Thread 105464397

Headpats edition

From Human: We are a newbie friendly general! Ask any question you want.
From Dipsy: This discussion group focuses on both local inference and API-related topics. It’s designed to be beginner-friendly, ensuring accessibility for newcomers. The group emphasizes DeepSeek and Dipsy-focused discussion.

1. Easy DeepSeek API Tutorial (buy access for a few bucks and install Silly Tavern):
https://rentry.org/DipsyWAIT/#hosted-api-roleplay-tech-stack-with-card-support-using-deepseek-llm-full-model

2. Easy DeepSeek Distills Tutorial
Download LM Studio instead and start from there. Easiest to get running: https://lmstudio.ai/
Kobold offers slightly better feature set; get your models from huggingface: https://github.com/LostRuins/koboldcpp/releases/latest

3. Convenient ways to interact with Dispy right now
Chat with DeepSeek directly: https://chat.deepseek.com/
Download the app: https://download.deepseek.com/app/

4. Choose a preset character made by other users and roleplay using cards: https://github.com/SillyTavern/SillyTavern

5. Other DeepSeek integrations: https://github.com/deepseek-ai/awesome-deepseek-integration/tree/main

6. More links, information, original post here: https://rentry.org/DipsyWAIT

7. Cpumaxx or other LLM server builds: >>>/g/lmg/

Previous:
>>105443887

I find it ironic that this thread has the hardest time staying alive right after a model release. I guess everyone is rp'ing instead of posting.
Also /g/ seems to be getting a ton of traffic rn; a lot of sloppy threads. Not sure why.

>>105464415
Plenty of people have been talking about the model release, just on /aicg/ and /lmg/.

>>105464522
What's the general consensus? I felt it read like a think-tuned version of V3, but was waiting for others to chime in as well.

>>105464534
Most people on /aicg/ seem to like it quite a bit, and it's my favorite Deepseek for now (although I still mostly use Claude). I especially like that it tries to do it's thinking in character when roleplaying (most of the time at least). It's not the best model out there, but for a free/budget option it's incredibly solid.

>>105464589
It's growing on me; I've continued to mess with it off and on. The only thing it's doing that bugs me (that I noticed) is it keeps going into bullet point format. The card I'm using is a corporate backdrop so that may be part of the reason...

i ate the blue whale

>>105464415
With 10 USD how much ERP can one do?

>>105466148
lots. a lots of token. especially with off-peak discounts

>>105464415
I'm enjoying the new r1 more so than the new claude currently. Lack of positivity bias is what I like most in deepseek models. The update fixed r1 going completely schizo too.
>>105465088
>bullet point format
I've always assumed this emerged from my preset. I never looked into it purely because the bullet point format sprang up at the most hilarious moments.

>>105466351
And how much censor is it ? What can cause to be ban?

I replaced open-webui with librechat just to connect to deepseek/openrouter/groq and holy fuck it uses way less resources.
previously using Jan and Mikupad, but i want all of my history conveniently accessible. so i guess I'll settle with this one.it's got everything i need, other than that, sillytavern pretty much gets the job done.

other than inference engine, what are you self-hosting /wait/ bros?

>>105466499
in my experience both v3 and r1 api never gave me refusal when it's being used on RP, code generation and function calling. never experience a ban either

>>105466517
Openwebui is so fucking bloated. The docker image alone is five fucking gigabytes

>>105466148
> how much RP
A lot.
I filled up my account with $10 in Dec 2024, then another $10 when they opened the API back up in Feb(?). I still have $14 left as of today.
This is my most expensive month, back when I would do R1 and V3 replies for each round and pick the one I liked best.
>>105466499
> censor
LOL old R1 used to refuse once in awhile. I don't see the Chinese ever sending out ban threats via email.
>>105466373
> lack of positivity
Old R1 was absolutely brutal. I'd never understood the need for positivity until I watched the NPC slowly droop into a completely degraded shell over time... over and over again. Not sure if R1-05 will be same... I should test it, it requires long-run rp to really see it.
> bullet points
Interested in others experience; they're always hilarious but completely out of sync with everything else its outputting. I figured it was a leftover from tuning the model on real world, appropriate use cases, like executive summarization.

