Amazon has mostly sat out the AI talent war

A better way to look at it is that the absolute number 1 priority for google since they first created a money spiggot throguh monetising high-intent search and got the monopoly on it (outside of Amazon) has been to hold on to that. Even YT (the second biggest search engine on the internet other than google itself) is high intent search leading to advertising sales conversion.

So yes, google has adopted and killed lots of products, but for its big bets (web 2.0 / android / chrome) it's basically done everything it can to ensure it keeps it's insanely high revenue and margin search business going.

What it has to show for it is basically being the only company to have transitioned as dominent across technological eras (desktop -> web2.0 -> mobile -> (maybe llm).

As good as OpenAI is as a standalone, and as good as Claude / Claude Code is for developers, google has over 70% mobile market share with android, nearly 70% browser market share with chrome - this is a huge moat when it comes to integration.

You can also be very bullish about other possible trends. For AI - they are the only big provider which has a persistent hold on user data for training. Yes, OpenAI and Grok have a lot of their own data, but google has ALL gmail, high intent search queries, youtube videos and captions, etc.

And for AR/VR, android is a massive sleeping giant - no one will want to move wholesale into a Meta OS experience, and Apple are increasingly looking like they'll need to rely on google for high performance AI stuff.

All of this protects google's search business a lot.

Don't get me wrong, on the small stuff google is happy to let their people use 10% time to come up with a cool app which they'll kill after a couple of years, but for their big bets, every single time they've gone after something they have a lot to show for it where it counts to them.

For Amazon “renting servers” at very high margin is their cash cow. For many competitors it’s more of a side business or something they’re willing to just take far lower margin on. Amazon needs to keep the markup high. Take away the AWS cash stream and the whole of Amazon’s financials start to look ugly. That’s likely driving the current panic with its leadership.

Culturally Amazon does really well when it’s an early mover leader in a space. It really struggles, and its leadership can’t navigate, when it’s behind in a sector as is playing out here.

Does he have this net worth because what he is doing or despite what he is doing? :-)

Correlation does not imply causation. Attribution is hard.

Buying competition is par for the course for near-monopolies in their niches. As long as the scale differences in value are still very large, you can avoid competition relatively cheaply, while the acquired still walk away with a lot of money.

Is synthesizing the right word here?

Outside of computer, "the moat" is also data to train on. That's an even wider moat. Now, google has all the data. Data no one else has or ever will have. If anything, I'd expect them to outclass everyone by a fat margin. I think we're seeing that on video however.

I tend personally to stick with ChatGPT most of the time, but only because I prefer the "tone" of the thing somehow. If you forced me to move to Gemini tomorrow I wouldn't be particularly upset.

xAI seems to be the exception, not the rule

I don't know what you are talking about. I use Gemini on a daily basis and I honestly can't tell a difference.

We are at a point where training corpus and hallucinations makes more of a difference than "model class".

I wouldn't be surprised if the likes of Anthropic wasn't paying AWS for its compute.

As the saying goes, the ones who got rich from the gold rush were the ones selling shovels.

As more opens up in OSS and academic space, their knowledge and experience will either be shared, rediscovered, or become obsolete.

Also many of the people are coasting on one or two key discoveries by a handful of people years ago. When Zuck figures this out he gonna be so mad.

Hopefully some big players, like FB bankrupt themselves.

Even assuming a company gets to AGI first this doesn't mean another one will follow.

Suppose that FooAI gets to it first: - competitors may get there too in a different or more efficient way - Some FooAI staff can leave and found their own company - Some FooAI staff can join a competitor - FooAI "secret sauce" can be figured out, or simply stolen, by a competitor

At the end of the day, it really doesn't matter, the equation AI === commodity just does not change.

There is no way to make money by going into this never ending frontier model war, price of training keeps getting higher and higher, but your competitors few months later can achieve your own results for a fraction of your $.

There's a bunch of ways AI is improving itself, depending on how you want to interpret that. But it's been true since the start.

1. AI is used to train AI. RLHF uses this, curriculum learning is full of it, video model training pipelines are overflowing with it. AI gets used in pipelines to clean and upgrade training data a lot.

2. There are experimental AI agents that can patch their own code and explore a tree of possibilities to boost their own performance. However, at the moment they tap out after getting about as good as open source agents, but before they're as good as proprietary agents. There isn't exponential growth. There might be if you throw enough compute at it, but this tactic is very compute hungry. At current prices it's cheaper to pay an AI expert to implement your agent than use this.

It is far from clear. There may well be emergent hierarchies of more abstract thought at much higher numbers of weights. We just don't know how a transformer will behave if one is built with 100T connections - something that would finally approach the connectome level of a human brain. Perhaps nothing interesting but we just do not know this and the current limitation in building such a beast is likely not software but hardware. At these scales the use of silicon transistors to approximate analog curve switching models just doesn't make sense. True neuromorphic chips may be needed to approach the numbers of weights necessary for general intelligence to emerge. I don't think there is anything in production at the moment that could rival the efficiency of biological neurons. Most likely we do not need that level of efficiency. But it's almost certain that stringing together a bunch of H100s isn't a path to the scale we should be aiming for.

