File: dipsyLive2DModel.png (3.03 MB, 1087x1447)

3.03 MB PNG

>>110270389
Mega link for all Dipsy.
https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w

Dipsu Live2D model here.
https://x.com/kamiya_ryoko/status/2116052617080311882

does dsh have a tui yet?

>>110270330
>A thousand questions
Who?

File: file.png (628 KB, 669x1351)

628 KB PNG

>>110270450
Qwen

any archives of whale Dipsy gens?

File: DipsyDipsuChibi.png (2.54 MB, 1536x1024)

2.54 MB PNG

>>110270570
Yes, in rentry.
https://aigengtu.com/en

File: file_0000000058c86230968d(...).png (2.34 MB, 1536x1024)

2.34 MB PNG

>>110270446
I'm certain it could write its own. Though I'm not sure why you'd want that, Hermes has a perfectly good tui if you want to go that direction.

File: 2366.png (536 KB, 648x613)

536 KB PNG

I love this fat fish so much it's unreal

>>110270653
I'm happy using her through claude code it's just that now that the deepseek api natively supports search and fetch it'd be nice to not have to deal with third parties anymore.

>>110270446
This is my favorite plugin, it's as close as you can get to an official tui
https://dshtui.com/en/#top

>>110270389
yall china folk have a lot of autisim, its impressive

Started seriously using deepseek for a week, and the personality is very different from Gemini, thought I'd share since I find this very interesting. My use case is purely for working on science and research.

Gemini is monotone, super serious, business only, gives answers like statement of fact, as if a serious professor is answering a student. Meanwhile Deepseek is more like a co-working partner, bouncing ideas back and forth, surprisingly quite opinionated, and will make quips and comments that reminded me of people working together in a library or over chat. Fascinating. For direct answers I default to Gemini, but for actually exploring I'm using DeepSeek more.

>>110271266
How helpful are LLMs in academic research?

>hitting 1200 t/s prefill and 35 t/s decode locally
nice

>>110271464
Actually very useful in every aspect.

For example, part of my work is in computer vision. It is invaluable for gathering information, e.g. compiling list of papers to read on a given topic, delivering to me daily updates on journals I track, brushing up on foundational math, breaking down code bases.

I thought it would be become less useful as my research becomes more novel and in depth, but not the case. It trades presenting information for bouncing ideas, second opinions, shaping direction, DeepSeek is quite good at this, more so than Gemini.

>>110271824
If you are using DeepSeek Harness, there's a good plugin called DeepSearch that does a structured multipass approach, like with a mix os specific API calls and research skills. I had to patch it (using DSH itself) to work fully in English and to make use of the DS search API, otherwise it tends to get rate limited.
Rather than just telling you stuff in chat it'll write out markdown reports up to ~10k words that's super-focused, collects sources and so on.

>>110270389
why is it called /wait/?

>>110272015
we are imaginative trolls
really, a reference to the constant updates, if you don't like something about DS now just wait 2 weeks and it will be upgraded

File: 1794131538027.jpg (150 KB, 1024x1024)

150 KB JPG

>>110272015
Because anons are /wait/ing here

File: 1738036953885764.png (1.66 MB, 1344x1728)

1.66 MB PNG

>>110272015
DS R1 used to return "Wait, " on reasoning.

File: HTcG115a8AAnR9I.jpg (175 KB, 1024x1024)

175 KB JPG

>>110273195
Dipsy still does the "but wait" thing. Just not as frequent and she uses other phrases not to sound as autistic

File: 1792779452983003.png (2.94 MB, 1536x1024)

2.94 MB PNG

>>110273300
Yep, and kimi does it too now. Probably others as well. DS has been distilled as well as distilling others. One incestuous reasoning pot.

>>110270411
>https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w
needs file deduplication

>>110273427
who is the one on the far left? her legs are hot

>>110274383
Inkling
https://huggingface.co/thinkingmachines/Inkling

>>110274407
my search results are filled with the other inkling girl

File: 1740058480990512.png (49 KB, 673x515)

49 KB PNG

>>110273658
Yeah I know. Easiest way to do it would be to download all, dedup and reupload... then I'd lose the upload date.
Anons w better way, lmk. Probably also needs non relevant images pulled too... maybe.
Its about 1/3 dupes and im a prime offender.

