site banner

Friday Fun Thread for January 31, 2025

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

2
Jump in the discussion.

No email address required.

I keep being somewhat amazed by what R1 can do. It's really, really good at writing. Anyway:

Question: how do I buy API access? Deepseek's order page seems down. Perplexity.AI is really expensive going the obvious route. Anyone knows something about that.

Answer: probably the best way is using OpenRouter.ai . There's some free API access too, but it's kind of slow.

Question: did anyone manage to get the R1 model, which uses some novel form of selective quantisation and can allegedly run at a modest speed on a 4090 running?

Answer: .. yeah, you can run a dynamically quantised R1 but the speed is going to be <1 token.

Some fun stuff:

Youtube video: How to make an AI girlfriend smarter than you using deepeseek.

Must see. Guy jailbreaks Deepseek by telling it to roleplay as Gemini and talk about Tianamen. It shitposts with the best.


After following the guide: setting up WSL2, installing uv, installing vllm we got DeepSeek-R1-Distill-Qwen-14B running on a 4090. Must have made some major error as it runs at only .6 t/s which is only sufficient for very, very old people. Should run at perhaps 4-6 at least judging by what others have seen. Update: no error. Unless you have god-tier CPU/memory you can expect <3 tokens from big models. You need GPUs for more token /s.

Also people, just use Docker images with Text Generator Inference.. it's probably as good as vllm and it can also (allegedly) handle dynamically quantized LLMs which means you can load a better model into less memory and run it. vllm struggles with that.

For api access, openrouter should work if you're not doing anything sensitive.

... you think you'd get busted if you used openrouter to help you with writing malware etc?

Likely no. But if you fed a bunch of high-value data through openrouter for natural language processing purposes, I think there's a decent chance said high-value data finds its way into future training datasets.

Oh yeah... I think that any data you enter into such a service should be treated as published..