Amazon’s ecommerce business has summoned a large group of engineers to a meeting on Tuesday for a “deep dive” into a spate of outages, including incidents tied to the use of AI coding tools.
The online retail giant said there had been a “trend of incidents” in recent months, characterized by a “high blast radius” and “Gen-AI assisted changes” among other factors, according to a briefing note for the meeting seen by the FT.
Under “contributing factors” the note included “novel GenAI usage for which best practices and safeguards are not yet fully established.”
“Huge rich company responsible for hosting like half of the fucking internet spent the last year pushing code to global-scale production without so much as a review by a senior engineer.”
That’s how I read that headline.
I read it as “now a senior developer will be at fault for all AI code.” Do you think they will have time to review all that code properly and do their jobs.
One of my first big jobs at NASA was as a lead engineer on a multi-experiment platform to fly on the space shuttle. I checked all the work and compiled all the data and trotted my 27 year old self down to Johnson to present my case to the Safety Board. When I stood up to present, the head of the panel asked if I knew why I was there. I confidently told him that I was there to walk them through my evaluation of each of the payload components and show that the payload was safe to fly. He smiled. He then said “You’re here because if something goes wrong on this mission, there had to be one ass to kick. Proceed.”
Everyone needs an ass to kick, and AI doesn’t offer that function.
That sounds like an almost refreshing “you’re one of us now / welcome to the real thing” type of brutal honesty.
Did it have a friendly tone and/or serve as an ice breaker before your presentation?
They will save time by making them go pee in bottles
The preview for the reply notification for this comment started getting my brain so excited when my eyes scanned over the beginning. Screen grab:
What is AI good at? Creating thousands of lines of code that look plausibly correct in seconds.
What are humans bad at? Reviewing changes containing thousands of lines of plausibly correct code.
This is a great way to force senior devs to take the blame for things. But, if they actually want to avoid outages rather than just assign blame to them, they’ll need to submit small, efficient changes that the submitter understands and can explain clearly. Wouldn’t it be simpler just to say “No AI”?
AI’s greatest feature in the eyes of the Epstein class is the ability to shift responsibility. People will do all kinds of fucked up shit if they can shift the blame to someone else, and AI is the perfect bag holder.
Just ask the school of little girls in Iran which were likely targets picked by AI with out of date information about it being a barracks. Why bother confirming the target with current intel from the ground when no one’s going to take the blame anyway?
If you ask a writer what is Ai good for? They will say it’s good for art. But never use it for writing, because it’s terrible at it.
If you ask a artist what is Ai good for? They will say it’s good for writing. but never use it for art, because it’s terrible at it.
Conclusion… it’s good at neither… or am I missing your point?
The output looks good to people who are poorly versed in the segment for which AI is being asked to perform, but often inefficient or fails in ways that an expert in the field would never miss.
—ignore this part, I’m just rambling from here on Depending on the context, you’ll almost certainly get something that looks correct on first glance, especially if you’re not an expert. If you’re an expert, you wouldn’t need to ask for such a task and, if you did to save time, you’d probably end up adjusting, correcting, or fixing several things to produce a production-ready output. I use it regularly for code because the last language I had any training in proper syntax was Fortran 77. And eventually the simple tasks I ask it to code for me work. I’ve asked it to do some excel calculations (I’m not an excel expert, I do a lot of mathematic manipulation in custom sheets) and some of them work, but most are either wildly convoluted or relay on obscure calls/functions rather than simply using standard logic and mathematic operations which are easy to edit and change. I’ve also asked it to do some graphical illustration (because I’m not a graphic artist) and it has produced nice looking illustrations with zero basis in reality - i.e. “draw me an outline of Scotland in the style you’d see on a tourist map and label, with a star, these four cities”. It produced what I would expect an average American would estimate the outline of Scotland looked like and was equally as accurate with the location of the four cities (i.e. utterly incorrect).
Or I suppose add extra work by walking an AI tool through making small incremental changes.
In my experience, LLMs suck at making smart, small changes. To know how to do that they need to “understand” the entire codebase, and that’s expensive.
Yeah that’s what I mean by extra work. I can make the change myself or I can argue with claude code until it does what I want.
How in the glorious fuck was this not a thing from the start? In a system this big and this critical all code should be reviewed by cognizant individuals. Anyone who thought an LLM would be perfect and not need code reviews has their heads so far up their asses they can see through their pee hole.
