“A part of any deep learning or machine learning project is to explore the data because there's probably 10 or 20 different models you could use, but it's how you choose them for the machine to draw that best-fit line, and there's different ways: nearest neighbour, linear regression, pick one of a hundred thousand of those.”

“You need to work out what the answer is at small scale, quickly, and then once you're going to industrialise that process generally, you do that on-premises afterwards, because it becomes very expensive to continually re-train your machine, which you should absolutely be doing, in the public cloud.”

“The other key lesson is never to take your eye off the ball, because drift or A/B testing or canary testing, that's always keeping it real so that your assumptions on day one may not be valid on day 90 because maybe you've changed your customer base, maybe you've improved your product, whichever the different vector is, that you always have to keep coming back to: 'is my model still relevant?'”