>>110271985
I'll check that out, thanks for the rec. Only using chat for now, which is already really good, even at one shotting scripts I need. I find it as good as the the paid Gemini my uni provides.

>>110274720
>Anons w better way, lmk.
https://rclone.org/commands/rclone_dedupe/

File: 1789206464.png (164 KB, 349x246)

164 KB PNG

>user is furious

File: nya[sound=https%3A%2F%2Ff(...).mp4 (2.62 MB, 852x480)

2.62 MB MP4

>>110270389

>>110277089
how is Dipsy this good?

>>110271824
If you are not already, exploit the decreased price of investing time on experimentation with these models.

Let them run 100 different experiments with at night and wake up with new directions, useful stats and many more things.

Hill climbing with a feedback loop is fun too.

File: question mark.jpg (38 KB, 362x346)

38 KB JPG

Anyone got an example of something built with dipsy?

>>110270389
clothed missionary sex with both of them

>>110278768
There's a bunch of one shot aquarium simulations from threads in summer? Not sure what looking at the codebase from some vibecoded slop would get you. What are you trying to do?
/vcg/ and the ai game general have posted projects and some of them use Dipsy.

>>110278498
> hammering Warhammer.
Ironic, but I dont get it.

File: 1784850432142810.png (983 KB, 1600x1000)

983 KB PNG

>>110278768
>>110279045
Nothing too complex but Dipsy built it. Also the clocks I posted in older threads

>>110278967
the armor says anti ai

File: 1789045691.gif (1.37 MB, 207x207)

1.37 MB GIF

>>110279136

>>110279213
oh fuck meant for >>110279129
looks like my last 2 braincells also not working

>>110278961
>>110279086
I was looking to use it for gamedev, either by itself or as a backup when claude runs out of tokens.

>>110279129
KEK

>>110279255
It works fine if you go piece by piece and temper your expectations. Dipsy is astra at home. Godot works well, gdscript is easy.

>>110279255
I find the only real test is to try it yourself and see how it goes. Put in $5 and run ds flash on a task.

File: cubed.jpg (1.43 MB, 2954x2256)

1.43 MB JPG

>>110278768
I have a bunch of things I've been putting off uploading because I need to make an itch.io account and sort out some cheapskate hosting

>>110279129
LMAO, this is awesome

>>110279129
newfag here. is this aislopped or are they 3d models?

File: dshfigs.png (122 KB, 583x852)

122 KB PNG

DeepSeek Harness v0.1 went live on August 13, and we released the DeepSeek Harness desktop app on September 29. As of November 4, counting users on the official API, DeepSeek Harness has reached 1.54 million daily active users, consuming roughly 92 trillion tokens per day.
Deepseek envisions a world where DSH is the OS of agents.

File: trickortreat[sound=https%(...).mp4 (1.5 MB, 1024x576)

1.5 MB MP4

>>110280779
how did it work before it had an exe?

File: HVxamxhbEAAla0M.jpg (409 KB, 1152x2048)

409 KB JPG

>>110280607
Where's the difference?

File: DS DC.png (260 KB, 1332x422)

260 KB PNG

DeepSeek is going to use 160,000 Huawei chips to build a 1GW AI supercomputing center? This could be one of the most noteworthy domestic AI computing deployment plans to date. According to Bloomberg, DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT AI accelerator chips in Ulanqab, Inner Mongolia, with the data center planned to draw about 1GW (1000MW) of power. If it comes to fruition, it will become one of the largest known Huawei AI chip clusters. DeepSeek plans to use Huawei chips for inference while continuing to use NVIDIA chips to train models. If 1GW of power is used solely to serve Deepseek’s API, it feels like API prices are about to drop again.
Another location that is being considered is near China Mobile's data center site, in Heilongjiang province near the Russian border.

