I was with Amazon but wasn't part of Alexa. I was working closely with the Alexa team however.
I remember vividly the challenge of building centralized infra for ML at Amazon: we had to align with our organization's "success metrics" and while our central team got ping ponged around, and our goals had to constantly change. This was exhausting when you're trying to build infra to support scientists across multiple organizations and while your VP is saying the team isn't doing enough for his organization.
Sadly our team got disbanded eventually since Amazon just can't justify funding a team to build infra for their ML.
giancarlostoro 58 minutes ago [-]
> Amazon just can't justify funding a team to build infra for their ML.
Sounds like they didn't plan it out correctly. It should have been done in phases, one team at a time, starting with the Alexa team, or the smallest team with the smallest amount of effort as a test bed, while keeping the other teams informed in case they have suggestions or feedback for when their turn comes along.
didip 58 minutes ago [-]
Isn't this a challenge for any big tech companies. Success metrics tend to be attached to a particular product and yet... central infra is a necessary foundational block for everyone, but it is without a specific success metrics.
coredog64 3 hours ago [-]
Not this article, but the one that it references:
> And most importantly, there was no immediate story for the team’s PM to make a promotion case through fixing this issue other than “it’s scientifically the right thing to do and could lead to better models for some other team.” No incentive meant no action taken.
Oof!
simplesimon890 3 hours ago [-]
Unfortunately it aligns with internal pressure. If you are not working on something that has clear quantifiable/promotional benefits that can be realized within 2-3 quarters, you are at risk of the PIP train. Having buy in from senior management can help but in a company that re-orgs regularly, your manager can change, so the risk is higher.
wnevets 7 minutes ago [-]
> No incentive meant no action taken.
sounds like late stage capitalism
draw_down 3 hours ago [-]
[dead]
PaulHoule 4 hours ago [-]
Some of it is the rapid progress in fundamental research.
If in Q2 2025 a company like AAPL or AMZN decides to invest in a current top of the line neural network model and spend 18 months to develop a product, whatever they develop might be obsolete when it is released. Holds for OpenAI or any incumbent -- first mover advantage may be neutralized.
Secondly there are a lot of problems in ambient computing. Back in the early 00's I often talked with an HCI expert about ideas like "your phone (pre-iPhone) could know it is in your backpack" or "a camera captures an image of your whole room reflected in a mirror and 'knows what is going on'" and she would usually point out missing context that would make something more like a corporation that gives bad customer service than a loyal butler. Some of Alexa's problems are fundamental to what it is trying to do and won't improve with better models, some of why AMZN gave up on Alexa at one point.
mrweasel 4 hours ago [-]
It's also not entirely clear what Alexa was suppose to do, nor Siri for that matter. Being a personal digital assistant turned out to be much less useful than many imagined and being a voice controlled Bluetooth speaker is mostly a gimmick outside the car or kitchen.
That's not to say that Alexa and others can't be useful, but just not to enough people that it justifies the R&D cost.
cle 3 hours ago [-]
Meanwhile multiple non-technical people that I know pay $20/mo to OpenAI and have long, verbal conversations with ChatGPT every day to learn new things, explore ideas, reflect, etc.
These are obviously what voice assistants should do, the research was just not there. Amazon was unwilling to invest in the long-term research to make that a reality, because of a myopic focus on easy-to-measure KPIs. After pouring billions of dollars into Alexa. A catastrophic management failure.
mrweasel 3 hours ago [-]
Are they talking to ChatGPT, or are they typing? More and more we're seeing that user don't even want to use a phone for phone calls, so maybe a voice interface really isn't the way to go.
Edit: Oh, you wrote "verbal" that seems weird to me. Most people I know certainly don't want to talk to their devices.
ljf 3 hours ago [-]
My wife paid for ChatGPT and is loving it - she only types to it so far (and sends it images and screenshots), but I've had a go at talking to it and it was much better than I thought.
If I'm alone I don't mind talking if it is faster, but there is no way I'm talking to AI in the office or on the train (yet...)
throwaway314155 2 hours ago [-]
> If I'm alone I don't mind talking if it is faster
When is talking faster than text? I only ever use it when my hands are tied (usually for looking up how to do things while playing a video game).
Retric 2 hours ago [-]
When you can talk at your normal pace?
People talk at about 120WPM - 160WPM naturally, few can type that fast which is why stenographers have a special keyboard and notation.
PaulHoule 5 minutes ago [-]
I feel tiredness in my throat when I talk to bots like Alexa as you have to enunciate in a special way to get across to them.
throwaway314155 6 minutes ago [-]
I struggle to have naturally flowing conversation with an AI for much the same reason people don't use most of Siri's features - it's awkward and feels strange.
