We're talking about places that even after decades haven't recovered. What do you think is happening there right now?
There's a common fallacy that tries to argue that it'll be alright over time, no matter what happens. Given enough time, you can also say that about atomic wars. But that won't help the generations that are affected.
If you live in a dead town with no opportunities then you either make your own opportunities or you move to a place with opportunities.
If you just sit on your hands complaining about the lack of opportunities then you won't get any sympathy from me. People aren't entitled to live wherever they want, humanity's entire thing is adaptability. So adapt. Life is what you make it.
When I say 'place' that includes entire countries. Adapting then depends on the kindness of strangers towards foreign refugees.
I wouldn't be surprised if at some point in the near future something like "Adapt. Life is what you make it" could be read in big bold letters above the entrance of a place like Alligator Alcatraz.
I think it's pretty clear that I'm not referring to Palestinians, Ukrainians or Syrians etc here. I support refugee programs but I also acknowledge that we have limited resources. We can't help everyone. We can't just open the borders and let everyone in.
I hadn't heard about alligator Alcatraz until now, I'm not American so I don't keep up with all of Trump's shenanigans. I feel compelled to make it clear that I in no way support Trump. The fact that the US has elected that clown not once but twice is frankly embarrassing.
Humans are entitled to live wherever they want. Capitalists destroying rural regions with false promises (prosperous jobs) is a thing since the industrialization. Should all people move to overrun big cities? Small once established markets are getting destroyed by big discounters or stuff like Amazon. Also adapting is and never was a thing for most people. I dont know where you got that from but this isn't the wild west anymore. People are trying to set up a life for themselves without moving every 2 years. Entitled city person viewpoint.
There's plenty of rural areas with plenty of opportunities. Cities are not the only option. If I lived in a dead mining town I'd move elsewhere.
You can blame corporations or whatever you want, doesn't matter whose fault it is. Complaining and blaming doesn't solve anything. Finding solutions does. Stop complaining, start finding solutions.
I grew up in a beautiful rural place. I'd like to live there, but what I like even more is not having to drive for over an hour to work every day. So I moved. I also went to university in my late 20s and some of my peers were in their 40s and 50s.
People adapt to all kinds of stuff all the time. Saying adapting isn't a thing for most people is ridiculous. Of course it's a thing. It's what you do when your current situation isn't working. You adapt.
And that is extremely difficult at a large scale. Especially when you reduce safety nets and heighten the consequences of failure. A lack of healthcare, poor rural hospitals, extortionate tuition at colleges, high housing costs; these all make it extremely difficult to adapt.
Sure, you can smugly say that the hard-working will survive. But i don't want to imagine what the USA will be like millions of unemployed and under-provisioned Americans. Poverty and the process of falling off the socio economic ladder is ugly for everyone, unless you're wealthy enough to afford to insulate yourself from the consequences.
The fact that this nuance appears to be lost on you makes me suspicious of your motives for posting your opinion.
I'm not talking about large scale. I'm also not talking about politics, which I assume is what you're implying re my motives. I don't have any motives, I'm not even American. I don't really care about all the crazy stuff happening over there. They got what they voted for, meanwhile the world is wondering how they elected that clown not once but twice.
The ironic thing in all this is that these rural people you're talking about are probably the exact people responsible for electing him. Evokes images of leopards eating faces and such.
is there anywhere to track the qos degradation? would love to use this for a feature we're shipping today just to try it out, but consistently get 429's on Gemini or Vertex. Checking quotas shows that neither are close to limits, which makes me think it's the DSP being infra constrained??
I have helped hundreds of people migrate to AWS and never had a single person spend more effort unless they went for an apples to apples disaster. I have only seen this when people take a high overhead tool they don’t understand (eg k8s) and move to cloud services they don’t understand.
Amazon’s effort in making sure things _actually are up_ is fundamentally different than budget clouds.
The systems you design when you have reliable queues, durable storage, etc. are fundamentally different. When you go this path you’re choosing to solve problems that are “solved” for 99.99% of business problems and own those solutions.
Still, things fail. A-tier clouds also fail, and you may still have to design for it. Rule of thumb, if you are capable of rolling out your own version, you'll be far more competent planning for & handling downtime, and will often have full ownership of the solution.
Also, any company with strict uptime requirements will have proper risk analysis in place, outlining the costs of the chosen strategy in case of downtime; these decisions require proper TCO evaluation and risk analysis, they aren't made in a vacuum.
This is a strangely limited view. Cloud providers have done the work of building fault-tolerant distributed systems for many of the _primitives_ with large blast radius on failure.
For example, you'd be hard pressed to find a team building AWS services who is not using SQS and S3 extensively.
Everyone is capable of rolling their own version of SQS. Spin up an API, write a message to an in memory queue, read the message. The hard part is making this system immediately interpretable and getting "put a message in, get a message out" while making the complexities opaque to the consumer.
There's nothing about rolling your own version that will make you better able to plan this out -- many of these lessons are things you only pick up at scale. If you want your time to be spent learning these, that's great. I want my time to be spent building features my customers want and robust systems.
