ai.bythebay.io: Chris Fregly Interview
Recording: ai.bythebay.io: Chris Fregly Interview
you [Music] you [Music] uh AI it's the ability to find sort of like finer grain classifications or and write like these days so I'm starting to see a lot of people trying to just go straight to like AI right now so as far as it relates to my role I spend a lot of time thinking about right like deploying so there's the early pipeline part of it where you're actually training and stuff that that's being tackled by like so many people but the ability to actually deploy them and so tensorflow has a different model of deployment than say like p.m. ml or scikit-learn so been spending them just a lot of time I'm like trying to optimize things after models have been trained like for example and and just the ability to do experimentation which I'm actually seeing a lot of people really compare AI models to their traditional scikit-learn ml models so yeah like a lot of experimentation and performance comparison things like that yeah I think one of the things that's put right like you can always think back and say okay spark spark had a lot of examples when it first came out it was very approachable there was a whole ton of support it sort of had this academic background and you know people got excited about that I think one thing that made spark pretty fun was the the focus on machine learning right there's a lot of etl and everything yeah so really obviously with tensorflow google has just right like tons and tons of like examples and resources and that kind of thing and I think something they could do better they there's this tool called tensor board right which kind of helps you try to reason through the actual learning process about how it's actually learning things it's kind of a fundamental problem with AI is just trying to understand rather how did it get rightly to these different classifications like how to each node learn one particular you know piece of clothing or one stripe or one edge or something like that it's detecting so I mean I don't think it's something even like the best tool designers can help build right now the current state of tensorflow I think there's quite a bit of there's still quite a bit of confusion as to what's the actual best like high level library to use for it because yeah comfortable itself is super low level which kinda reminds me of spark as well when the first came out you know spark had kind of all these like really really low level operators and then they started to build more sequel like semantics to kind of appeal to more and more people but yeah yeah well I mean from a competitor type thing I mean I view services like Google's cloud ml you know which are starting to dabble in this experimentation and yeah so Google obviously knows tensorflow early in and out right there the creators of it they're they're getting better about the the serving of it and sort of what what happens after you built the model right and so their services like the Amazon Riley ml service and you know just I guess making it like easier to actually deploy these things and just have a couple clicks to say all right yeah this is experiment a experiment be if you start doing multi-armed bandit Riley type of experiments where you can slowly shift traffic throughout the experiment to the rarely winning model and just sort of from like the entire experiment like you can sort of overall rather optimized for that and not just the absolute winning sitting at the end of it kind of a waterfall approach to testing which is just wait you know three weeks and then flip to the new one versus kind of slowly shifting traffic yep [Music] yeah it's fine I like the AI spin I've seen a couple of the same speakers that I've seen throughout the other by the base series but they've got new talks they've got talk specific to a I that was one thing I was right like starting to struggle with a little bit as I had a lot of material and spark and I really wanted this to be read like tensorflow specific so it's fun to see the same principles even you know that are there in spark and other technologies that that are being transferred over and different ways to do it in different tools and stuff so yeah it's good to see all this okay so any sweet man [Music] you