> Anyone who has ever used Mercurial knows very well what a good versioning tool UX looks like...
So true. I used Mercurial back in the day and also used Darcs before it, and it helped me realize that the best versioning tool UX that exists is still the one Git provides.
PS: Also CVS, SVN, Perforce, and Clear Case professionally, and gave a try to Fossil. None of them even close to Git usability-wise.
This subsidized inference is just a marketing ploy to increase prices and profit.
If common people can have a DIY setup with an open source model cheaper than those behemoths with a scale advantage, it's clear that we have been played.
Time to either self host a Chinese open source model or to just pay the cheap Chinese providers.
Yeah, local is clearly the future. Even beyond the cheap Chinese models you can install the apfel[1] stuff if you're on a mac and want a quick available onboard cli option. And I'm sure people will adapt the Flash-MoE[2] integration to be even better soon as well.
Don’t forget the hate speech laws. It’s just ridiculous. A state in Germany wants to criminalize questioning a certain country’s existence, with penalties of up to four years in prison
A strange view. The trade-off has nothing to do with a specific ideology or notable selfishness. It is an intrinsic limitation of the algorithms, which anybody could reasonably learn about.
Sure, the exact choice on the trade-off, changing that choice, and having a pretty product-breaking bug as a result, are much more opaque. But I was responding to somebody who was surprised there's any trade-off at all. Computers don't give you infinite resources, whether or not they're "servers," "in the cloud," or "AI."
He was surprised because it was not clearly communicated. There's a lot of theory behind a product that you could (or could not) better understand, but in the end, something like price doesn't have much to do with the theoretical and practical behavior of the actual application.
So that leads to a question: Is there a physical box I could buy that an amortize over 5-7 years to be half the API cost?
In other words, assuming no price increase, 7 years of that pricing is $15k. Is there hardware I could buy for $7k or less that would be able to replace those API calls or alternativr subs entirely?
I've personally been trying to determine if I should buy a new GC on my aging desktop(s), since their graphic cards can't really handle LLMs)
You can't realistically replace a frontier coding model on any local hardware that costs less than a nice house, and even then it's not going to be quite as good.
But if you don't need frontier coding abilities, there are several nice models that you can run on a video card with 24GB to 32GB of VRAM. (So a 5090 or a used 3090.) Try Gemma4 and Qwen3.5 with 4-bit quantization from Unsloth, and look at models in the 20B to 35B range. You can try before you buy if you drop $20 on OpenRouter. I have a setup like this that I built for $2500 last year, before things got expensive, and it's a nice little "home lab."
If you want to go bigger than this, you're looking at an RTX 6000 card, or a Mac Studio with 128GB to 512GB of RAM. These are outside your budget. Or you could look at a Mac Minis, DGX Spark or Strix Halo. These let you bigger models much slower, mostly.
Thanks. That is what I suspected. The 3090's in my area seem pretty expensive for a several year old second hand card - they are the same price as a new 5080.
5090 is pretty expensive (~$4000) to justify it over a $10-50 sub. I guess the nice thing is the api side becomes "included", if I ever want to go that route. But if I have a GHCP $40 sub vs a $4000 GC to match it, just on hardware, pay off is at 8 years. If I add in electricity, pay off is probably never.
Sure, the sub can go up in price, but the value proposition for self-running doesn't seem to make sense - especially if I can't at least match Sonnet on GHCP or something like that.
I hope to self-run some not useless LLMs/Agents at some point, but I think this market needs to stabalize first. I just don't like waiting.
For what it's worth, eBay in the US currently has some used 3090s for about $1,300, including some marked "Buy it now." I got mine used for about $1,000, and I'm really happy with it—it's a very solid gaming card for Steam on Linux (if you don't need ray tracing), and it allows me to experiment with models up to about 35B parameters. I'm not saying it's a good investment for you in particular, of course! But it's solid at that price, and you can just chuck it in any consumer gaming rig and get a really fun AI "home lab".
As for models, I'm really genuinely impressed with Gemma4 26B A4B and Qwen3.6 35B A3B right now. Between them, I've seen solid image analysis, good medium-image OCR on very tough images, very good understanding of short stories, good structured data extraction from documents, extremely good language translation, etc. If you wanted to build a custom tool which summarized your inbox/RSS feeds/local news every day, or extracted information from emails and entered it into a database, or automatically captioned images, those tasks are all viable locally. The quality of the results is up dramatically in the last 12 months. At this point, my old personal non-agentic LLM benchmarks are "saturated": All the current leading models score extremely well on literally anything I was asking last year.
It's the true agentic coding workflows where the big models really stand out. And those models are all large enough that the hardware needs to amortized over enough users to run 24 hours/day.
> or a Mac Studio with 128GB to 512GB of RAM. These are outside your budget.
M3 ultra with 80GOu cores and 256GB of ram is $7500 - that’s right at the edge of the budget, but it fits.. if you can get an edu discount through a kid or friend you’re even better off!
You can buy a roughly $40k gpu (the h100) which will cost $100/mo in electricity on top of that to get about 30-80% the performance of OpenAI or Anthropic frontier models, depending what you're doing.
Over 5 years, that works out to ~$45k vs ~$10k, and during that duration, it's possible better open models will come available making the GPU better, but it's far more likely that the VC-fueled companies advance quicker (since that's been the trend so far).
In other words, the local economics do not work out well at a personal scale at all unless you're _really_ maxing out the GPU at close to 50% literally 24/7, and you're okay accepting worse results.
As long as proprietary models advance as quickly as they are, I think it makes no sense to try and run em locally. You could buy an H100, and suddenly a new model that's too large to run on it could be the state of the art, and suddenly the resale value plummets and it's useless compared to using this new model via APIs or via buying a new $90k GPU with twice the memory or whatever.
Note that the (edit: US) postal system is a for-profit system.
Given the trends of the capitalist US government, which constantly cedes more and more power to the private sector, especially google and apple, I assume we'll end up with a state-run model infrastructure as soon as we replace the government with Google, at which point Gemini simply becomes state infrastructure.
> Note that the (edit: US) postal system is a for-profit system.
That's not correct. If USPS makes more revenue than their expenses for a year, they can't pay it out as profits to anyone.
It's true that USPS is intended to be self-funded, covering it's costs through postage and services sold, and not tax revunue. That doesn't mean there's profit anywhere.
> Note that the (edit: US) postal system is a for-profit system.
Pricing in the US postal system is not based on maximizing profit. Ths US postal system is not a for-profit system, at all. It is a delivery system (more or less) that happened to start turning a profit (2006) until PAEA. After that, the next time it made a profit was 2025.
For something like OpenClaw you realistically only need rather slow inference, so use SSD offload as described by adrian_b here: https://news.ycombinator.com/item?id=47832249 Though I'm not sure that the support in the main inference frameworks (and even in the GGUF format itself, at least arguably) is up to the task just yet.
You can get quite good models running on a Mac Studio, but these will not rival a frontier model.
$3,699.00
M4 Max 16c/40c, 128GB of RAM, 1TB SSD.
LM Studio is free and can act as a LLM server or as a chat interface, and it provides GUI management of your models and such. It's a nice easy and cheap setup.
Forgejo, on the other hand, is a drop-in replacement for GitHub.
reply