[alexathrowawa9]

Throwaway here but I used to work in Alexa org at Amazon and was amazed by how big the org was (thousands and thousands of people) considering it doesn't seem to be a big revenue generator

I remember constantly hearing about projects other teams were working on thinking "why would anyone use that" or "how would that ever make money"

Just trying to shoehorn alexa into as many domains as possible

It was like empire building at its finest

I would joke that the canary tests were the biggest customer for a lot of services

And the way amazon works with SOA even what seems like a small feature ends up being a couple services, a pizza team of 10 devs + SDM, the overhead is huge

Back when it was announced alexa org was being hit harder by layoffs that did not surprise me

[asdasdsddd]

I've interviewed many people on Alexa before. From what I gather, its just a giant switch statement, and each individual "path" takes a bunch of effort to support and there are thousands of paths for music, ordering, commands, etc. It's peak AI == if statement architecture.

[ihkasfjdkabnsk]

Throwaway, used to work at the NLU unit of Alexa about 5 years ago. There is some ML going on but as with all ML projects I have worked on people want control. This means you add rules for the "important" stuff. You also add test cases to make sure the ML works. But if you already have those test cases, why not just match on them directly? There are also advanced techniques for generating examples (FST for example).

What this culminated in is a platform where 80% of request, and pretty much 99% of "commands" are served by rules built with a team of linguists.

[novok]

IMO with my experience with siri being AI-style unreliable in many ways, like bit flips when saying turn off the lights makes the dimmer go to %100, I think it's better to do the switch statement for the dozen or so query types that probably represent %90 of traffic, like weather, music, home control, unit conversions, etc in exchange for way more reliability.

[laidoffamazon]

I think describing the NLU Engine as a switch statement is underselling it a bit. Determining domain and intent alone requires more than that (frequently, at least).

[SahAssar]

Do you mean actual coded if statements as in actually human written code like

    if (question.match(/^what is (.*)/)) return wikipedia.search(question)

or something more automated?

[visarga]

> I've interviewed many people on Alexa before.

I got a recruiting message once for a ML engineer position in Alexa, ignored it.

[brazzy]

I guess that's what happens when you're generating absurd amounts of revenue and want to "reinvest" it all: anything that even vaguely smells of "innovation" gets money thrown at it like crazy, and you end up incentivizing bullshitting.

[bloggie]

My optimistic side tells me that Blue Origin was founded as a tech lab / app lab / skunkworks and not necessarily as a company with the goal of putting humans in space, that has suffered from said bullshitting. But my cynical side tells me the latter.

[deepfriedrice]

> Just trying to shoehorn alexa into as many domains as possible

It happened outside of Alexa too. Every team with a public facing product was directed (it seemed) to come up with some sort of Alexa integration. It was usually dreamed up and either a) never prioritized or b) half assed because nobody (devs, PMs, etc.) actually thought it made any sense.

[spike021]

I once heard about a feature dogfooding invite that was sent out specifically for people with babies because they wanted to use Alexa always-listening to activate when a baby was crying and automatically order diapers or something ridiculous like that.

[bagels]

Doesn't even make sense. Babies cry all the time for a multitude of reasons, none of which are informed by how many diapers are in the house.

[alexathrowawa9]

I remember getting that invite!

You could use "baby crying detected" as an automation trigger

[laidoffamazon]

It wasn't for diaper ordering, Alexa has a feature to detect particular sounds (like glass breaking, dog barking etc) for monitoring and alerting purposes. Basically it involved adding hotword recognition for not just the N number of hotwords but also Y number of sounds for particular devices.

[bobnamob]

> I would joke that the canary tests were the biggest customer for a lot of services

This is true for a surprising number of amzn/aws products

[Balgair]

Hey, I gotta ask a few questions here, given this oportunity:

How many people actually 'worked' there? Like, really did something all day?

What was the pay like?

What were the internal politics like?

Any good stories?

[alexathrowawa9]

> Any good stories?

One time I walked into a dark room with like 50 test devices to get something and somehow accidentally triggered them and all 50 started talking at the same time

Was both hilarious and creepy

[baxtr]

I would love to see the initial “vision doc”. How were they envisioning to make lots of money with it in the first place?

[SomeCallMeTim]

I am actually surprised that there are thousands of people working on Alexa.

WTF are they all doing?! It's pretty much unchanged from the outside in any way that's relevant to me compared to where it was in 2014. And the few changes I've noticed have been things breaking.

I used to, for instance, have a script (I forget the Alexa term) that would turn off a few lights and then play a Pandora radio station when I gave it a "bedtime" command. Worked great for about a year, and then the Pandora plugin suddenly refused to take any combination of commands that I could figure out to play a particular Pandora station in my account. This is true from outside of the automation as well, by the way. It's just completely broken.

The weather app integration is annoying too. I wanted weather to use a different weather source, and instead of just giving me results from that weather source, it would always preface it with "Weather from BlueSky" or whatever. Maybe it's their fault and they wanted the ad blurb? But as a consumer, it sucked. I just wanted more localized weather, not an ad every time I asked for the weather.

And the "AI" behavior of the app...it was just awful. I could get better answers from Google Home devices across the board. The best Alexa would do if I asked it a question is to read the first paragraph of a Wikipedia entry, and it was about a 1 in 4 chance it would actually choose the correct Wikipedia entry.

OH, and don't get me started on the Android Alexa app (!!!). Again, the most major change was a UI update where the most important feature I ever use was hidden behind another layer of menus for no particularly good reason. And the "Kindle Accessibility" feature of reading Kindle books is so flaky I doubt anyone on the team ever uses it, from random pauses to sudden jumps back to read from the beginning of the section of the book you started on 10 minutes ago, looping forever on those same 10 minutes.

Sorry. I know it wasn't your fault. But I finally gave up on using Echo devices, and the only reason I still even have the Alexa app on my phone any more is so I can have it read a Kindle book while I'm driving, and it's so amazingly frustrating to use that it would likely be better if it didn't even exist. It's more "customer frustration" than useful.

[xnx]

> Just trying to shoehorn alexa into as many domains as possible

The original AI technology spamming: Alexa Integration

[conductr]

My take on it was that Alexa was developed at a time they were trying really hard to solidify their continued existence as a tech company; AWS and kindle came around then too. The stock market values tech companies much higher than retail companies. That’s what justified the headcount expense

[nickm12]

AWS and Kindle were developed a decade before the Echo.