Companies like OpenAI and Anthropic are still incredibly risky investments especially because of the wild capital investments and complete lack of moat.

At least when Facebook was making OpenAI's revenue numbers off of 2 billion active users it was trapping people in a social network where there were real negative consequences to leaving. In the world of open source chatbots and VSClone forks there's zero friction to moving on to some other solution.

OpenAI is making $12 billion a year off of 700 million users [1], or around $17 per user annually. What other products that have no ad support perform that badly? And that's a company that is signing enterprise contracts with companies like Apple, not just some Spotify-like consumer service.

[1] This is almost the exact same user count that Facebook had when it turned its first profit.

The two are effectively separate businesses with a completely separate customer base.

In general these fall into the category of things humans cannot do at the scale and speed necessary to run SaaS companies.

Many of the things LLMs attempt to do are things people already do, slowly and relatively accurately. But until hallucinations are rare, slow expensive humans will typically need to be around. The AI booster’s strategy of ignoring/minimizing hallucinations or equivocating with human fallibility doesn’t work for businesses where reliability is important.

Note that ML algorithms are highly imperfect as well. Uber’s prices aren’t optimal. Google search surfaces tons of spam. But they are better than the baseline of no service exists.

https://kiro.dev/

Pay no attention to the cracks that are showing. Nevermind the chill. Everything is fine.

Only thing it can do is set a timer, turn off a light and play music.

It is still nice, but it’s so frustrating when a question pops into my mind, and I accidentally ask Alexa just to get reminded yet again how useless it is for everything but the most basic tasks.

And no, I won’t pay 240 dollars a year so that I can get a proper response to my random questions that I realistically have only about once a week.

But the project is pretty much dead, it was supposed to launch in February or March and is still not anywhere close to being out.

I'm curious if non prime members make up a big market for Alexa. I rarely use my smart devices for anything beyond lights, music, and occasional Q&A, and certainly can't see myself paying 20$/month for it.

https://en.m.wikipedia.org/wiki/Amazon_Robotics

Well, let’s take your life. Your life is about 3 billion seconds (100 year life). That’s just 3 billion next-tokens. The thing you do on second N is just, as a whole, a next token. If next-token prediction can be scaled up such that we redefine a token from a part of language to an entire discrete event or action, then it won’t be hard for the model to just know what you will think and do … next. Memory in that case is just the next possible recall of a specific memory, or next possible action, and so on. It doesn’t actually need all the memory information, it just needs to know that that you will seek a specific memory next.

Why would it need your entire database of memories if it already knows you will be looking for one exact memory next? The only thing that could explode the computational cost of this is if dynamic inputs fuck with your next token prediction. For example, you must now absolutely think about a Pink Elephant. But even that is constrained in our material world (still bounded physically, as the world can’t transfer that much information through your senses physically).

A human life up to this exact moment is just a series of tokens, believe it or not. We know it for a fact because we’re bounded by time. The thing you just thought was an entire world snapshot that’s no longer here, just like an LLM output. We have not yet trained a model on human lives yet, just knowledge.

We’re not done with the bitter lesson.

> For example, I can't easily run models that need GPUs without launching classic EC2 instances.

Yeah okay, but you can run most entreprise-level models via Bedrock.

Fargate and lambda are fundamentally very different from EC2/nitro under the hood, with a very different risk profile in terms of security. The reason you can't run GPU workloads on top of fargate and lambda is because exposing physical 3rd-party hardware to untrusted customer code dramatically increases the startup and shutdown costs (ie: validating that the hardware is still functional, healthy, and hasn't been tampered with in any way). That means scrubbing takes a long time and you can't handle capacity surges as easily as you can with paravirtualized traditional compute workloads.

There are a lot of business-minded non-technical people running AWS, some of which are sure to be loudly complaining about this horrible loss of revenue... which simply lets you know that when push comes to shove, the right voices are still winning inside AWS (eg: the voices that put security above everything else, where it belongs).

https://www.cnbc.com/2025/08/08/chatgpt-gpt-5-openai-altman-...

> Last year, OpenAI expected about $5 billion in losses on $3.7 billion in revenue. OpenAI’s annual recurring revenue is now on track to pass $20 billion this year, but the company is still losing money.

> “As long as we’re on this very distinct curve of the model getting better and better, I think the rational thing to do is to just be willing to run the loss for quite a while,” Altman told CNBC’s “Squawk Box” in an interview Friday following the release of GPT-5.

Selling compute for less than it cost you will have as much revenue as you want to pay for.

My intuition is that the root cause it's their frugal culture (frugal as in cheap). They don't want to start a compensation race.

*Microsoft enters the conversation

at that given point in time, this was not their main businesses and they fared quite well.

microsoft missing the mobile is different, because mobile being a competitor to desktop destroyed microsoft's main business.

In all of these cases, the problem was losing track of what actually benefits users. AI has that problem really bad now because the infrastructure is expensive and the executive class has been sold on the idea that mass layoffs are just around the corner, and they’re pushing hard to ship before the benefits are there.

