my user name AT google's mail service
I generally don't much effort into my grammar when commenting on HN, so I apologize in advance.
- > They are obviously making awful lot of revenue.
It's not hard to sell $10 worth of products if you spend $20. profit is more important than revenue.
- > Inference is almost certainly very profitable.
It almost certainly is not. Until we know what the useful life of NVIDIA GPUs are, then it's impossible to determine whether this is profitable or not.
- > Right now models have roughly all of the written knowledge available to mankind, minus some obscure held out private archives and so on.
Sorry for the nit, but this is a gross oversimplification. Most private archives are not obscure but obfuscated and largely are way more valuable training data then the publicly available ones.
Want to know how the DOD may technically tracks your phone? Private.
Want to know how to make Coca Cola at scale? Private.
Want to know what the schematic is for a Google TPU? Private.
etc etc.
- > I understand why it's unpopular to relate the fact that applications for serious surveillance/espionage CNE work at NSA are competitive, and that the USG is very open about soliciting those applications.
And I'm not challenging you on that, at all. I actually believe you are correct because you typically provide very thoughtful answers from a position of authority and usually bring evidence to back your assertions. In this case, you're not. Your comments are childish and makes your position way less believable and hence why I'm pushing you.
> It remains a fact.
Just because you say it's a fact doesn't mean it is. The fact that you've done nothing but say "I know its a fact so therefore it is" doesn't help your position either.
> What I'm not going to do is write you an apology for confronting you with that information.
Yeesh. Chill out. I never asked for an apology. Your discourse is belittling and unproductive.
- Flippant again. Nice.
- > If you think otherwise, you're in a filter bubble.
Belittling. Excellent way to get your point across.
> They actively recruit on elite engineering campuses over it. It is super fucking interesting work and candidates compete for the opportunity to do it.
This seems like it should be an easy thing to verify with some sort of reference. This is exactly what the parent comment is suggesting and you still flippantly are avoiding it as "trust me bro". I actually believe you, so why don't you share some evidence then?
- Tesla UX gripes (may be specific to Model S 2020):
- Yolk steering is terrible
- Lack of physical knobs. Haptics are nice but haptics don't work well for cars.
- Tesla menus are getting stuffed more and more with options, making affordance and UI crowding much worse
- The horn is a button press...no one in a emergency situation is looking for a button press..
- Native apps (for example Spotify) are inferior to just using my phone via bluetooth
- Calendar integration / notification is too chatty
- > The bottleneck is still knowing what to build, not building.
My hot take - LLMs are exposing a whole bunch of developers to this reality.
- Same. If this is the situation then what is the use case for most "average" consumers?
- From a purely technical view, skills are just an automated way to introduce user and system prompt stuffing into the context right? Not to belittle this, but rather that seems like a way of reducing the need for AI wrapper apps since most AI wrappers just do systematic user and system prompt stuffing + potentially RAG + potentially MCP.
- Oracle's growth and value is in SaaS apps (NetSuite) and their cloud offering, not DB licensing. The economic impact of enterprises moving off Oracle DB is massively overstated here.
- > Or maybe these benchmarks are all wrong
You must be new to LLM benchmarks.
- "There are only two hard things in Computer Science: cache invalidation and naming things."
- I own a Rivian too, and previously owned a Tesla. While I too have my gripes about the UX on the Rivian, it still beats the cr*p out of a Tesla.
- You completely missed the point of my comment...
- Objectively yes that's the case. However, people are irrational and their irrationality is being exploited.
- While there are justifiable comments here about how LLMs behave, I want to point out something else:
There is no consensus on what constitutes a high quality codebase.
Said differently - even if you asked 200 humans to do this same exercise, you would get 200 different outputs.
- > There's a significant blind-spot in current LLMs related to blue-sky thinking and creative problem solving. It can do structured problems very well, and it can transform unstructured data very well, but it can't deal with unstructured problems very well.
While this is true in my experience, the opposite is not true. LLMs are very good at helping me go through a structure processing of thinking about architectural and structural design and then help build a corresponding specification.
More specifically the "idea honing" part of this proposed process works REALLY well: https://harper.blog/2025/02/16/my-llm-codegen-workflow-atm/
This: Each question should build on my previous answers, and our end goal is to have a detailed specification I can hand off to a developer. Let’s do this iteratively and dig into every relevant detail. Remember, only one question at a time.
- > Had the cost of building custom software dropped 90%, we would be seeing a flurry of low-cost, decent-quality SaaS offering all over the marketplace, possibly undercutting some established players.
Aha. Are developers finally realizing that just writing code doesn't make a business? We actually have a ton of SaaS companies being born right now but they're not making headway, because functionality and good code don't necessarily mean good businesses. Building a business is hard.
How do you know this?
> You don't need perfect certainty about GPU lifespan to see that the spread between cost-per-token and revenue-per-token leaves a lot of room.
You can't even speculate this spread without knowing even a rough idea of cost-per-token. Currently, it's total paper math on what the cost-per-token is.
> And datacenter GPUs have been running inference workloads for years now,
And inference resource intensity is a moving target. If a new model comes out that requires 2x the amount of resources now.
> They're not throwing away two-year-old chips.
Maybe, but they'll be replaced by either (a) a higher performance GPU that can deliver the same results with less energy, less physical density, and less cooling or (b) the extended support costs becomes financially untenable.