[bottom999mottob]

Did you link a card with Pay-as-you-go? It seems like OCI puts free tier users in a low-priority queue, so that people have resorted to using auto-clicker scripts for weeks to even provision an instance.

That repo looks cool, if I really wanted to learn Kubernetes. My usage right now is for running inference on CPU RAMs because I refuse to upgrade my 980-ti, so I'm not sure it would work for me

[SOLAR_FIELDS]

You have to link a card if you don’t want to have a bot that spams looking for open instances. I tried without linking a card but it’s basically impossible without automation to secure an instance, and they can arbitrarily just nuke your instance at any time if they feel like you are being “abusive” or even just idling the machines and not using resources, plenty of complaints about that. Adding a card removes all of those concerns and you can instantly provision the resources and never worry about anything being arbitrarily shut down.

I have only had one 1 cent charge since I got it because I provisioned some volume storage before I realized that it would incur costs.

If you just use the repo I linked and otherwise never touch resources or auto provision anything on the infra side you’ll never be charged.

I would highly recommend this use case for learning kube, but also it’s kind of limiting because several things you might want to use with kube will also incur additional costs. You can avoid the costs if you simply just don’t do anything that triggers a cost event, but people who are not familiar with how kube can provision infra resources under the hood might footgun themselves and wire into a cloud load balancer instead of using MetalLB for instance.

Probably inference is not really a good use case here, the compute they give is considered extremely good in terms of free, but still kind of anemic for inference use cases.

[bottom999mottob]

Yeah I refuse to give Oracle a dime, so I'll keep that in mind.

I only run the tiny models for fun (the CPU inference is enough) and any complex coding tasks go through GPT or Claude. Mainly just using the one instance as another backup server, run scrapers, host a bunch of containers, and experiment with networking services. I don't see a reason to split up my one instance, but maybe in the future.

[SOLAR_FIELDS]

Yeah the main reason that I’m interested in using OCI is because it’s free and it actively takes money away from them by me doing this. So I will happily continue to take advantage of Larry, I know he would do the exact same to me but many times worse. So it’s a great solution as long as it lasts. I get to make Larry inconsequentially poorer every second my stuff runs, plus I get some free cloud compute. Win win

[kstrauser]

This is the first thing I’ve ever read that has me considering spinning up an OCI setup.

[SOLAR_FIELDS]

It’s a use case half driven by value and half driven by spite, and 100% about selfishly taking things at the detriment of others - which just feels so perfectly on brand for everything having to do with Larry Ellison. A microdose of his own medicine, if you will

[bottom999mottob]

Making Larry lose money while you sleep, truly a reverse Uno :)

[ipaddr]

What model are you running? I found it to be too slow.

[bottom999mottob]

SmolLM v2 1.7GB Q8 has been running pretty fast. Only needs about 3.5GB of memory. I think it's fine for chatting, but it's only really good for Python code compared to larger models.

Deepseek-coder is pretty slow, even with the quantized models.

https://huggingface.co/blog/smollm

[shiroiushi]

>Yeah I refuse to give Oracle a dime

The parent only paid $0.01. Even $0.10 should be fine, even if you strongly dislike Oracle: the cost to Oracle for charging your card $0.10 is more than that, so they're actually losing money by charging you such a small amount!