Three years of putting LLMs into Software - Lessons learned

After GPT 3.5 came out, many wanted to augment their software with LLMs - or, AI, as most call it. We built a lot of such features into different software, with transcription, commands, live feedback, different services, that made us sometimes feel superhuman and sometimes humbled us down. So, after three years, its time for a recap - I’ll show real results, talk about (exactly) how we made it, what the customers liked about it and what they disliked. I’ll show the differences between commercial and self-hosted, discuss prompting, fine-tuning and when-not-to-finetune. I’ll talk about the time we needed to add those functionalities, and how well it did work - and the talk will be kept up-to-date until the last week before the event ;)