
In plain language: Do you like an image on our website that was created with an LLM? Then feel free to copy it and use it freely - without proof.
The same applies to our podcast episodes: copy, redistribute, no problem.
The situation is different for texts. Unless otherwise stated, our contributions are licensed under a Creative Commons license - at least with attribution.
Things get exciting with services like Suno. A commercial license for AI-generated music is only available there with a paid subscription. The cover songs we uploaded from totallyfreesong by Mark Silverman had such a subscription.
So are we not allowed to commercialize the derived works and publish them as public domain?
Honestly? We don't know exactly.
We do it anyway - out of conviction and in the spirit of Mark, whose original idea we greatly appreciate and want to carry on.
Note: This is of course not legal advice, but our personal assessment and position on the subject.