As such I can maintain about five minutes of slow pace before giving up and typing. I have to believe others have similar experiences. But perhaps I'm an outlier.
taeric 3 hours ago [-]
I continue to be baffled that they are going to cannibalize the "voice controlled radio and timer" market in the chase of some "magic assistant" one.
It would be one thing if they were just adding extra "smart home" features to connect new terminals. I can see benefit of some of the smart screen calendar and weather things. No, they seem dead set to completely kill what they had.
taormina 4 hours ago [-]
Even if the word AI was not involved, this just sounds like Amazon's standard, well documented toxicity.
alex-mohr 3 hours ago [-]
And you could write a similar blog post about why Google "failed" at AI productization (at least as of a year ago). For some of the same and some completely different reasons.
- two competing orgs via Brain and DeepMind.
- members of those orgs were promoted based on ...? Whatever it was, something not developing consumer or enterprise products, and definitely not for cloud.
- Nvidia is a Very Big Market Cap company based on selling AI accelerators. Google sells USB Coral sticks. And rents accelerators via Cloud. But somehow those are not valued at Very Big Market Cap.
Of course, they're fixing some of those problems: brain and DeepMind merged and Gemini 2.5 pro is a very credible frontier model. But it's also a cautionary tale about unfettered research focus insufficiently grounded in customer focus.
BryanLegend 4 hours ago [-]
The new Alexa+ is super great, if you've been invited to it.
The voice recognition is at a whole nother level, much much faster. Controlling lights is easily an entire second faster.
The TTS upgrade is a trip. She sounds younger and speaks faster.
WalterBright 3 hours ago [-]
> Controlling lights is easily an entire second faster
We easily use Alexa 20x a day at my house. Announcements, kitchen timers, lights & music. Even checking store hours or weather.
Replaces a phone in many cases.
otterley 2 hours ago [-]
That works great if you're in the room, but not so much if you're not home (or aren't even in the same room), or want to control landscape lighting and want it to automatically adapt to the seasons.
diggan 4 hours ago [-]
> Controlling lights is easily an entire second faster
How long time did it used to take VS how long time does it take now? I'm not sure "an entire second faster" is sarcasm here, big improvement or what.
BryanLegend 1 hours ago [-]
There used to be second or two between finishing a command and it getting executed. We easily use Alexa 20x a day at my house. Announcements, kitchen timers, lights & music.
I think the voice recognition is async now. It's streaming the data to a model. Before it would wait until the command was finished then send the .wav file off to a model.
nsonha 3 hours ago [-]
is it "ChatGPT in a box" level yet? If not why has that not been a thing?
BryanLegend 1 hours ago [-]
It is. It will make up a story if you ask it to. Answers questions with it's own knowledge now instead of using the amazon answers site.
PartiallyTyped 3 hours ago [-]
The problem with ChatGPT in a box and similar is that they are not made for live interactions. If you've tried chatGPT or claude on your phone with voice conversations you will see that it takes a while to think.
Humans on the other hand start processing the moment there's a response and will [usually] respond immediately without "thinking", or if they are thinking, they will say as much, but still respond quite quickly.
throwaway314155 2 hours ago [-]
Surely whatever solution they came up with amounts to "ChatGPT in a box" (using an llm with fewer parameters for speed).
awsthrowaway1 1 hours ago [-]
(Throwaway account because I work at Amazon)
Everyone at Amazon is focused on AI right now. Internal and external demand for GPU resources and model access is off the charts. The company's trying to provide enough resources to do research, innovate, and improve business functions, while at the same time keeping AWS customers happy who want us to shut up and take their money so they can run their own GPUs. It's a hard problem to solve that all the hyperscalers share.
4 hours ago [-]
adolph 1 hours ago [-]
This introduced an almost Darwinian flavor to org dynamics where teams
scrambled to get their work done to avoid getting reorged and subsumed into
a competing team.
To the extent that an organization is so wealthy and vast that it can fund redundant efforts, isn't getting reorged into the "winning" team a good thing?
Symmetry 16 minutes ago [-]
If a successful research program would take longer than it would take to be absorbed that can be a problem.
SpicyLemonZest 59 minutes ago [-]
It's not the worst thing in the world, but it's usually much better for your career to have other people reorged into your team. Even if the reorg doesn't reset promotion progress for everyone who moves (which it often does!) that way you get to demonstrate leadership ramping the new folks up on how things work in your neck of the woods.
aaroninsf 36 minutes ago [-]
There is a small pleasure to be had in the fact that the relentless decent into dystopian surveillance capitalism was apparently momentarily retarded by such venal banalities as employee recognition and compensation schemes having entirely devolved at FAANG into "career hacking" gimicks.
ywxdcgnz 2 hours ago [-]
[dead]
aaroninsf 42 minutes ago [-]
What's most striking to me in this article is how perfectly this sums up the ongoing collapse of America's political system and civil society:
"In the paper Basic Patterns in How Adaptive Systems Fail, the researchers David Woods and Matthieu Branlat note that brittle systems tend to suffer from the following three patterns:
- Decompensation: exhausting capacity to adapt as challenges cascade
- Working at cross-purposes: behavior that is locally adaptive but globally maladaptive
- Getting stuck in outdated behaviors: the world changes but the system remains stuck in what were previously adaptive strategies (over-relying on past successes)"
Painfully apt.