If you do this, you signal the AI isn’t ready for production capabilities, which limits your sales groups capability to market it. Which is in reality the actual case and AI sucks and should never be trusted.
Yes, so now when there’s a success, it gets attributed to AI. When there’s an outage, that’s the fault of humans not reviewing correctly. These senior engineers will get fucked in all scenarios.
Precisely. From Cory Doctorow’s latest, very insightful essay on AI, where he talks about the promise of AI replacing 9 out of 10 radiologists:
“if the AI misses a tumor, this will be the human radiologist’s fault, because they are the ‘human in the loop.’ It’s their signature on the diagnosis.”
This is a reverse centaur, and it’s a specific kind of reverse-centaur: it’s what Dan Davies calls an “accountability sink.” The radiologist’s job isn’t really to oversee the AI’s work, it’s to take the blame for the AI’s mistakes.
I don’t think it’s fair to compare LLM code generation to machine vision in this way. These are very different "AI"s. Not necessarily disagreeing with Doctorow, but this is an important distinction.
How the machines work does not matter. The situation is using a machine to replace human expertise while ensuring a human still takes responsibility for things that human is not responsible for. It is not the owning class who is at risk for their machines mistakes, it is the owning classes wage slaves who are at risk.
My understanding is that the tumor detecting machine vision is generally thought useful in addition to the radiologist’s expertise. It basically outputs “yes”, “maybe”, and “no”, which is more expertise respecting than generating somewhere thereabouts code, which the coder has to (now) validate.
This is why I wouldn’t equate these tools. LLM code generation is marketed to do much more than machine vision for tumor detection.
Cory Doctorow actually goes more in depth on the radiologist example in a post from last year:
'If my Kaiser hospital bought some AI radiology tools and told its radiologists: “Hey folks, here’s the deal. Today, you’re processing about 100 x-rays per day. From now on, we’re going to get an instantaneous second opinion from the AI, and if the AI thinks you’ve missed a tumor, we want you to go back and have another look, even if that means you’re only processing 98 x-rays per day. That’s fine, we just care about finding all those tumors.”
If that’s what they said, I’d be delighted. But no one is investing hundreds of billions in AI companies because they think AI will make radiology more expensive, not even if that also makes radiology more accurate. The market’s bet on AI is that an AI salesman will visit the CEO of Kaiser and make this pitch: "Look, you fire 9/10s of your radiologists, saving $20m/year, you give us $10m/year, and you net $10m/year, and the remaining radiologists’ job will be to oversee the diagnoses the AI makes at superhuman speed, and somehow remain vigilant as they do so, despite the fact that the AI is usually right, except when it’s catastrophically wrong.
“And if the AI misses a tumor, this will be the human radiologist’s fault, because they are the ‘human in the loop.’ It’s their signature on the diagnosis.”
This is a reverse centaur, and it’s a specific kind of reverse-centaur: it’s what Dan Davies calls an “accountability sink.” The radiologist’s job isn’t really to oversee the AI’s work, it’s to take the blame for the AI’s mistakes.’
In short, we definitely could (and indeed should) be using tools like tumor detecting machine vision as something that helps humans build a better world for humans. But we’ve seen time and time again, across countless fields that it never works out that way.
That’s because this isn’t a problem with the technology of AI, but the fucked up sociotechnical and economic systems that govern how this tech is used, who gets to use it, who it gets used on, whose consent is required for those uses and most significant of all: who gets to profit?
!Not us, that’s for sure!<
The kind of AI doesn’t matter with this situation. Hell, It could be a magic talking rock™ and it change nothing of Mismanagement using a person to avoid blaming their shiny and expensive new toy.
“this is an important distinction”
it really isn’t
If my job ends up being reviewing AI code spammed at me by vibe coding juniors all day, I’m joining a nunnery.
If nunneries are as gay as I always imagined in my head, I’m in.
or hear me out, they can build it themselves so they don’t have to chase hallucinations. as a matter of fact, let’s cut the ai out of the project and leave it to summarizing emails.
This 1000x. You think that senior dev got to that level hoping one day all they’d have to do is evaluate randomly generated code? No! They want to create, build, design, integrate, share. Cut out the middle, useless step and get back to the work these professionals have dedicated their careers to.
AI is an assistant, not a replacement. It amazes me that Amazon, Microsoft, Google, and all these “tech leader” companies are going to make the same tech fuckup multiple times.
