I'm not at all close to knowing their internals with books in particular but for most products Amazon has made a significant effort to remove themselves from direct product stocking wherever possible (instead preferring to warehouse other people's stock for efficient shipping). I would guess that they aren't particularly well positioned with publishers at this point... Also, most publishers probably loathe Amazon at this point for all the times Amazon has tried to fleece them w.r.t. Kindle.
This is what I've heard from my friends both there and at Meta. It definitely doesn't carry the prestige it once did. It's more like, "willing to get yelled at, be afraid of being put on a PIP every day, and trying to ensure output outpaces your peers". That's sad, but it's where the industry is at so many places right now.
Amazon offers good signing bonuses that pay out over two years, so it's very common to peace out right at the 2-year mark when your take-home pay abruptly goes way down and the grind is no longer worth it. That's what I did, and I noticed it almost everywhere: there's a constant revolving door of inexperienced people. I don't understand how they keep it going with so much turnover, but somehow they've built a system that kinda withstands it, except in frontier AI apparently.
If you chart money in your pocket per unit time, it takes a dip around 18 months where your average pay is less than your first month’s pay. (When I was there, using RSU numbers for that time.) Guess when average quitting time was?
HR swore it wasn’t on purpose and I was somehow measuring pay “wrong” by looking at money in my pocket over time; but knowing their HR, I’m sure the valleys in transition from signing bonus to vesting are intentional.
This is where Bayes bites in practice: we can know that impatient and long-minded behaviors are correlated to outcomes, without the outcome meaningfully informing which case we have.
That’s because people are complicated and do things for many reasons, so people quitting when we cut their pay might not tell us they’re impatient.
I suspect the intent is what you say, but I don’t think the outcome is.
Because all humans are 100% fungible with all other humans. So, you can take whatever old humans you have, throw them away (or make them want to leave), and then replace them with all new humans, with zero negative effect on anything. Rinse and repeat.
Human experience means nothing. Anything that does have meaning gets encoded in the system for the mindless humans to implement.
That’s my take on the view from the top, based on my time working at AWS.
Not an employee, just have a "shit ton" of Amazon stock that has been beating inflation very well. The last few quarters of volatility has allowed me to buy and sell, often fully covering my monthly expenses. I'm very greatful for that. Far from a company "nobody" wants to work for. You have a better performing company than Amazon to recommend that I can use instead? I'm all ears and I'll even give you a kickback for your time!
Your comment was likely focused on dysfunctional parts of the culture there (and that seems to be how most replies to you interpreted it as well). There's truth to that, sure. But I think respect for data sensitivity and security is also a big part of what makes the culture ill-suited for frontier AI lab work.
When I was at AWS, we had data access locked down very tightly. I was trying to train a model to predict how long various dynamic workloads running on behalf of customers would take. Just getting access to anonymized production data from a DB that my team owned took well over a month and had me compromise heavily on what attributes I could access. I also spent a decent amount of time in meetings with lawyers and compliance folks to explain and justify all our data classifications for PII, retention, access controls, etc. We would review our data policies quite frequently. We needed break-glass procedures to access prod data in our own systems, and it was uncommon to need to resort to that. Even when you broke glass, certain classes of sensitive data would still not be visible to you. On the plus side, We did get quite good at operating and maintaining systems without needing to see sensitive info.
If you contrast that with Meta's disregard for privacy or the way frontier labs disregard the law, it's easy to see why Amazon would struggle in this space.
I read your comment as “amazon acts too responsibly to be a frontier lab,” and that worries me. That they even bothered to do all that stuff is good. That others don’t is bad.
I did indeed mean it that way, but I agree with you. It sucks, but it's just the unfortunate reality of the world. Many of the most disruptive and successful startups made it there by ignoring legal or moral limitations.
Frankly reading your comments was a relief. I'd rather they take my privacy (and security) seriously. I don't need or want any AI features or technology from Amazon, thanks very much.
Amazon has a history of fostering off-shoot companies that are given more freedom to innovate[1][2]. Keeping AI development in-house under Bedrock was probably their mistake. There's a big difference between running AI and developing new AI models.
Incentives can produce locally rational and globally harmful behavior. (See also: "Moloch", although I'm less of a fan of the concept lately since it focuses exclusively on the harmful aspects of emergent systems, deleting the neutral and positive ones in the process.)
A relevant metaphor is the nuclear arms race. Nobody wants a world full of nukes pointed at each other, but if the other guy is gonna do it...
The argument in this case is, we can't slow down, because China certainly won't, and do you want those guys to build the machine God first? So it's "rational" to maximize p(accidental_skynet).
> That they even bothered to do all that stuff is good. That others don’t is bad
good and bad isn't what Meta and AWS is running with. If AWS trains willy nilly on user data like Meta does, AWS is toast. AWS is already taking a huge risk by now using their own take of AWS FDEs competing against their established ecosystem of consulting companies that used to recommend all 583 of AWS products to host a static webpage. Turning of more customers by training on their data is too much of a risk. That's the "good" - a risk analysis.
