LM Studio got a mention, so I downloaded it.
It starts like this … only the “skip” at the top and the “legal stuff” at the bottom would un-gray themselves … ???
#1 … What the hell would “Skip ‘Get Started’ “ even mean?
#2 … I think what un-grayed the “Get Started” button was clicking the link to T&Cs … but I’m not entirely sure. Maybe there was a timer or something. This is weird design.
For automation / following simple instructions / agentic, if the subject is not too complex, smaller models may be good too. To do some relatively complex SQL against an ERP DB, as an example, it would probably need to be trained on the DB structure, etc., and be bigger than 15b to start with, otherwise it would be getting too confused all the time.
But again, there’s also prompt engineering: you can get much better results from better prompts. So instead of programming in Pascal, we may all soon enough end up writing detailed prompts – for AI to do the actual Delphi programming Conceptually, prompts are like programs, and what we mostly do now is typing, you see where I’m going with this, right?
Current commercial models don’t even advertise their sizes anymore, BTW, but they are all likely to be over 200b, perhaps 400b. – that’s well outside of small business capabilities (i.e.: 500k server as a start, to run a 200-400b model only = in rental terms that’s around AU$80 per hour, to train that would probably have to be an ~5mil server).
They all initially had the “download” icon next to them.
I think the Qwen3:4B model was about 2GB in size. Downloaded fine.
Both the Qwen3:8b and DeepSeek models were about 5GB in size … and they show them downloading on the screen when you choose a new model and ask it something …
Both of them downloaded (according to Task Manager) at around the 10 - 100 MB/s rate … until 4.6GB … when it slowed to a crawl.
Trying things, the first one in the end got enough bursts of data by disabling and then re-enabling the wifi device driver a couple times. The second one died with errors from Cloudflare.
Stopping and reselecting the same model … started the download from scratch again.
However Google, or maybe a hunch, revealed that there is a command-line interface as well.
And it is able to continue failed downloads … a technology we celebrated in 1986 with ZMODEM
LMStudio does allow you to dowload other models, I don’t have access to it at the moment but I was able to download and load a bunch of different models.
Yeah, I guess the “Get Started” is just about downloading your first model to play with and skipping that means you’re going manual/selecting your own.
Did the “Finding best options” suggest any alternate models? I don’t remember seeing the Get Started screen when I installed but that was a while ago. Perhaps it defaults to one of the current popular models (OpenAI open weight gpt in this case) but then lists some that are appropriate for your RAM/GPU?
Conceptually, prompts are like programs, and what we mostly do now is typing, you see where I’m going with this, right?