If only the lessons were painful for them and not just us/the workers.
Wonder what the turnover rate in executives is. I bet it is about 8 years.
“Everyone must use AI.”
…
“No! Not like that!”
Junior and mid-level engineers will now require more senior engineers to sign off any AI-assisted changes, Treadwell added.
So instead of getting a human to write it and AI peer reviewing it you want the most expensive per hour developers to look at stuff a human didn’t write and the other engineers can’t explain? Yeah, this is where the efficiency gains disappear.
I read stuff from one of my Jr’s all the time and most of it is made with AI. I don’t understand most of it and neither does the Dev. He keeps saying how much he’s learned from AI but peer programming with him is the pits. I try to say stuff like, “Oops! Looks like we forgot the packages.” And then 10 secs of silence later, “So you can go to line 24 and type…”
Just to add to this:
- When a senior dev reviews code from a more junior dev and gives feedback the more junior person (generally) learns from it.
- When a senior dev reviews code from an AI, the AI does not learn from it.
So beyond the first order effects you pointed out - the using of more time from more experience and hence expensive people - there is a second order effect due of loss of improvement in the making of code which is both persistent and cumulative with time: every review and feedback of the code from a junior dev reduces forever the future need for that, whilst every review and feedback of the code from an AI has no impact at all in need for it in the future.
Given enough time, the total time wasted in reviews and feedback for code from junior devs is limited - because they eventually learn enough not to do such mistakes - but the total time wasted in reviews and feedback for code from an AI is unlimited - because it will never improve.
Seniors reviewing code is fine but only when, as someone else mentioned, the code writer is learning from the review. The AI doesn’t learn at all and the Jr Dev probably learns very little because they didn’t understand the original code. Reviewing AI code often turns into me rewriting most of it.
Exactly.
The best way to learn is to have done the work yourself with all the mistakes that come from not knowing certain things, having wrong expectations or forgetting to account for certain situations, and then get feedback on your mistakes, especially if those giving the feedback know enough to understand the reasons behind the mistakes of the other person.
Another good way to learn is by looking through good quality work from somebody else, though it’s much less effective.
I suspect that getting feedback on work of “somebody” else (the AI) which isn’t even especially good, yields very little learning.
So linking back to my previous post, even though the AI process wastes a lot of time from a more senior person, not only will the AI (which did most of the implementation) not learn at all, but the junior dev that’s supposed to oversee and correct the AI will learn very little thus will improve very little. Meanwhile with the process that did not involve an AI, the same senior dev time expenditure will have taught the junior dev a lot more and since that’s the person doing most of the work yielded a lot more improvement next time around, reducing future expenditure of senior dev time.
Lol I would be your Jr, except instead of 10 seconds of silence it would be 10 seconds of me frantically clacking on the keyboard “add a block to this for these packages with proper syntax, I forgot to include it” to claude. Then I’d of course be all discombobulated and shit so I wouldn’t even bother to open code, I’d just ctrl-c about 100 lines somewhere around the general area of where I think the new code block should go, then ctrl-v the whole thing into the chat box because why not the company is paying out the dick for these tokens so might as well use them.
And two weeks later half our website crashes which results in you having to go to a meeting where management tells you to keep a closer eye on me. Which is basically what you had been already doing before AI but now you get to babysit me and claude!
It’ll be temporary, a gut reaction to add more experienced engineers in the loop. These folks will try to codify and then push better checks/guardrails into CI/CD and tooling to save themselves time. Given how new this all is, it’s almost the blind leading the blind though.
Amazon might also have some poor system boundaries, leading to non-critical systems/code impacting critical systems. Or they just let junior devs with their AI tools run wild on critical components without adequate guardrails… also likely. :-P
I read stuff from one of my Jr’s all the time and most of it is made with AI. I don’t understand most of it and neither does the Dev. He keeps saying how much he’s learned from AI but peer programming with him is the pits. I try to say stuff like, “Oops! Looks like we forgot the packages.” And then 10 secs of silence later, “So you can go to line 24 and type…”
So what kind of code is that? Code lyoko? Are they using more advanced code than their training should make one think?
It’s usually fine code but it just doesn’t follow the same conventions and flow. It’s kind of like reading a novel typed in block letters written in 3rd person then suddenly it’s cursive letters and 1st person.