Meta doesn't care if you stop using Facebook, Whatsapp etc. Frankly, if you are embedded in the Facebook, Whatsapp ecosystem, you don't have a way out as a normie. Even if you are technically capable, good luck publishing your custom Mattermost build for iOS.
Meta has a significant amount of investment in hardware, originally due to Metaverse. The way to monetize that investment is inference and training. That's a do or die decision for Meta. Meta isn't releasing open-weight models because they are "good".
Meta is releasing open-weight models for the same reason Alibaba, NVIDIA or IBM does. Or Google releases Gemma.
There's no "good" or "bad" - the moment Alibaba, NVIDIA, IBM or Google's objective functions change, "Do no evil" will quickly morph into "Do as much evil as necessary without the costs of getting caught exceeding E(outcome)"
Their core selling point is customer trust that their product works as advertised.
From my interactions with them, is that they really care about customer success, and everything including AI is a means to an end to them.
I think building a cool product with undefined reliability and capabilities, then selling fantasies about what it will possibly be able to do down the line isn't really what they do - or what their customers are paying for.
I agree, from my personal experience. AWS support at most tiers I’ve used is excellent. Granted, the larger account you have, the friendlier the service, but even when using it for smaller personal projects, I’ve always been pleased with the level of support and devotion to customer experience (except the console experience, that is not fun).
Don’t know to comment about the other areas of their business but AWS has always been excellent for me.
I interned there for a summer working on evaluating text generation and despite the fact that I used no consumer data whatsoever, but rather a well-known public research dataset, I also couldn't egress from the machines I ran training on. Not even reports on the training (admittedly, this was 2021 when they had just introduced this new system). So it was even more of a damper on research progress than you might imagine.
That explains a lot. They are anti “move fast and break things” which is great news for enterprises or anyone seeking stability but terrible for innovation.
I don’t condone the use of pirated content for training. But anything that was bought is above the board as far as I’m concerned. Can let the lawyers fight out the minutia and inevitable licensing specifics after the fact (book publishers are hawks).
Fun fact, did you know that every library in the world has an independent contract with every major book publisher. In some cases I am familiar with, they stipulated that a physical book can only be lent X times. After that, the library much purchase a new copy (even if the original book is in pristine condition). Yeah.
this. the culture that makes AWS the gold standard in cloud security (yes, it is still the best in the business despite the one-off embarrassing CapitalOne incident) with a strong emphasis on locking down data, minimizing PII, and saying “no” by default, is the very opposite of what building frontier LLMs is all about, which is really around hoovering up vast amounts of other peoples data and moving fast before the rules catch up.
What you are citing appears to me as a business use case of classical machine learning models not necessarily LLMs, and demand forecasting. Frontier AI is clearly not dependent on such data, and has many outside data sources and I'm even sure sensitive AWS data would be of good use as training data, not to mention the legal/contractual obligations AWS has in protecting that data, what am I missing?
That was just an example anecdote to highlight the culture around respecting privacy and security. Obviously AWS would never allow for training on customer data, but even for metadata or data of our own internal systems, things are super strict.
Public datasets that are liberal with licensed access is one thing, but the best frontier labs clearly went well beyond that.
Was s3 the last time Amazon showed lab grade innovation? Lambda?
I could be wrong because I haven't paid much attention but Amazon has just fast followed everyone else innovation for a awhile so it's in like one ecosystem yeah?
I'm not trying to be snarky, I just haven't thought about aws as having any innovation for a while and I always figured it was because ruthless companies that churn through people as disposable commodities generally don't do tech innovation.
I'm talking on the AWS side, on the logistics side they clearly are doing insane things on supply chain perf optimization.
> I could be wrong because I haven't paid much attention but Amazon has just fast followed everyone else innovation for a awhile so it's in like one ecosystem yeah?
A lot of innovation at AWS is invisible on the outside. They pioneered hyperscale datacenters with publicly accessible capacity, software-defined networks, secure VMs, and so on.
> I'm not trying to be snarky, I just haven't thought about aws as having any innovation for a while and I always figured it was because ruthless companies that churn through people as disposable commodities generally don't do tech innovation.
I worked at AWS before COVID. It was really a varied experience, but it was not _bad_ overall. We prided ourselves on making sure our APIs are solid, stable, and backwards compatible. When we had to make a backwards-incompatible change, we actually contacted every affected customer and ensured that their workflows wouldn't be broken.
Oh for sure, these many years AWS has set a high bar in the service space for conscientiousness in offerings. When I compare to my personal work experiences with gcp and azure the amount of times I've been rug pulled over the years has been essentially zero for AWS.
Dunno about that, but my experience with AWS is their stuff generally works, the other big 2 doesn't.
Generally speaking, on AWS you can buy compute, storage, network and other basic 'goods' and just like a supermarket where you can get either flour and eggs, a cake mix or a ready-made cake, AWS is willing to serve it to you in various levels of ready-ness, and you can start at the level you want, and build your own layers on top if you don't like AWS's solutions.
