> Who decides what gets to be reverse engineered and what doesn't?
whoever owns the hardware
currently running 2x DGX Sparks and never thought to use a commercial model again
> Who decides what gets to be reverse engineered and what doesn't?
whoever owns the hardware
currently running 2x DGX Sparks and never thought to use a commercial model again
Curious about your setup if you don’t mind elaborating a bit for us? What models are you using and how do you find they compare? The 128gb model costs less than one months salary for a medium level engineer here, they look quite attractive.
A month and a half salary after tax here. That is still particularly expensive, and besides its sinking money into a steeply depreciating asset.
That’s about five months of wages up here in rural nowhere NY doing IT, for what it’s worth.
Some of us are just out of the AI race completely.
You mean appreciating? They have doubled in price since release...
i've been running DeepSeekV4 flash but looking at the other models being suggested around, best part is switching to test has 0 cost :)
i felt like commercials LLMs were "good enough" for my needs about 2-3 generations ago, so this clicked into my work/life needs perfectly
i got them right before they went up in price, but honestly, for the ability to know that i wont have my thinking "questioned"... priceless
Best model on 2xSpark currently is GLM 5.3 flash ... but really does require 2x Spark. 25 to 60 tok/s with MTP.
2x Spark? As in two DGX Spark machines? That is obscenely expensive.
Is it? How much would it cost to hire a person to do the same work two of these can do?
Maybe fair to compare to using AI subscription ~20$/m (not Anthropic neither Openai)
Especially now with OpenAI constantly changing (and always reducing) what the subscription can do ... At least it will keep working.
Right now I need 4 subscriptions to do what 1 used to do. And they're tightening it further, without announcing:
https://news.ycombinator.com/item?id=50320588
(technically it isn't even legal, since changing what people get after they've agreed to the sale is obviously highly illegal)
Not that they've ever actually announced how much a subscription buys you. In fact the token use for a task on the API is also not consistent, so even on the API you don't know. And then there's Sam Altman in interviews, where he says what he wants to do: you submit a prompt, ChatGPT sends you a price and just charges it.
> whoever owns the hardware
Yes and this is why self-hosting is more important than ever. The issue with LLMs is the cost of hardware if you want something close to the frontier. But there is no excuse not to host your own files, email, media center, etc.
Email is, given the issues with connectivity, something you can justifiably make someone else's problem.
There's no reason not to have your own domain though. I've been using Migadu and the amount of flexibility that gives you is ridiculous. The only downside is the lack of push notifications, but if your email is so important that you can't wait for a refresh then I can't help.
same as self-hosting email/webservers though, you dont need world-class hardware for personal projects. frontier models may be more capable but a smaller model can just as easily work through a project you break down for it