AI at Amazon: A case study of brittleness

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.

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.

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!

(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.

Even if the word AI was not involved, this just sounds like Amazon's standard, well documented toxicity.

I remember Google getting rid of a large part of it's Assistant team at the same time Amazon laid off a large part of its Alexa team ~18 months ago. You can even read the "is Amazon closing down Alexa?" rumours in that time.

It's pretty clear that LLMs with action hooks were going to take over from the old bespoke request->response methods so i guess they were trying to make sure there was no old guard holding on in the changeover. Only just now are the Alexa+ and Google home Gemini integrations becoming available to pick up where they dropped off in 2023.

Apple had a few Siri layoffs but it seemed to keep the ship steady. It'll be interesting to see which was the better long term approach though.

Worked at AWS in the past. This resonates. Let's of missed chances. For example, the Alexa for Automotive team initially impressed with a well-produced demo video that painted a compelling vision of in-car voice integration. However, that vision unfortunately never translated into real-world execution. Despite having a major opportunity to collaborate with one of Asia’s largest automotive manufacturers—an ideal partner to bring Alexa into the vehicle ecosystem—there was little to no follow-through. The opportunity was essentially ignored.

From my perspective, one of the core issues was cultural. The Alexa teams were often staffed by long-timers with substantial RSU grants, many of whom appeared more focused on preserving internal influence and career security than driving bold external partnerships or innovation. It indeed felt less like a team pushing the envelope, and more like a collection of fiefdoms guarding their territory.

In the end, it was a missed opportunity—not just for Alexa, but for Amazon to play a central role in the connected car revolution.

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.

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.

The elephant in the room is that Amazon sucks to work at.

Every ex-Amazon employee I’ve worked out talked about their burnout culture.

They under-compensate compared to their peers and as this article touches on with the discussion about customer focus, their corporate culture is the most draconian and abnormal.

I did an early stage Amazon interview and they basically wanted me to memorize every detail of their company culture and relate every single one of those aspects to a piece of work I did in the past. They wanted me to demonstrate that I had essentially joined their cult before I even had the opportunity to join it!

I have no idea how someone is supposed to honestly complete that interview process without outright lying.

Does anyone know what we learned? Looks like the second problem (decentralization) is the antidote to the first (centralization) which was rather amusing.

Probably one of those things which are actually perfectly fine trade-offs, operational challenges and in no way causal to the demise. If Alexa had found a market, the same article would probably be called "AI at Amazon: a case study of <insert management buzzword>" and explain to us how the same processes paved the path to success.

A good CEO would have forced alexa to become the conversational AI in the world once GPT4 dropped.

    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?

Indeed, how does a "customer obsessed" organization cram features down their customers throats that they don't want?

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.

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.