And AWS's solutions are generally built like that as well, there's no magic to Lambda or Fargate or EKS, its just they run the same EC2s, but with some proprietary magic that lifts some complexity from your shoulders, at the cost of generally some markup, and loss of some otherwise desirable features.
But the core thing is that AWS's offerings are generally solid, so as AI evolves and companies become more willing to take on more complexity, they will move to lower level services, but not necessarily out of AWS itself.
their innovation seems to be limited to renaming the forced sidebar on amazon.com every few months and not allowing me to dismiss their useless features.
Ublock hides the sidebar from me, but it does still take up the whitespace on the side of the screen. Now it won't remember that I dismissed it and it pops back in on every new screen. I am just used to the screen being narrower now. But it is still annoying.
I think the forced-chat price history thing is a dirty way of trying to get people used to talking to the chatbot. Can't think of any other reason it would be designed so stupidly
The frontier of AI has gotten so spicy that the delivery risk of pursuing this rabbit is astronomical. Knowing when to quit and use someone else's solution is critical at this scale.
If someone started from zero, how long do we think it would take for them to reach the cost/capability front currently provided by gpt-6.1-terra? How long until OAI figures out a way to distill that down again for pennies on the dollar?
Apple seems to have this sorted. You won't find them chasing OAI, Google and Anthropic around the playground.
It doesn't make too much sense for Amazon to invest too much of their own money into AI.
They already sell a lot of the shovels for training and inference and make ample money from that, and there are enough open weight models on the market to make it infeasible to build a large enough moat.
The remainder of your quote matters more, I think:
> Instead, labor and computing power are being diverted toward a singular “frontier-model effort” led by Pieter Abbeel, director of the Berkeley Robot Learning Lab whose robotics company Covariant was hoovered up by Amazon in 2024. At the time Abbeel joined the tech giant, Covariant’s founders were working on AI models for robots, which lines up pretty well with Amazon’s push to automate everything it possibly can.
Amazon can't hope to compete in the LLM genAI (what is colloquially known as AI anyway) space. That is toast, the race is over for them, too much competition, no way to make money.
But industrial automation? Globally, Amazon employs about 1M people in fulfillment/operations roles, the utter majority of their headcount. If there is one company on this world that has the money and compute capacity to spare for a high performance, real world application of fusion between modern machine learning and robotics, it's Amazon - and to make life easier for Amazon, unlike every other competitor they don't have to pay or partner with another company for real world testing. They can just go and deploy a robot in their warehouse if they want to.
Achieve a breakthrough in a fully automated picker robot, and they can save half a million people on their payroll, if not more, they'd save shrink loss and handling mishaps as well. That is one hell of an opportunity to save money by R&D. And if they patent everything they do, they can lock their competitors out of that technology for a loooong time.
Well except for the regular HN anti-AI crowd. The same people that have wished for communism and never ever looked in detail into the 20+ states that failed that way.
Values drive behavior. Who owns / runs futurism and what are their values?
Journalism / media as activity needs to be sustainable and unfortunately in our current world it's seems very difficult to sustainably remain high integrity. Sometimes it may be greed, but most of the time I think it's survival.
In this case AI is part of what is threatening journalist survivability, so it kind of makes sense.
It's possible for both a new technology to be hugely beneficial in the long run & for early investors to lose a lot of money funding it (if they over-invested / over-paid).
I wonder if the best play now is not to play and wait until the landscape is clear. They are making partnerships now (e.g OpenAI), and investments (e.g. Anthropic).
It would seem that they can experience most of the benefit of AI without most of the cost; anthropic is operating at the frontier and shows that even while supplying the best models and token sales to enterprise, they lose massive amounts of money.
Amazon makes 60B+ in profit per year, they don't need to tank that by losing it on training AI / selling tokens.
There are plenty of open weight models out there for it to use with most of the benefit and very little of the downside.
Wait so Microsoft gutted their AI Division and Microsoft looks like it did as well from the list of "departments" that are in the approved buyout list. Does this mean they are just going to be resellers of the big models companies going forward?
Amazon Nova is their model? I worked for a company that was being pushed into trying to use AWS's "Nova Act" [0], which I guess was their python-based wrapper to do things in Playwright. I didn't think it used Nova the model under the hood though, or maybe it did. Either way, Nova Act was horrible in terms of performance - it was so...damn...slooooow. Much easier to just use Playwright.
Either you are getting the smartest engineers, or offer the most freedom to pursue interesting goals that aren't on the frontier. Amazon has neither so they naturally aren't going anywhere in that domain.
Can you name 3 ways is positively impacted the average non tech persons life in a meaningful way? Meaningful would be more than the invention of the microwave, but not as important as a vaccine.
We're at the dot com boom stage of AI. You could say it didn't deliver because boo.com didn't make a profit or because the average person still used paper maps, but that was just because of momentum.
It's the same with AI. The technology is here now. My friends in financial advice and market research have not lost their jobs immediately, because of momentum.
