Google Updates in 2021 and What To Expect in 2022 with Lily Ray 14 December 2021 at 17 H CEST (Paris)
In this clip, Jason Barnard and Lily Ray discuss some of the limitations of Machine Learning and possible overreach by Google. We are all sometimes guilty of imagining that the machine can do more than it actually can (even Google engineers). It is helpful to remember that the machine is purely predictive: when it does get something wrong, things go south very fast. Jason also mentions the dance between humans and machines (as described by Andrea Volpini from Wordlift).
00:00 Machine Learning
00:47 Machine is purely prediction
00:57 Lessons to manage machine learning
Transcript from: How Good is Google’s AI (Machine Learning)?
They’ve got less adventurous or they’re letting the machine less loose. And I think that’s part of the ML (machine learning) thing is that maybe sometimes, we all get the impression that the machine can do more than it actually can and the machine, when it gets it wrong, it gets it phenomenally wrong. And Andrea Volpini from Wordlift talks delightfully about the dance between humans and machines and not to forget that the machine… What was it? The machine can’t count to three. It doesn’t actually know what it’s saying.
That’s reassuring! And it’s a predictive thing whereby you give it one word and it predicts the next word, then the next word. Got it! And obviously, when that goes wrong and it starts to go wrong, the machine just gets it increasingly horribly wrong as it moves forward, because it’s purely prediction. Right! Word on word. So it doesn’t actually know what it’s saying and that is scary. So you’ve got to be incredibly careful.
Now I’m giving lessons to Google about how to manage that machine learning. I do apologize to anybody from Google who might be watching this because obviously, I know nothing. I always want to be clear. What they’re doing is it’s extraordinarily hard work. And of course, it makes sense that there are errors and things that need to be rolled back. But to me, (the Meta Title debacle of Fall 2021) was one of the more extreme examples of things going wrong that I’ve seen.