I always saw a code review like a dissertation defense. Why did you choose to implement the requirement in this way? Answers like ‘I found a post on Stackoverflow’ or ‘the AI told me to’ would only move the question back one step; why did you choose to accept this answer?
I was a very unpopular reviewer.Likely, but you did not let poor code pass. That is valuable.
Well that’s going to make your senior developers quit.
Exactly. If you’re too stupid or lazy to adequately vet what your LLM puts out yourself, it shouldn’t be somebody else’s job to wade through the sewage you’re producing. You either shouldn’t be using one or, if you can’t do your job without it, you shouldn’t have that job.
—Someone who doesn’t use genAI but has spent way too much time digging through LLM slop
You know what my favorite pizza topping is? Bleach.
Dominoes REFUSES to put bleach on my pizza, so I gotta do it myself. I found out about it from AI. Now my pizza tastes great! The downside is having to go to the hospital to get a stomach pump everytime.
If you use antifreeze, it will make it sweet. Perfect for pineapple pizza!
Bleach Boys - https://open.spotify.com/track/0o6zZmPn5a3FJMNjINjZIB
Here to save you from the comments about spotify.
People: Just take this Link. You do not need to comment how much you hate spotify
Thanks, that’s a really cool site! Echoing the other comment about no YouTube, and also I thought it was hilarious that Spotify was the first option.
I was going to praise you, the link, the site, the everything. Fucking great idea. Someone build a bot. But no You Tube Music. :( Big, evil company - yes. But music and NO fucking adds makes it worth every penny, cent and rupie.
I hate spotify
I mean honestly yeah, I’m not going to waste my time with some junior developer who can’t explain how the code works and how it interacts with whatever framework I’m working on. I ain’t got time for that nonsense, especially when the code I deal with involves safety critical sections of code.
Honestly if my work ever decided to allow unfettered AI code generation into my code base, I would immediately look for a new job at that point.
I am not a developer, but:
I told the owner of the company recently that, and I quote, “I will fucking kill myself if my job becomes reviewing AI output”
Never kill yourself for something that’s somebody else’s fault.
It’s going to make snr devs get fired, surely?
They either refuse to sign off when boss wants them to and get fired or sign off and get fired when ai code they signed off on causes issues.
Or quit/find new jobs. I suspect that’s by design by Business Idiots.
*Get rid of the most expensive engineers and the cheaper ones can just use AI to make up the difference in output. And we can make the lower engineers the fall guy when convenient and replace them at our leisure *
The disdain bosses have for average people is astonishing.
Bingo.
Maybe not outright fired, but absolutely open them up to career limits based on what you described.
All of Amazon’s code undergoes code reviews already. Accepting a PR is already spiritually a sign off.
This is just explicitly a threat, explicitly trying to find someone to hold accountable because you can’t hold ai accountable. What are they gonna do, fire the ai? Sign here to be the fall guy. Fuck off.
Senior devs are expensive.
- If they don’t quit, they are an accountability sink.
- If they do quit, amazon saves money on labor arbitrage.
Win, win! Brilliant! For amazon.
What are they gonna do, fire the ai?
Well… yes?
Hold the Fucking C-suites accountable
How can they fire them ?
They’re not employing them to look pretty you know.
Easy, it’s next quarter’s problem.
it’s pretty fucking stark right? these are the devs that stayed after management mandated they USE the shit in the first place, now they want the same devs to become responsible for what the shit does to their codebases.
You’re absolutely right!
“You are now a senior auditor.”
Couldn’t they, I don’t know, just go back to people writing the code, and stop using AI to do something it clearly can’t handle? Just an idea.
I guess they’ve invested (thrown) so much money at this thing, they’re determined to make it work. Also, I know they’ve gone into insanely deep debt and if it doesn’t work they’re going to lose an eye watering amount of money, and perhaps the bubble bursting will be the catalyst to bringing down the entire world economy.
Oh, so yeah, they do have great incentive to make this work, but I don’t see it happening. As usual, they fuck up and the rest of us pay the bill. None of the billionaires will suffer any more than loss of face over this. Even if they’ve broken laws, all they ever get is a small fine and a slap on the back, “Better luck, next time, ol’ boy!”
…
Do the senior engineers NOT sign off on changes to systems that can take down the production servers? Even if we take out the LLM created code, this sounds like a bigger problem
We may start to see people realize that “have the AI generate slop, humans will catch the mistakes” actually is different from “have humans generate robust code.”