- File taxes and find issues that "accredited", "certified" entities missed. In one instance, my federal refund increased by nearly $5k. At $20/mo, just the refund is two decades worth of value.
- Appeal health, home and car insurance claims successfully
- Find out details of interesting plants without forking out tens of dollars for an app that claims to do the same but fails to do so most of the time
- Summarize never ending email threads with 20 disinterested parties who would rather not be on the email but have to respond to show they are "engaged"
- Write a convincing email to my mother in law why I cannot meet her this weekend and hold her hand through the tough times she's having with her ethnic neighbor who annoys her with their "ethnic" music and "smell of fish" BBQ
- Locate the user manual for my microwave and what that rapid 3 beeps every 1s for 2 minutes that starts at 10AM since the last 2 days mean
GPT 3.5 is 2022. In the past ~18 months the majority of all human software development has been wiped out and the models keep getting better rapidly. AI has already radically overdelivered and the biggest returns are in the next 10, 20, 50 years.
There is definitely an unanswered question as to what the metric, but no, JavaScript is not 90+% by any useful metric. Most code is not a website, and JavaScript penetration outside of the client-side code on a website and some GUI applications on various operating systems is almost nonexistent.
Remember that all of the code in operation includes code like the microcontroller in your mouse, the hundreds of chips in your car doing things like keeping the engine running or "window goes up when I push the button," the DVD menu (which is actually Java code), not to mention all of the staid systems responsible for things like writing your paycheck every other week.
Models are finally good enough that I'm confident that I can repeatably build pretty much any software reliably in less than 2% of the time that I could before 2022. I wouldn't have made that claim 6 months ago.
Even quantized local models like Qwen3.8 are good enough to handle a lot of drudgery.
The caveat is that I developed software for 30 years, so I understand most of what it produces and can guide it to make fewer mistakes and fix what it can't do. The biggest challenge has been keeping my requirements focused - it's very easy to gold plate everything and prematurely optimize.
The biggest question I ask myself is why am I even still doing this. The smart people are probably already trying to figure out how to pipe raw matter and energy to their homes to build replicators.
The future is already here, it's just not evenly distributed.
Build experiences that are user dependent. The value isn't in the software it's in the users. That's the ultimate moat. It's what isn't properly grasped about GPT's position, insofar as they don't mess up big: they have captured a zillion users, they're Kleenex. Ask GPT = Google it.
You don't want to be in the CD selling or radio business with the Internet & digital & streaming & smartphones & iPods bearing down on you. Go for concerts. Notice that humans are spending more than ever on live concerts. Disney World and Disneyland still command extraordinary pricing power, despite everybody hiding inside on their phones.
LLM clones of eg Photoshop will end up being worthless, like ideas generally. The software will be worthless for the same reason ideas are: they're too easy. Users are hard, software will be (relatively) easy.
Is the value in McDonalds in the food? Everybody knows it's not. McD captured consumers, their heft is extraordinarily difficult to displace. Ie habits, brand awareness, location, etc.
"LLM clones of eg Photoshop will end up being worthless, like ideas generally. The software will be worthless for the same reason ideas are: they're too easy. Users are hard, software will be (relatively) easy."
I doubt that. If a clone out there offers all the Photoshop's features for free or much lower cost, without additional learning curve, then many professionals will switch to save money.
McDonalds would be gone too if somebody opened a restaurant chain serving similar food for free or 80% lower price. But that's impossible to do profitably, while freeware software has existed for a long, long time.
Photoshop users could switch to GIMP tomorrow. They don't want to.
The many professionals are the ones that could most easily understand how to switch to GIMP, which is free to use. They could save enormous amounts of money. They prefer to spend money and stick with Photoshop instead. That has persisted for a very long time now.
And in before ‘JumpCrissCross chimes in about inference being profitable while ignoring all the investment costs needed to get inference to where it is.
The full balance sheet captures the info. If the AI companies suck up all of our salaries and spend that money on R&D/training, they still aren’t profitable, right?
I mean, if anything summarizes where we have gotten to as an industry this quote is sure it. From Uber using VC funds to subsidize growth to where the AI players are at...
Why have they needed to worry about profitability?
They have been chasing extreme growth with cheap money. That enabled them to hyper scale at a pace the world has rarely seen before. They have made the correct choice so far.
Market capture didn't work so well for Uber, and it won't work for AI, and even less so for AI because you can't get a driver in China to take you to LAX airport, but you can use cheaper Chinese AI.
Uber has to pay back all the billions of dollars it wasted trying to get people to use their platform, and now I can't afford most Uber or Lyft rides, it's way too expensive. So did market capture really work if they had to raise their prices so much as to make the service unworkable for most people? The only time I take uber is if my company pays for it, and that is getting less often.
AI is also going to have to become way too expensive so they can pay back the billions (trillions?) it is borrowing. And, when everyone is laid-off or fired because AI took their job, then who's going to be paying for AI? I guess the companies that lay off all the humans will, but when the wealthy business owners realize they've replaced everything with AI and nobody is buying what they are selling because there's no more jobs, then what?