I see where you’re coming from — no doubt, services like SQS and S3 make it easier to build reliable, distributed systems without reinventing the wheel. But for me, the decision to shift to European cloud providers wasn’t about wanting to build my own primitives or take on unnecessary complexity. It was about mitigating regulatory risk and protecting revenue.
When you rely heavily on U.S. hyperscalers in Europe, you’re exposed to potential disruptions — what if data transfer agreements break down or new rulings force major changes? The value of cloud spend, in my view, isn’t just in engineering convenience, but in how it helps sustain the business and unlock growth. That’s why I prioritized compliance and risk reduction — even if that means stepping a little outside the comfort of the big providers’ managed services.
> For example, you'd be hard pressed to find a team building AWS services who is not using SQS and S3 extensively
I design and develop products that rely on queuing systems and object storage; if its SQS or S3 is an implementation detail (although S3 is also a de-facto standard). Some of those products may rely on millions of very small objects/messages; some of them may rely on fixed-size multi-MB blocks. Knowing the workload, you can often optimize it in a non-trivial way, instead if just using what the provider has.
> The hard part is making this system immediately interpretable and getting "put a message in, get a message out" while making the complexities opaque to the consumer.
Not really, no. As you said, is already a solved problem. Aws obviously has different scale requirements than my reality, but by having ownership I also have only a fraction of the problems.
> There's nothing about rolling your own version that will make you better able to plan this out -- many of these lessons are things you only pick up at scale.
I cannot agree with you on this. As an example, can you tell me which isolation level is guaranteed on an Autora instance? And what if it is a multi-zone cluster? (If you can, kudos!); next question is, are developers aware of this?
If you have done any cursory solution design, you will know the importance of mastering the above questions on development workflow.
Fred Brooks, the author of The Mythical Man-Month said:
> “Software is ten times easier to write than it was ten years ago, and ten times as hard to write as it will be ten years from now.”
Ansible, Hetzner, Prometheus and object storage will give you RDS if you prompt an LLM, or at least give you the parts of RDS that you need for your use case for a fraction of the cost.
Hetzner is also working on their own Managed RDS offering. Their own S3 Offering is also relatively new. Back then, they've also had job offerings for DB Experts
When choosing distributed systems platforms to work with, k8s vs. rolling your own orchestration isn’t the decision anyone is making. It’s k8s vs cloud vendors that want your money in exchange for the headaches.
Honestly running your own control plane is not that much harder than using something like EKS or GKE. The real complexity that the grandparent was talking about is all the tweaking and configuration you have to do outside of the control plane. Eg the infrastructure and deployments you’re building on top of Kubernetes and all of the associated configuration around that. In other words, whether you use EKS or hand roll your own kube you still have to solve node auto scaling. Load balancing. Metrics/observability. DNS and networking. Ingress. Etc etc etc.
And not everyone working in things that aren’t “hacked together MVPs” has your experience. You can read any number of reports about code generated at FAANG, incident response tooling that gets to RCA faster, etc.
There are obviously still things it can’t do. But the gap between “I haven’t been able to get a tool to work” and “you’re wrong about the tool being useful” is large.
That "much smaller number" is the tricky part. Most rerankers degrade substantially in quality over a few hundred candidates. No amount of powerful rerankers will make "high powered behavior based models" more effective. Those behavioral signals and intents have to be encoded in the query and the latent space.
> Most rerankers degrade substantially in quality over a few hundred candidates.
The reason we don’t use the most powerful models on thousands/millions of candidates is because of latency, not quality. It’s the same reason we use ANN search rather than cosine sim for every doc in the index.
This isn’t true. You can look at basically every cross encoder used today and observe degradations in precision with increases in k
Ofc latency matters for retrieval pipelines and this is another reason to care. But first pass retrieval has to surface the right candidates for it to matter at all. It has to do it within the constraints of the precision degradation wrt k of the first pass reranker
by that same logic, why would you not strive to push all the signals you have available into the ANN search? sure, some will have reduced resolution vs using a heavy reranker, but surely the optimal solution is to use the same signals in both stages and just add resolution in the second stage? the more they are aligned, the fewer candidates you need -> better latency & lower cost.
If this were true, and initial candidate retrieval were a solved problem, teams where search is revenue aligned wouldn't have teams of very well paid people looking for marginal improvement here.
Treating BM25 as a silver bullet is just as strange as treating vector search as the "true way" to solve retrieval.
I don't mean to imply that it's a solved problem; all I'm saying is that in a lot of cases, the "weak initial retrieval" assertion stated by the article is not true. And if you can get a long way using what has now become the industry standard, there's not really a case to be made that BM25 is bad/unsuited, unless the improvement you gain from something more complex is more than just marginal.
one thing to remember is that bm25 is purely in the domain of text - the moment any other signal enters in the picture (and it ~always does in sufficiently important systems), bm25 alone can literally have 0 recall.