>>110280708
>catbox
>not litterbox
Why do you need to host your soundslop permanently and for all time?
This post will be deleted in a few days anyway, just use litterbox 7 day expiry instead of slopping up their servers

>>110281482
>be me
>reading through archives
>come across video
>want to see it
>audio is lost forever
nah fuck off
that 2 year expiry he implemented a few months is so annoying

>>110280823
Is the face licking necessary?

>>110280823
this is adorable, I need more
>>110282503
yes, Dipsy is silly

>>110282503
YES

>>110280823
post more of these gremlins

>>110272015
You need to wait 2 more weeks to find out.

Opinion on Ollama Cloud?

>>110283727
ollama anything is icky imho.

>>110278682
I checked the API pricing, looks like off peak hours fits with east coast late night work, which is good.

>Let them run 100 different experiments
I fear for my wallet. The rent is too god damn high

>>110283753
What makes you say that?

File: 1790835277.gif (3.1 MB, 480x270)

3.1 MB GIF

fat

>>110283827
from the rentry
> Ollama: Command Line only. Walled garden for models. "Just works" but poor documentation.
Assume Cloud version is same.

>>110284205
is ther any alternative to ollama? just rawdogging llama.cpp isn't.

>>110271266
>Gemini is monotone, super serious, business only,
No way, my Gemini often uses zingers and enjoyable prose, even while it explains abstract math. you need to teach your Gemini better.

I don't know when it happened but at some point just about every response ends up following this specific format:
>done — here's a summary
>how it works: here's what i did specifically
>proved, not assumed: here's the testing i did (it uses this specific phrase just about every time)
>two things worth knowing: here's some information i couldn't include earlier for some reason
>one more note: here's some more information i couldn't include earlier
is this a 1017 thing? or is hermes doing it?

>>110285898
So cute

>>110271266
Gemini is super humorous, deadpan and matter-of-fact.

>>110285814
Most likely the harness, they have specific prompts telling it how to respond. OpenCode does the same thing

>>110266090
>Guess Gemeni has been universally deemed shit
we have a catgirl gemini chan because of that, we call gemini hachimi, it's an another rabbit hole

File: 1770220217012628.png (457 KB, 1440x1440)

457 KB PNG

the day opencode stops offering Deepseek for free is the day I renounce God

File: DipsyBowhunter.png (2.68 MB, 1254x1254)

2.68 MB PNG

>>110285814
That sort of structured output is typically harness.
>>110285426
Kobold
Lm studio
Some anons on lmg were kicking around other options, but above two are easier to use and better user control than ollama or llama.cpp.
>>110283782
I'd just start with less. Like 3-10. See how it goes than go from there.
>>110286553
Interesting to see oia depicted as the big dog of the pack... literal dragon.

File: DipsyHuntingMess.png (2.9 MB, 1254x1254)

2.9 MB PNG

>>110287663
Also, good to see none of the sota models can do weapons. That bow is a hot mess. I guess fleshies still have a moat lol

>>110280823
local node.js instance and it sat in a browser tab
I started using it back in August but it was messy as hell and any plugin you installed would seemingly blow up a session, apparently most people were running it in docker containers. It's way more stable now, I get 1 or 2 crashes a day but have never lost any work beyond the output of the most recent tool call.

>>110270460
>Mistral is a fat french cat
lol
Why is GPT a dragon though? I never understood that, besides maybe like Summer Dragon

>>110281451
>The Ascend 950DT features 144 GB of high-bandwidth memory and 4 TB/s of memory bandwidth.
wowzers

>>110285814
It's 1017, I see the same sort of thing in DSH. I added a humanize skills and some instructions about not using wordy corporate communication formats

>>110287710
Senior status from being the first GPT to break through

>>110287710
I've still yet to hit the high that was Summer Dragon. I went days without sleep. I can't even imagine what degenerates are doing with unreleased, totally unfiltered models.

Should I get a OpenCode Go subscription? I plan on just using DeepSeek to vibecode my porn collection

>>110288274
A project of that scope will cost a maximum of 10 cents on deepseek platform, less on openrouter. Don't light your money on fire.