Sounds like they should worry about profitability, and a lot of other things too.
Uber isn't dealing with a shrinking market. They're becoming critical transportation infrastructure and their business keeps expanding year after year.
$32b -> $37b -> $44b -> $52b -> $57b
That's Uber's sales growth year by year. And now they're solidly profitable.
The sole threat to Uber is the Waymo / Tesla automation angle. A merger is most likely there.
Huh? uber is an amazing success story given their market dominance and enterprise value created. of course the next kid on the block (Waymo) is going to do the same thing
The price of a given point on the intelligence frontier is only going to get cheaper, likely 10-20x over the next few years.
but profitability is a very narrow view on the labs, and investors are pricing in future free cash flow
you need to take into account the derivative, what is their growth? their opex and capex? their revenue growth? their moat? New product lines and applications?
Not saying that it's a homerun, but there is a case to be made for the labs being wildly profitable in 4-10 years
I'm not sure if I'm more productive, but in my role where most of my deliverables are technical documents or emails I think my output quality has gone up. Now that I typically use LLMs as a research assistant and editor before sending anything significant to co-workers or customers I'm sure I make fewer mistakes. AI catches a lot of gaps and errors. That's a significant qualitative improvement although I can't quantify that in terms of profit impact.
I think you're letting the LLM write emails that are overly verbose that are shortened on the other side by another LLM and so there are 2 minus habens to both side of a communications that are communicating the same stuff as before but now are paying for altman's yacht in order to do it
Nah, that's not what I'm doing at all. Did you even read my comment? I use it more like a fact checker. It's not perfect but it catches a lot of my errors.
Maybe you could use an AI assistant to help you write better HN comments.
AI is not about productivity. It's about weaponizing the ability to ship fast go-to market features in the blink of an eye when it becomes necessary. If trends change fast and you can't adapt quickly with AI: you're dead. So you need to have AI ready to go, just in case.
We ship features faster and workers are more independent. They can get the answers they need faster than ever. And i'm talking about the business itself if the business has enabled that. There are a ton of productivity gains.
But make no mistake. Being more productive does not equal higher profits because your customers think you've added value. Both value and productivity are highly subjective terms which is why it is measured differently depending on where you work.
Most businesses adapted to AI by shipping a lot more features in smaller increments of time. What they should have done is become much more ambitious and build moon shots in two quarters instead. You need to go big or go home if you're going all in on AI.
i know multiple stealth startups with 9 figure revenues and profitable and growing
lab arr is climbing
given the recent crackdown on token spend and from speaking to execs they are very focused on the exact ROI they will get from AI
Note that business revenues in competitive markets may actually not change that much, because if everyone is using the same models there's no competitive advantage from using AI. but if one of them didn't use AI they would fall hopelessly behind
Too dangerous to release was 2019. Today it's only very useful for coding and other niche industries. It promised to automate all work and to usher in a utopia. Instead it helps programmers further enshittify Big Tech offerings. That's well short of any promise or expectation set out by boosters, paid and otherwise, for the last 7-ish years. Where's the civilization upending change? Where's the Beef(tm)?
Actually AI has been over promising and under delivering for about 70 years. Whether or not you think the juice is finally worth the squeeze, the distance we've come is impressive.
181 comments
[ 2.7 ms ] story [ 19.2 ms ] thread> Amazon Is Gutting Its AI Division [July]
HR swore it wasn’t on purpose and I was somehow measuring pay “wrong” by looking at money in my pocket over time; but knowing their HR, I’m sure the valleys in transition from signing bonus to vesting are intentional.
That’s because people are complicated and do things for many reasons, so people quitting when we cut their pay might not tell us they’re impatient.
I suspect the intent is what you say, but I don’t think the outcome is.
.. and then got let go at 22 months in favour of the AI budget.
Gotta respect the hustle.
Human experience means nothing. Anything that does have meaning gets encoded in the system for the mindless humans to implement.
That’s my take on the view from the top, based on my time working at AWS.
|----------------|------------------------------:|--------------------:|
| United States | 850,000 | 750,000–950,000 |
| India | 115,000 | 100,000–125,000 |
| United Kingdom | 68,000 | 55,000–75,000 |
| Canada | 40,000 | 30,000–45,000 |
| Germany | 36,000 | 30,000–40,000 |
| Spain | 27,000 | 23,000–30,000 |
| France | 23,000 | 19,000–25,000 |
| Italy | 17,000 | 14,000–19,000 |
| Poland | 17,000 | 14,000–19,000 |
| Japan | 13,000 | 11,000–16,000 |
| Mexico | 13,000 | 10,000–15,000 |
| Australia | 6,000 | 4,500–8,500 |
| Ireland | 6,000 | 5,500–6,500 |
| Singapore | 2,400 | 2,000–2,500 |
That's a lot of "nobody"
Not an employee, just have a "shit ton" of Amazon stock that has been beating inflation very well. The last few quarters of volatility has allowed me to buy and sell, often fully covering my monthly expenses. I'm very greatful for that. Far from a company "nobody" wants to work for. You have a better performing company than Amazon to recommend that I can use instead? I'm all ears and I'll even give you a kickback for your time!