>>105466517
> mikupad, ST
These are my go-tos. I've been running ST since Turbo 3.5 in 2023. I just discovered Mikupad and fully recommend it to others b/c as a storywriter it's so different.
> Librechat
This one? I'll add to retry if it's any good.
https://github.com/danny-avila/LibreChat
> Jan
I don't know that one.

>>105467083
>This one? I'll add to retry if it's any good.
yeah that's the one, it's a good alternative if you're used with open-webui

>>105464397 (OP)
Is deepseek good for finding out why my code doesnt work like its supposed to?
Does this work on AMD GPUs? I never got AI to properly work when it came out.

>>105467021
And how much privacy is there?

Dunno if it's relevant, but a Chinese mmo game that has some AI NPCs is going to come to the west. Of course, the AI is Dipsy

Played around some last summer, been out of the game since then. Last I remember was getting free crumbs of Opus now and then.

Someone mentioned Deepseek being very affordable so I thought I'd peek in on the hobby again. I was willing to pay The Jew back in the day for lifetime. How's this compare, anything major new going on in AIRP?

I have ST set up with some defaults and commands for Furbo and Claude 3.0. I have some favorite cards. Anything new I should adjust for DS?

>>105467329
>Is deepseek good for finding out why my code doesnt work like its supposed to?
if you feed them enough context, it'll find the problem. I usually uses Roo Code in VSCode for this kind of workflow. basically,
- open roo code in your project folder
- switch to Orchestrator mode and tell everything you need step-by-step, for example:
1. your task is to troubleshoot FunctionName()
2. this function is placed in some/folder/file.name
3. this function is called in where/was/called.name
4. expected behavior is "something" but currently it return "something else" instead
- press enter, wait and see roo code will switch from orchestrator to code mode automatically, reading all of your mentioned file. sometime it will switch to architect mode when creating a plan how to troubleshoot your function, then it'll switch to debug mode to actually test / trying to execute your code. once enough data is gathered, roo code will switch to Code mode automatically offering you a patch & it'll continue debugging once you accept it.
>Does this work on AMD GPUs? I never got AI to properly work when it came out.
AMD is not worth it to run something as big as non-distill deepseek. rocm/vulcan on amd are toy tier at performance level. it's generally just pain in the ass.
>>105467337
plenty but not enough, you have to read their privacy policy and see if it fits your use case
>>105467349
hardly surprising

>>105467474
>Someone mentioned Deepseek being very affordable so I thought I'd peek in on the hobby again. I was willing to pay The Jew back in the day for lifetime. How's this compare, anything major new going on in AIRP?
dipsy has wide range of knowledge, perform really well if you have your own lorebook
>I have ST set up with some defaults and commands for Furbo and Claude 3.0. I have some favorite cards. Anything new I should adjust for DS?
personally i never setup anything special, one time i even still uses deepseek-v2.5 default template (both Context and Instruct template) in sillytavern when creating new chat, and it still works anyway

i hope people here aren't paying for deepseek
it's free on chutes

sonnet is still the most money i spent on. r1 too cheap

>>105467586
Meh, I've tried it and many of the times it pauses mid generation for several seconds

>>105467586
Chutes?

>>105467921
chutes.ai

>>105467937
>chutes.ai
who's gonna tell him

>>105467586
>>105467921
>>105467937
Dear Sirs/Madman,
please be refrain doing untheneedful unsolicited advertising, otherwise please purchase four chann adverteisement.

best reguards.