Not only that, but writing code is so much easier than understanding code you didn’t write. Seems like either you need to be able to trust the AI code, or you’re probably better of writing it yourself. Maybe there’s some simple yet tedious stuff, but it has to be simple enough to understand and verify faster than you could write it. Or maybe run code through AI to check for bugs and check out any bugs it finds…
I definitely have trusted AI to write miniature pointless little projects - like a little PHP page that loaded music for the current directory and showed a simple JS player in a webpage so I could share Christmas music with my family and friends. No database, no file uploading or anything. It worked decently, although not perfectly, and that’s all it needed to do.
This is true not just with code, but with many types of complex outputs. Going through and fixing somebody’s horrible excel model is much worse than building a good one yourself. And if the quality is really bad, it’s also just faster to do it yourself from scratch.
I’ve been writing a slightly larger project with frontend, bff and backend and I need to take it in small batches so that I can catch when it misunderstands or outright does a piss job of implementing something. I’ve been focusing a lot on getting all the unit tests I need in place which makes me feel a bunch better.
The bigger and more complex the projects get, the harder it is for the LLM to keep stuff in context which means I’ll have to improve my chunking out smaller scoped implementations or start writing code myself I think.
All in all I feel pretty safe with my project and pleased with the agents work but I need to increase testing further before bringing anything live.
Security testing will be the most important.
I’ve done a couple of tiny projects that I didn’t feel like coding. So far, I have not been terribly impressed. Well, it is impressive that it can make something functional at all, and in one case, what it made was fine enough to use as the temporary project it was intended (sharing christmas music with friends/family - reading files from a directory and writing a javascript player to play them in a shuffled order).
In the other case, replicating a simple text-based old DOS game with simple rules (think a space-based game around the complexity of checkers or so), it failed to think of so many things that while it did what I told it for the most part, it wasn’t a playable game. It was close, and fun enough for a nostalgic moment, but I had to work with it on logic like “If two fleets of ships arrive at the same planet in the same turn, you have to see how the first battle goes. If the first battle captures the planet, the second fleet is not attacking the first fleet’s ships - we won the planet at that point”. Very simple concepts that sure, you’d have to think of as a programmer, but if you were telling another person about how the game should work, were things I felt another person would think about.
I hope AI works well for you. Anywhere security it needed like database sanitation or user credentials… I hope you test thoroughly and I hope you can tell it enough to remind it to implement things like sanitation and other safety measures. An app can certainly appear to be working, but give many many fronts for attack. That’s my main worry with AI code. I worry enough on the little projects I do if I’m being secure enough myself.
Yeah I hope I am cautious enough. I use strict db models that were man written and have type checking and sanitation. That along with unit tests that cover everything I’ve been able to think of that can go right or wrong combined with the classic “obscurity===security” motto.
Of course there are always vectors one hasn’t thought of, but that goes for man made projects as well. If I decide to bring it live and scale up I’ll probably order a pen test.
Sounds like you’re 1) thinking about it and certainly 2) doing way the fuck more than most utilizing AI.
My approval means quite little, but you have it anyway <3
❤️
Yeah, initially writing the code never was the time sink.
*LLM
I guarantee there’s so much pressure on those engineers to deliver code that they rubber stamp a ton of it with the intention of “fixing it later”
Source: I’ve worked in software for 20+ years and know a lot of folks working for and who have worked for Amazon
That’s basically the story at all the big tech companies, from what I’ve heard. In my time at Facebook, I felt like the only person who actually read the merge requests that people sent me before hitting it with “LGTM”
If companies are going to place increasing reliance on review due to having lower-quality submissions, then they should probably evaluate employees weighting review quality (say, oh, rate of bugs subsequently discovered in reviewed commits or something like that).
Sure. i’ll review your code favourably if you do the same with mine.
That is also a way to get no bugs at all.
When I worked there 20% of the work we had to do had to go through a senior engineer. And getting his time was like pulling teeth.
More of the time he would just nitpick grammar in docs and then finally rubber stamp work. It was awful.
the way private companies work is that they require their employees to produce more than is reasonable given the work quality that is expected.
when this discrepancy is pointed out, it’s handwaved away. when the discrepancy results in problems, as it most obviously will, somebody is found to place the blame on.
it’s not the developer’s faults. it’s a management decision.
source: I’m talking out of my ass I’m just a salty employee who is seeing this happen at their own workplace when it didn’t used to, at least not to this level
Morged.