When I was at AWS, we had data access locked down very tightly. I was trying to train a model to predict how long various dynamic workloads running on behalf of customers would take. Just getting access to anonymized production data from a DB that my team owned took well over a month and had me compromise heavily on what attributes I could access. I also spent a decent amount of time in meetings with lawyers and compliance folks to explain and justify all our data classifications for PII, retention, access controls, etc. We would review our data policies quite frequently. We needed break-glass procedures to access prod data in our own systems, and it was uncommon to need to resort to that. Even when you broke glass, certain classes of sensitive data would still not be visible to you. On the plus side, We did get quite good at operating and maintaining systems without needing to see sensitive info.
If you contrast that with Meta's disregard for privacy or the way frontier labs disregard the law, it's easy to see why Amazon would struggle in this space.
[1] https://en.wikipedia.org/wiki/Annapurna_Labs
[2] https://en.wikipedia.org/wiki/Amazon_Lab126
A relevant metaphor is the nuclear arms race. Nobody wants a world full of nukes pointed at each other, but if the other guy is gonna do it...
The argument in this case is, we can't slow down, because China certainly won't, and do you want those guys to build the machine God first? So it's "rational" to maximize p(accidental_skynet).
good and bad isn't what Meta and AWS is running with. If AWS trains willy nilly on user data like Meta does, AWS is toast. AWS is already taking a huge risk by now using their own take of AWS FDEs competing against their established ecosystem of consulting companies that used to recommend all 583 of AWS products to host a static webpage. Turning of more customers by training on their data is too much of a risk. That's the "good" - a risk analysis.
Meta doesn't care if you stop using Facebook, Whatsapp etc. Frankly, if you are embedded in the Facebook, Whatsapp ecosystem, you don't have a way out as a normie. Even if you are technically capable, good luck publishing your custom Mattermost build for iOS.
Meta has a significant amount of investment in hardware, originally due to Metaverse. The way to monetize that investment is inference and training. That's a do or die decision for Meta. Meta isn't releasing open-weight models because they are "good".
Meta is releasing open-weight models for the same reason Alibaba, NVIDIA or IBM does. Or Google releases Gemma.
There's no "good" or "bad" - the moment Alibaba, NVIDIA, IBM or Google's objective functions change, "Do no evil" will quickly morph into "Do as much evil as necessary without the costs of getting caught exceeding E(outcome)"
Metaverse was 4-5 years ago already wasn't it? That hardware is verging on obsolete.
From my interactions with them, is that they really care about customer success, and everything including AI is a means to an end to them.
I think building a cool product with undefined reliability and capabilities, then selling fantasies about what it will possibly be able to do down the line isn't really what they do - or what their customers are paying for.
Don’t know to comment about the other areas of their business but AWS has always been excellent for me.
I don’t condone the use of pirated content for training. But anything that was bought is above the board as far as I’m concerned. Can let the lawyers fight out the minutia and inevitable licensing specifics after the fact (book publishers are hawks).
Fun fact, did you know that every library in the world has an independent contract with every major book publisher. In some cases I am familiar with, they stipulated that a physical book can only be lent X times. After that, the library much purchase a new copy (even if the original book is in pristine condition). Yeah.
Public datasets that are liberal with licensed access is one thing, but the best frontier labs clearly went well beyond that.
They're a good data point on where the robotics labor frontier is.
like pee bottles.
I could be wrong because I haven't paid much attention but Amazon has just fast followed everyone else innovation for a awhile so it's in like one ecosystem yeah?
I'm not trying to be snarky, I just haven't thought about aws as having any innovation for a while and I always figured it was because ruthless companies that churn through people as disposable commodities generally don't do tech innovation.
I'm talking on the AWS side, on the logistics side they clearly are doing insane things on supply chain perf optimization.
A lot of innovation at AWS is invisible on the outside. They pioneered hyperscale datacenters with publicly accessible capacity, software-defined networks, secure VMs, and so on.
> I'm not trying to be snarky, I just haven't thought about aws as having any innovation for a while and I always figured it was because ruthless companies that churn through people as disposable commodities generally don't do tech innovation.
I worked at AWS before COVID. It was really a varied experience, but it was not _bad_ overall. We prided ourselves on making sure our APIs are solid, stable, and backwards compatible. When we had to make a backwards-incompatible change, we actually contacted every affected customer and ensured that their workflows wouldn't be broken.
Generally speaking, on AWS you can buy compute, storage, network and other basic 'goods' and just like a supermarket where you can get either flour and eggs, a cake mix or a ready-made cake, AWS is willing to serve it to you in various levels of ready-ness, and you can start at the level you want, and build your own layers on top if you don't like AWS's solutions.
