A real failed case of applying ML to a real life scenario A rare specimen. Contrary to the ordinary mantra of just "adding more layers" some info from the real project: - https://andlukyane.com//blog/medical-chat-bot Takeaways: - Too complex - Data too fuzzy labeling too complex - Pre-trained models useless for real life cases - Real life annotation is full of pain and gotchas - They also used fast-text like tokenization - Simple models - Wild, deep, complex in-domain data engineering rules PS - So sad to see trillions of rubles spent on chasing trendy topics like image generation and public datasets ... are nowhere - I wonder what % of trendy projects reported by AI firms are like this - Lack of leadership and care in corporate world I appreciate the author's courage in coming out with this ... after he has moved to another country, lol.