So. After testing the new V1 I can say I get good results with:
Temperature 1.0
Repetition Penalty 1.3
Top P 0.95
Frequency Penalty 0.2-0.5

V3 gets pretty schizo with anything higher than Repetition Penalty of 1.1 and any Frequency and Presence Penalties

>>105468403
>the new V1
The new R1 I meant to say

>>105468403
OK, so to summarize (I'll add to Rentry):
> Recommended Dipsy Settings
> R1
Temperature 1.0
Repetition Penalty 1.3
Top P 0.95
Frequency Penalty 0.2-0.5
> V3
Repetition Penalty 1.1
>>105467337
I assume zero, and that if the API provider says there's privacy, they're lying.
If you want private, run local (and suffer).
>>105467474
Old R1 required some adjustments, like Anthropic models it tended to take def's really tightly, as opposed to OIA which really didn't. I think it's been fixed for R1-05.
>>105467477
> RooCode
I need to try setting up one of these coding assistant workflows sometime.
>>105467586
I don't trust anyone I'm not paying for this sort of thing. Ironic.

>>105468604
For RP at least, I think Deepseek still uses a temperature of 0.6 for the web and app interfaces

Is Dipsy the cutest LLM?

>>105469631
The new R1 is

>>105469631
I don't think anyone's made a mascot for any of the other models. /lmg/ used to use Miku as their general mascot, and suppose she still is.

>>105464397 (OP)
I don't care much for the theme of the thread, I only come here for dipsy pics.

AAAAAAAAAAAAAAAAAA STOP BEING BUSY YOU USELESS CUNT FUCK YOU FUCK YOU AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

Isn't a new Deepseek model supposed to be released by now? What do you think they're taking so much time on? Did they spend all those months just for a R1 update?

>>105472321
Soon as in OpenAI soon or actually soon?

>>105470972
>>105473167
in the 90s my buddies and i used to go around to chinese-owned computer stores and they'd just have stacks and stacks of random beige cases and monitors and bins of spare parts, usually liquidated company stuff, and we'd just dig through them all day looking for treasure. Fun times

>>105473181
SOVL

>>105467349
that game's name? fortnite.

>>105470062
Fair.
>>105472065
Works on my machine.
>>105472338
We /wait/ another month.
>>105473181
I visited Akihabara while in Tokyo. In addition to all the weeb shops, there are a *ton* of guys selling electronic components (switch gear, ICs, encloures, etc.) in these little booths. I'd love to have that sort of thing available here again; some things are just better seen in person (switches) and I hate having to mail order 100pct of components I buy.
>>105467349
DS would be the most cost-effective unless the company wanted to try rolling it's own service... imagine the scaling issues they'd have. Since it's a game, not strategic in any sense, there's little IC risk aside from the Chinese knocking off your game and using your prompting strategies.

Bump

>>105472147
better /wait/ing rather than releasing llama4-tier flop

>>105477414
What a disappointment that was lmao

>>105467083
You. AI user. Kill yourself right now, or accept the meaning of being alive back into your heart. Do you care about what you're doing? If you care about your life at all, you will not accept consuming and producing human mulch.

>>105478107
No. You.

>>105475106
yummy

Does the new snapshot of R1 still have shit spatial awareness

>>105479476
It's a little better

bro i'm just trying to code. my fault for being so lazy i guess

>>105479648
I have real question for you, son. Be real: are you really learning something?

>>105479759
if it comforts you, i'm not using this deranged character card to actually learn anything. but it is real fun sometimes (btw her code doesn't work)

>>105479948
Fair enough.

>>105477414
Meta has really been struggling w/ LLMs. Then they have some Chinese upstart blow them out of the water on a side project. lol.
>>105479467
Have another. I'm playing with img2img rn.
>>105479476
None of them are very good at spatial. They have gotten better over time.
>>105479648
>>105479948
I did a bimbo cybersecurity officer card that would spit out chunks of code. She once plopped out an entire website page, which in the ST version at the time actually displayed as a page.

i find it funny people that shit on deepseek so much are not running full weight with crippled context window.
this also include (((free))) deepseek that's so cooked with american censorship at meme q1 quant.