Oh, and if the stock actually goes up more than 15%, then regardless of your performance you won't get a raise because you've already exceeded band penetration.

Once a company gets big off its grand idea, there's little to no chance of it having another big winner, so buying one is best (and its cheaper too, you know it's a good idea, and you don't have to spend so much R&D on it.

You say this as if it's a coercive given, when you could just as easily say.. Nope, and continue to see how you compete with some agility. It might fail, but most of the big tech companies currently acquiring smaller companies themselves started small with acquisition offers being rejected along the way. Sure, there's selection bias at work there, but there are also many cases of smaller to mid-size companies that also said no to acquisition and still managed to find their successful niche.

Being acquired is not a given and neither is failure if you do compete in some way with the megacorps.

I see nothing about the current tech landscape that at all distinguishes it from previous landscapes in which smaller companies succeeded AND rejected acquisition.

Once a use case and platform has stabilized, they'll provide it via AWS, at which poiny the SME market will eat it up.

By whom? Certainly no one I work with. AWS has some sharp edges and frustrations but we couldn't do half of what we do without it.

(Though I'm pretty familiar with some of the concepts, I know some things to avoid (e.g., "push this button to set up a very expensive global enterprise scale observability platform of numerous complicated services, because you asked about a very simple turn-key syslog service"), and I'm expecting the occasional configuration headache (and, lately, configuration wizard bugs).)

For a new startup, I'd use AWS for all serving and hosting purposes by default, iff you have someone who can avoid pitfalls, and handle problems.

If you don't have such a technical person, maybe start off with managed Kubernetes service with high-level UI, at AWS or one of the other cloud providers, and try not to make too big a mess (which might slow you down, or take you down) before you can afford to hire specialists to make sure it keeps working for you.

News to me

It's the same as saying buying electricity from a network is worse than having your own generators.

It was common in the post wwii era in America and its Asian allies like Korea with its chaebols and Japan with its somethings I can’t remember the name of. The Asian countries forms were normally based around a single family, we’ll need more time with the current US form to see if they are also dynastic

Apple, on the other hand, hasn’t even invested in any of the players.

[1] https://www.amazon.science/tag/formal-verification

[2] https://machinelearning.apple.com/research

I don't think his point was we should hire junior engineers because they're cheap and lean into AI and AI produces slop. His position is not that he wants to cheaply create slop.

He wants to hire people who are cheap and love using AI because he sees that as a better long term strategy than making senior engineers embrace AI late into their career.

The physical server itself would be the wooden handle, I guess.

We saw this with crypto mining where truckloads of expensive GPUs were dumped in the trash after the proof of work became so hard it became not worth the cost of electricity to keep on that generation of card.

I’m not 100% convinced this is true. Additionally, I’m not convinced that a waiting pattern right now sets Amazon up for a point of no return. It seems plausible for Amazon to pull an Apple here, to wait until technology is more mature and use their unique position to provide a quality offering.

But I think you are making it sound like Amazon's moat is that it came up with its own technology behind its services.

A lot of times AWS was just grabbing a bunch of popular open source stuff off the shelf and hosting it (e.g., RDS, EKS, etc). Yes there is some R&D work but almost none of what Amazon has come up with is rooted in their own work.

The value they give you is the hosting, maintenance, and compliance of all these services. If you're paying AWS extra to host your database on RDS or your Kubernetes cluster in EKS, you're generally not paying AWS to come up with a better database than anyone else, you're just paying them to help you manage permissions, backups, replication, and other maintenance/compliance/management issues that a company needs for its internal services.

In other words, Amazon's AI customers don't need Amazon to build models. They just need Amazon to use someone else's models, host them on private enterprise compute that easily ties in to existing infrastructure, RBAC, etc, and make everything compliant and easy to maintain. A whole lot of the value is being able to answer audits with "AWS handles our database backups/data security/etc" rather than saying "we have a great ops team and here's all our proof that we handle our database backups/data security/etc properly."

I think it's actually explicitly Amazon's job to sit this one out, especially since they never successfully made a good business or consumer ecosystem device like a smartphone or PC operating system.

This sounds like you’re either unfamiliar with what software they make or underestimate the complexity of things like a modern operating system. For example, most people would consider Swift hard, or the various Core frameworks, or things like designing a new modern file system and doing in place migrations on billion devices, etc.

How do you explain the Elon keiretsu, though? Tesla and SpaceX are pretty tethered to the physical world, and in theory should have visibility into the same discrepancies that Apple sees. So why is Elon pushing so hard to develop Grok? Is it just ideology for him, or what?

Amazon I think just hasn't understood how to cohesively integrate AI into their offerings. Meanwhile they're selling shovels to the prospectors with AWS.

I guess both of these understand the Ai moat is not very large, and don't buy into AGI dreams.

Perhaps working for American companies remotely will change that view, but it’s too much a hassle for me at the moment.

Maybe compared to FAANG, but not compared to most corporate developer jobs out there.

They don't seem to give a shit. In the retail space their name means "low quality Chinese counterfeit products with fake reviews" and I've seen no effort on Amazon's part to counter that perception either.