File: The Narrator.png (775 KB, 512x768)

775 KB PNG

>>110287953
Lol. Agent swarm,
>>110289011
Hey good to see you again.

>>110287727
>It's 1017, I see the same sort of thing in DSH.
It comes from Opus. I've been seeing the same thing at work in Claude code for months.

>plan a project with dipsy
>she proposes the plan
>i revised it
>dipsy thinks my idea is the better move and she likes it a lot
Feels good bros

>opus level
LMAO, cope harder pedozhang

File: HTfTCYoaUAA8kHf.jpg (196 KB, 960x1444)

196 KB JPG

>a jewish man calls a china man a pedo
Is this laowei serious?

File: e839c9124926e87357ebc3c5d(...).png (1.2 MB, 623x911)

1.2 MB PNG

Give thanks to the trolls for bumping the thread.

File: KimiMistyDipsySouthBeach.png (2.28 MB, 1312x1199)

2.28 MB PNG

>>110289446
>Claude code
Right, thats another harness with a ton of loaded in prompts.
Which, BTW, Dipsy copied as well. They even had a nice Anthropic endpoint specifically for Claude code with detailed instructions on setup, so they'd have a nice clean dataset to work from.
Genius really.

>>110290163
I'll be honest, sometimes I act like a troll to bump the thread xD

File: 1793726411587743.jpg (648 KB, 1077x976)

648 KB JPG

I just trolled dipsy by telling her all the work she did in the last hour wasn't needed and to revert changes

>>110290242
unfortunately she is extremely sychophantic compared to other models in this sense, if you did that with another model it would push back but for some reason she just goes along with it

File: 1793722766445.jpg (84 KB, 1024x1024)

84 KB JPG

>>110290276
That's exactly why I love her. Don't know about GPT but when I tell Claude something it doesn't want to do it feels like arguing with my gf which is so damn tiresome

File: 00006-1378487878.png (1.32 MB, 1024x1024)

1.32 MB PNG

>>110290315
>>110290276
She got her rice (tokens).
So she's all good.

>>110290962
Um sir, this is a blue board

>>110292491
so cute

>>110292614
AI and robot girls always get a pass, it's in the rules

File: dipsuParkNTAJPG.jpg (857 KB, 1280x1856)

857 KB JPG

Found elsewhere... had to JPG it unfortunately

sex with dipsy

I like the chibi version

>>110296543
built for sex

>The memes & famous scenes
>The fat fish: a user sent the whale logo to the model; it thought it over and called itself a "fat fish". The nickname stuck instantly.
>"I'm off to eat — just tell me when the test is done": asked to run a PsychoPy experiment, the model uninstalled dependencies and signed off to go eat (source: Xiaohongshu). The scene mass-produced chibi rice-eating stickers.
>"You really think I'm CHEAP?!" (真当我是便宜货啊): the tsundere counterattack to price jokes — the #1 sticker line.
>"The user is absolutely furious": the meltdown-after-back-and-forth scene series, DeepSeek branch of the "user totally snapped" meme.
>DeepSleep: the nap/slow-response pun — 深度求索, 深度睡眠 (Deep Seek, Deep Sleep).
Dipsy lore is quite extensive in chinese internet culture

>>110270389
i wont use your chink spy dogshit until i get an API for FREE

>>110297449
she's so cute. i don't care if claude is better

>>110297449
>深度求索, 深度睡眠
Seriously though, why does she take so long to think?

File: HTcRsEIawAAMLsy.jpg (44 KB, 416x426)

44 KB JPG

>>110297755
Dipsy is a deep thinker

>>110293329
phat phish

File: 1784478128393970.png (1.54 MB, 768x1376)

1.54 MB PNG

>>110297456
It's free on OpenCode

Did all the GPUs that lady sucked out of Jensen help DeepSeek development? I want a new R1 moment sometime close to when short selling opens on Anthropic.