30 minutes ago [-]
Rendered at 19:02:01 GMT+0000 (Coordinated Universal Time) with Vercel.
I remember vividly the challenge of building centralized infra for ML at Amazon: we had to align with our organization's "success metrics" and while our central team got ping ponged around, and our goals had to constantly change. This was exhausting when you're trying to build infra to support scientists across multiple organizations and while your VP is saying the team isn't doing enough for his organization.
Sadly our team got disbanded eventually since Amazon just can't justify funding a team to build infra for their ML.
Sounds like they didn't plan it out correctly. It should have been done in phases, one team at a time, starting with the Alexa team, or the smallest team with the smallest amount of effort as a test bed, while keeping the other teams informed in case they have suggestions or feedback for when their turn comes along.
> And most importantly, there was no immediate story for the team’s PM to make a promotion case through fixing this issue other than “it’s scientifically the right thing to do and could lead to better models for some other team.” No incentive meant no action taken.
Oof!
sounds like late stage capitalism
If in Q2 2025 a company like AAPL or AMZN decides to invest in a current top of the line neural network model and spend 18 months to develop a product, whatever they develop might be obsolete when it is released. Holds for OpenAI or any incumbent -- first mover advantage may be neutralized.
Secondly there are a lot of problems in ambient computing. Back in the early 00's I often talked with an HCI expert about ideas like "your phone (pre-iPhone) could know it is in your backpack" or "a camera captures an image of your whole room reflected in a mirror and 'knows what is going on'" and she would usually point out missing context that would make something more like a corporation that gives bad customer service than a loyal butler. Some of Alexa's problems are fundamental to what it is trying to do and won't improve with better models, some of why AMZN gave up on Alexa at one point.
That's not to say that Alexa and others can't be useful, but just not to enough people that it justifies the R&D cost.
These are obviously what voice assistants should do, the research was just not there. Amazon was unwilling to invest in the long-term research to make that a reality, because of a myopic focus on easy-to-measure KPIs. After pouring billions of dollars into Alexa. A catastrophic management failure.
Edit: Oh, you wrote "verbal" that seems weird to me. Most people I know certainly don't want to talk to their devices.
If I'm alone I don't mind talking if it is faster, but there is no way I'm talking to AI in the office or on the train (yet...)
When is talking faster than text? I only ever use it when my hands are tied (usually for looking up how to do things while playing a video game).
People talk at about 120WPM - 160WPM naturally, few can type that fast which is why stenographers have a special keyboard and notation.
As such I can maintain about five minutes of slow pace before giving up and typing. I have to believe others have similar experiences. But perhaps I'm an outlier.
It would be one thing if they were just adding extra "smart home" features to connect new terminals. I can see benefit of some of the smart screen calendar and weather things. No, they seem dead set to completely kill what they had.
The voice recognition is at a whole nother level, much much faster. Controlling lights is easily an entire second faster.
The TTS upgrade is a trip. She sounds younger and speaks faster.
I just use the Clapper from the 1970s.
https://www.amazon.com/Clapper-Activated-Detection-Appliance...
Replaces a phone in many cases.
How long time did it used to take VS how long time does it take now? I'm not sure "an entire second faster" is sarcasm here, big improvement or what.
I think the voice recognition is async now. It's streaming the data to a model. Before it would wait until the command was finished then send the .wav file off to a model.
Humans on the other hand start processing the moment there's a response and will [usually] respond immediately without "thinking", or if they are thinking, they will say as much, but still respond quite quickly.
Everyone at Amazon is focused on AI right now. Internal and external demand for GPU resources and model access is off the charts. The company's trying to provide enough resources to do research, innovate, and improve business functions, while at the same time keeping AWS customers happy who want us to shut up and take their money so they can run their own GPUs. It's a hard problem to solve that all the hyperscalers share.
"In the paper Basic Patterns in How Adaptive Systems Fail, the researchers David Woods and Matthieu Branlat note that brittle systems tend to suffer from the following three patterns:
- Decompensation: exhausting capacity to adapt as challenges cascade
- Working at cross-purposes: behavior that is locally adaptive but globally maladaptive
- Getting stuck in outdated behaviors: the world changes but the system remains stuck in what were previously adaptive strategies (over-relying on past successes)"
Painfully apt.