And AWS's solutions are generally built like that as well, there's no magic to Lambda or Fargate or EKS, its just they run the same EC2s, but with some proprietary magic that lifts some complexity from your shoulders, at the cost of generally some markup, and loss of some otherwise desirable features.
But the core thing is that AWS's offerings are generally solid, so as AI evolves and companies become more willing to take on more complexity, they will move to lower level services, but not necessarily out of AWS itself.
Source: https://www.businessinsider.com/amazon-overhauls-ai-strategy...
Couple comments then:
https://news.ycombinator.com/item?id=49082344
If someone started from zero, how long do we think it would take for them to reach the cost/capability front currently provided by gpt-6.1-terra? How long until OAI figures out a way to distill that down again for pennies on the dollar?
Apple seems to have this sorted. You won't find them chasing OAI, Google and Anthropic around the playground.
https://blog.google/company-news/inside-google/company-annou...
They already sell a lot of the shovels for training and inference and make ample money from that, and there are enough open weight models on the market to make it infeasible to build a large enough moat.
> Instead, labor and computing power are being diverted toward a singular “frontier-model effort”
It’s a consolidation, not a withdrawal.
> Instead, labor and computing power are being diverted toward a singular “frontier-model effort” led by Pieter Abbeel, director of the Berkeley Robot Learning Lab whose robotics company Covariant was hoovered up by Amazon in 2024. At the time Abbeel joined the tech giant, Covariant’s founders were working on AI models for robots, which lines up pretty well with Amazon’s push to automate everything it possibly can.
Amazon can't hope to compete in the LLM genAI (what is colloquially known as AI anyway) space. That is toast, the race is over for them, too much competition, no way to make money.
But industrial automation? Globally, Amazon employs about 1M people in fulfillment/operations roles, the utter majority of their headcount. If there is one company on this world that has the money and compute capacity to spare for a high performance, real world application of fusion between modern machine learning and robotics, it's Amazon - and to make life easier for Amazon, unlike every other competitor they don't have to pay or partner with another company for real world testing. They can just go and deploy a robot in their warehouse if they want to.
Achieve a breakthrough in a fully automated picker robot, and they can save half a million people on their payroll, if not more, they'd save shrink loss and handling mishaps as well. That is one hell of an opportunity to save money by R&D. And if they patent everything they do, they can lock their competitors out of that technology for a loooong time.
https://news.ycombinator.com/item?id=49082344
This isn't a blip on anyone's radar.
COME ON, FLAG ME.
Latency is the #1 UX killer on voice apps, and the latency was terrible compared to Alexa. I used it for like 15 minutes and reverted.
All anyone was talking about was how much worse it was for common tasks.
It will always remain weird to me that a publication called Futurism is so outright negative towards our newest technologies.
Values drive behavior. Who owns / runs futurism and what are their values?
Journalism / media as activity needs to be sustainable and unfortunately in our current world it's seems very difficult to sustainably remain high integrity. Sometimes it may be greed, but most of the time I think it's survival.
In this case AI is part of what is threatening journalist survivability, so it kind of makes sense.
If true, Amazon is not really gutting its AI division.
“So… what happened to the previous team?”
This is something that requires an ethos completely different from how amazon works and the way jeff bezos sees the world.
Amazon makes 60B+ in profit per year, they don't need to tank that by losing it on training AI / selling tokens.
There are plenty of open weight models out there for it to use with most of the benefit and very little of the downside.
[0] https://nova.amazon.com/act
you want edge inference, from decision to small models.
That's it. It will be 2c, please.
It's the same with AI. The technology is here now. My friends in financial advice and market research have not lost their jobs immediately, because of momentum.
2. greatest hacking tool in history
3. greatest corporate espionage tool in history (stealing of Navier-Stokes solution)
4. greatest accountability transfer tool in history (whoops it was the AI that shot those missiles)
5. greatest capital raising tool in history (we use AI in our startup, and you should give us money)
Edit:Whoops, I didn't read "positively"
- File taxes and find issues that "accredited", "certified" entities missed. In one instance, my federal refund increased by nearly $5k. At $20/mo, just the refund is two decades worth of value.
- Appeal health, home and car insurance claims successfully
- Find out details of interesting plants without forking out tens of dollars for an app that claims to do the same but fails to do so most of the time
- Summarize never ending email threads with 20 disinterested parties who would rather not be on the email but have to respond to show they are "engaged"
- Write a convincing email to my mother in law why I cannot meet her this weekend and hold her hand through the tough times she's having with her ethnic neighbor who annoys her with their "ethnic" music and "smell of fish" BBQ
- Locate the user manual for my microwave and what that rapid 3 beeps every 1s for 2 minutes that starts at 10AM since the last 2 days mean
It's 1987: The PC is underdelivering.
It's 2000: the Internet is underdelivering.
It's 2010: smartphones are underdelivering.
What the hell does this even mean? Literally half of the code in operation I the world is still COBOL. The fuck is this nonsense wiped out
There’s no way this is even remotely true. JavaScript alone is probably 90+%, considering it’s on almost every website.
