I find this disingenuous. Yes, its important that the algos can perform well on real world data, but the framing of this post begins with an anecodote about one

I find this disingenuous. Yes, its important that the algos can perform well on real world data, but the framing of this post begins with an anecodote about one person who had a bad model, and implicitly extrapolates that these problems are generalized throughout all neural nets. One could say the same thing about programmers automating a task, or a number of other trivial examples. I would lean towards assuming deep mind has competent model validation teams vs. not, even if data science is hard.