AI has problems - and Google/YouTube's AI is awful at speech recognition

AI is all the rage - in theory. In practice, though, it can leave a lot to be desired.
Speech recognition is one of the earliest uses of AI, and dates back to 1972, but Google - the giant AI company - still has YouTube Auto-AI-Generated captions that are still pretty bad.

What does that say about the practical use of AI at scale? What can we learn from this as people who want to use AI in our projects?

00:00 AI isn't useless
01:06 The AI bug of the day
01:32 Introducing the bug
03:12 Examples
05:22 What might this mean about real-world AI usage?
08:13 Educated Speculation
11:33 Likely an issue with context - for AI - again
12:05 Guess-timating AI project success risk
13:57 You can learn this better than an LLM-based AI "software developer" can
14:22 Wrap-up

Links from the video:

How To Fix YouTube Captions Manually video:
https://www.youtube.com/watch?v=TCx6dJOiM_Y

AI Protein Folding won the Nobel prize in Chemistry in 2024: https://www.nature.com/articles/d41586-024-03214-7

First AI Find of 0-Day Security Vulnerability by Google: https://www.forbes.com/sites/daveywinder/2024/11/05/google-claims-world-first-as-ai-finds-0-day-security-vulnerability/

Computers recognizing speech:
https://www.senstone.io/a-very-brief-history-of-speech-recognition/
https://verbit.ai/ai-technology/from-audrey-to-siri-the-evolution-of-speech-recognition-technologies/

There's a phrase you may hear in computational search: TF-IDF
https://www.learndatasci.com/glossary/tf-idf-term-frequency-inverse-document-frequency/
https://www.emerald.com/insight/content/doi/10.1108/eb026526/full/html

Devin seems not to work out of the box, either:
https://www.answer.ai/posts/2025-01-08-devin.html

Only a tiny fraction of videos get very many views - graph:
https://pex.com/blog/what-content-dominates-youtube/