Remember that all of the code in operation includes code like the microcontroller in your mouse, the hundreds of chips in your car doing things like keeping the engine running or "window goes up when I push the button," the DVD menu (which is actually Java code), not to mention all of the staid systems responsible for things like writing your paycheck every other week.
Even quantized local models like Qwen3.8 are good enough to handle a lot of drudgery.
The caveat is that I developed software for 30 years, so I understand most of what it produces and can guide it to make fewer mistakes and fix what it can't do. The biggest challenge has been keeping my requirements focused - it's very easy to gold plate everything and prematurely optimize.
The biggest question I ask myself is why am I even still doing this. The smart people are probably already trying to figure out how to pipe raw matter and energy to their homes to build replicators.
The future is already here, it's just not evenly distributed.
You don't want to be in the CD selling or radio business with the Internet & digital & streaming & smartphones & iPods bearing down on you. Go for concerts. Notice that humans are spending more than ever on live concerts. Disney World and Disneyland still command extraordinary pricing power, despite everybody hiding inside on their phones.
LLM clones of eg Photoshop will end up being worthless, like ideas generally. The software will be worthless for the same reason ideas are: they're too easy. Users are hard, software will be (relatively) easy.
Is the value in McDonalds in the food? Everybody knows it's not. McD captured consumers, their heft is extraordinarily difficult to displace. Ie habits, brand awareness, location, etc.
I doubt that. If a clone out there offers all the Photoshop's features for free or much lower cost, without additional learning curve, then many professionals will switch to save money.
McDonalds would be gone too if somebody opened a restaurant chain serving similar food for free or 80% lower price. But that's impossible to do profitably, while freeware software has existed for a long, long time.
The many professionals are the ones that could most easily understand how to switch to GIMP, which is free to use. They could save enormous amounts of money. They prefer to spend money and stick with Photoshop instead. That has persisted for a very long time now.
And in before ‘JumpCrissCross chimes in about inference being profitable while ignoring all the investment costs needed to get inference to where it is.
I mean, if anything summarizes where we have gotten to as an industry this quote is sure it. From Uber using VC funds to subsidize growth to where the AI players are at...
They have been chasing extreme growth with cheap money. That enabled them to hyper scale at a pace the world has rarely seen before. They have made the correct choice so far.
Uber has to pay back all the billions of dollars it wasted trying to get people to use their platform, and now I can't afford most Uber or Lyft rides, it's way too expensive. So did market capture really work if they had to raise their prices so much as to make the service unworkable for most people? The only time I take uber is if my company pays for it, and that is getting less often.
AI is also going to have to become way too expensive so they can pay back the billions (trillions?) it is borrowing. And, when everyone is laid-off or fired because AI took their job, then who's going to be paying for AI? I guess the companies that lay off all the humans will, but when the wealthy business owners realize they've replaced everything with AI and nobody is buying what they are selling because there's no more jobs, then what?
Sounds like they should worry about profitability, and a lot of other things too.
Hate to say it, but in this case you’re not the market so you are irrelevant.
Here’s data from Chicago, USA, showing Uber and Lyft dominate the market now, so yes, it worked.
https://toddwschneider.com/dashboards/chicago-taxi-ridehaili...
It feels like a related statement to someone willing to burn their own country to rule over the ashes.
$32b -> $37b -> $44b -> $52b -> $57b
That's Uber's sales growth year by year. And now they're solidly profitable.
The sole threat to Uber is the Waymo / Tesla automation angle. A merger is most likely there.
The price of a given point on the intelligence frontier is only going to get cheaper, likely 10-20x over the next few years.
but profitability is a very narrow view on the labs, and investors are pricing in future free cash flow
you need to take into account the derivative, what is their growth? their opex and capex? their revenue growth? their moat? New product lines and applications?
Not saying that it's a homerun, but there is a case to be made for the labs being wildly profitable in 4-10 years
This isn’t the “hallucinate everything” style of model like the earlier GPT models…
We all feel more productive. But did your rust rewrite or my fancy react dashboard produce more value?
Maybe you could use an AI assistant to help you write better HN comments.
But make no mistake. Being more productive does not equal higher profits because your customers think you've added value. Both value and productivity are highly subjective terms which is why it is measured differently depending on where you work.
Most businesses adapted to AI by shipping a lot more features in smaller increments of time. What they should have done is become much more ambitious and build moon shots in two quarters instead. You need to go big or go home if you're going all in on AI.
i add more value to my company because it lets me ship much faster, debug issues faster, onboard new people faster
total enterprise ai spend is climbing https://ramp.com/data/ai-index#adoption#overall
i know multiple stealth startups with 9 figure revenues and profitable and growing
lab arr is climbing
given the recent crackdown on token spend and from speaking to execs they are very focused on the exact ROI they will get from AI
Note that business revenues in competitive markets may actually not change that much, because if everyone is using the same models there's no competitive advantage from using AI. but if one of them didn't use AI they would fall hopelessly behind