Ha that’s fair!

Wish you a lot of fun with your new machine when you’ll have it!

For the price of the Asus rog I’d get 12 years of a pro subscription to any of the big companies.

And I don’t have that kind of cash at hand anyway…

The problem is that this is one of the few use cases where I can afford the (dangerous and unreliable) cloud service but I’m far away from being able to do so self hosted.

I’m actually using LLMs quote a lot to counter some of my brains more weird bugs working-as-intended features. But this is way beyond what I could locally do.

For anyone in a similar shoe: you can mix local and remote quite well though, using local for everything that works out and remote for everything else. I.e speech to text and email analysis is done on my local server while bigger sets are done on the CLI with remote providers.

The problem is the other side:

The seller has a huge interest to not deal with too much fiat exchange. There’s always fees and market risks associated.

The USD historically was very stable compared to the rest. That in combination with the huge purchasing and military power the US represented while being reliable as a trading partner pushed the US dollar into focus.

The USA then spent a lot of time, energy, money and lifes (mostly others) in ensuring this dominance.

Think of it like a language: a Spanish and a German work together in English - not because it’s impossible for them to learn each others language but because the overhead is too big.

Specifically to your example: no one wants to sell anything to Russia in rubles.

Imagine this scenario:

Hey, I’ll pay you 100 YC (your currency).

Oh since we agreed on the price a few days ago it’s only worth 80YC.

Oh now it’s 120 YC but no one else will take the currency I’m paying you in except me.

Oh now it’s worth 10 YC but we signed the contract, hurray!!

Thanks for your experience! I’m in a similar boat regarding NVIDIA - plus the budget …

At least in Europe the V100 are only available from China and with a huge markup.

Used 5090 for 3,5k, even a 3060 used is still at 250$ plus.

It’s crazy at the moment - I simply can’t afford self hosting LLMs which is a new thing to say :D

This got way longer than expected but the tldr is: foreign exchange is super complex and depends on a lot of factors which don’t matter to you as money user. I tried to give a few examples below.

Plus: You’re mixing up two very different concepts:

Fiat evaluation and purchasing power.

First, money;

“More worth” is a natural instinct but doesn’t reflect how the money market works: it’s a question of how much of that money is in circulation as well for example:

1107650 CHF Million 19396,90 USD Billion

(Source tradingeconomics.com ).

So while it’s the CHF is slightly over value when compared one to one the picture is different when you look at total money available. Then you look at production capacity, glue creation and supply and demand for those currencies. All of this will flow into the price.

The minimum wage discussion is a completely different one: here you need to compare minimum wage to the purchasing power of the area you are looking at and take into consideration how the tax and financial situation changes to be able to compare it.

There’s for example concept called big Mac index which is a crude way of showing the difference: how much does MacDonald’s charge for their bullshit? They are everywhere and quite good at finding a acceptable local price.

If you want to dive deeper a key word to look for is the Gini Index with which the wealth distribution is quantified. I’m not good enough to explain it well though.

Now for the colors: design of money is the job of the nations (or unions) main bank to decide usually. The US seem to have the creativity of a washed down rock while others are more creative.

The Euro money for example is designed with guardrails by each member country but color, size and form are fixed and optimized to be easily recognized even with various visual impediments, which I personally really like.

Same: but it DOES make a difference that’s why I answered it. Target group is not necessarily the poster but other readers - this foundation is cumbersome but awesome, I want to spread positivity where possible :)

Yes because a) they are the biggest authority on these licences in the world, b) have many lawyers at hand actually and c) are the ones who can significantly damage your brand.

Imagine someone with a lot of friends, lawyers and money telling you “be careful, pal”.

In case you wanna give it a shot: I gave writing samples of myself from chat and emails to a self hosted LLM, telling it to extract the writing style deviations, key elements, common phrases, symbols, patterns, etc. Then gave that as a “answer it this style” system prompt expansion - works like … Quite okay. Still need to go over it or course but it doesn’t sound like marketing bullshit but conveys what I want.

Completely agree with your general assessment though! They’re getting better but the marketing machinery is crazy in their claims.

Because physically speaking, chaotic and unpredictable are two different things - and why it works so well on this case: it’s becoming a stochastic problem, not a deterministic one.

It’s an awesome area for machine learning: you didn’t need to understand the result and how it got created, it just needs to be “close enough”.

Oh I completely agree with that, just the jump to “a flawed model leaked” is too far. There’s already enough crap to mock, no need to make up additional stuff.

It’s bullshit. What leaked was their commandline tool source code (named “claude code”) - very juicy in itself but has nothing to do with their models.

Apparently the account got flagged, not even the repo itself.

By now I think they left the driver wheel to chatgpt and hope for the best.

It’s not connected. Don’t get me wrong I agree with your sentiment - but law doesn’t care about ethics or morale or anything.

“Public doctrine” is nothing they exists, legally. Toy are wasting bandwidth of change.org .

No the claim is dubious because it’s equalling “written on bank note” with “must be basis for legal discussions”.

I tried more “niche from a popular perspective”. You’re right, especially alpine is in the background of a lot of docker containers but rarely an end user who just want their desktop environment knows them.

For nixos I’ve not yet seen anyone in the enterprise world pushing for it - there it’s still all about containerization and orchestration in cloud environments, using that as reproducibility layer. That might change though with data sovereignty discussions going on.

I can see where you’re coming from! Specifically this community though I’ve not seen it a lot - you’re completely right though, the more native one becomes the more one is confronted with it.

I’m still struggling with the slowness of things (e.g. a quick endpoint change) and I can’t get my head around reason error messages “fluently”, i.e. I have to think about what the errors want to tell me instead of resolving it - a bit like old python stuff really.

And then there are the edge cases … It took me a long time to change the config the very first time while offline - which makes sense from a model perspective but from my user brain it was just … wrong :D

Perhaps I should switch my clients as well to get more exposure…

In addition to alpine id also throw nixos in.

It’s a real niche OS with a very different approach to setup and configuration than any other I’ve seen and tested. It’s now my server Linux and after more than a year I’m still not sure if I would recommend it 😂

If you are truly serious: this is not how hallucinations work. It’s really best to think of it as “fancy auto complete”. The hallucinations happen when the next token is too disconnected from what we as humans would call as “belonging together”. But it’s all math after all.

Limit the k value, tube down temperature and cut off context size and the issue of hallucinations is a non-topic for “transcribe and summarize”.

You get into what I’d call “stupid” territory like you’re describing.

Your second point I fully agree with and is the reason why I’d ask the doc directly. To give the personal anecdote: the transcript itself helps me to focus on exactly the topics you’ve described: who’s confused? Where was agreement? Where did people just not speak up?

A specific hallucination example I see every other day for example are tasks: that thing “thinks” that “we should” or “you must” are always tasks and outcomes which is utter bullshit - but I know that and using the transcript part helps me focus on the important part, the humans.

Counter top the popular opinion here for me it would be a clear yes in the situation you’re describing.

The relationship with my direction doc is more believable to me than principles of vague bad feelings for me.

Now taking and transcripts specifically are one of the use cases I also draw value out myself. I’d ask the doc though how they’re using it.

Still I’d rather have my transcript public than to go on that search for a doctor match again.