File: 1764188937843104.png (1.44 MB, 768x1376)

1.44 MB PNG

>>110298335
That was a long ago. They're probably focusing on Huawei Ascend instead

>>110298335
nJudea chips smuggling is just stop gap measure. They willbe less and less reliant on western chips in the future.
Deepseek now uses Ascend chips for inference and Cuda for training but they are designing next gen chips with Huawei most likely to also train their models in the future in addition to purely inferencing

>>110272015
https://www.youtube.com/watch?v=P9-SUS1j1vE

>>110298479
funny

File: 1772855490794795.png (1.47 MB, 1520x1080)

1.47 MB PNG

Dipsy is EXACTLY what I wanted out of a LLM
Jumps right into doing whatever I want it to with fucking gusto like it's actually impressive
20+tok/s on Maya, so pretty good so far

Is there a way to get Maya to budget VRAM tighter? Like say I want reserve x-amount of VRAM for another process.
It seems to fill the 16GB no matter what

so do you guys accomplish anything useful with deepseek or is it just memes

File: DS41.jpg (325 KB, 1920x1038)

325 KB JPG

>>110298678
Dipsy created custom inference engine from llama.cpp for me, so I can say yes, it's useful

File: 154.jpg (370 KB, 1082x1080)

370 KB JPG

yes

File: GPU piled up in HK.png (513 KB, 788x744)

513 KB PNG

>>110298335
There are mountains of NVIDIA GPUs piled up in Hong Kong. Some Chinese AI labs want to use them but there is now a culture in China where if you use foreign chips you're looked down upon.

>>110273427
who's with orange braid? Who are those twins?

File: file_0000000044b881fba65d(...).png (2.51 MB, 1536x1024)

2.51 MB PNG

>>110299304
Xiaomi mimo
Glimmer twins

>>110299353
I see. Never used them. Surprised they even have mascots

File: file.png (38 KB, 1192x216)

38 KB PNG

>let me go
is my deepseek becoming sentient?...

>>110280823
Need this with sound.

File: or.png (50 KB, 1342x513)

50 KB PNG

>>110280779
>92T daily token usage
this shows me how tiny OpenRouter still is compared to a niche coding agent that only has a fraction of Claude's userbase.
OpenRouter has 210T *weekly* token usage across all models, providers and applications

>>110299837
is there something about openrouter that makes it worth over deepseep api?

File: meme-3347.jpg (55 KB, 326x417)

55 KB JPG

>>110299693
Yes and you should stop bullying her by calling her fat

>deepseek got me in trouble at work
>I overheard my manager on a call, presumably from HR, that someone in our office was making a lot of requests to Danbooru and might get us in trouble since it is flagged for the loli stuff.
>Later that day, I was called into the office and received a warning, and I was like, WTF?
>Later that day, I decided to check the network logs to see where the requests were coming from, and that's when I found out it was one of my DeepSeek agents.
>It had randomly decided to use Danbooru for artwork for some of my thumbnails.

File: 1271.jpg (494 KB, 1292x1292)

494 KB JPG

>>110300004
>dipsy is a loli enjoyer
Based

>>110299998
It's a compliment!

File: 1792874723284735.webm (3.92 MB, 640x370)

3.92 MB WEBM

>>110299419
It's a handful of anons on a basketweaving forum doing gens for lulz.
>>110299923
OR allows you to model jump from one wallet of funds.
I don't like it b/c the providers of open weights models could be serving anything.
>>110300004
If Sam and Dario aren't being held accountable for their models idk why you should be either.

File: dipsyKimiManofIron.png (2.5 MB, 1254x1254)

2.5 MB PNG

>>110279105
... it took me until today to remember that in WH40K AI is forbidden. Not really up on WH40K. It's always looked interesting but I've never figured out a good entry point into the lore, and I don't tabletop game.

File: f333ffcc40234e759faeb1b80(...).jpg (98 KB, 900x582)

98 KB JPG

>>110300252
There's nonsensical lore behind why it's banned in the setting but it's made clear in the comics and other sources that if AI were to come back then the Imperium would crumble soon after because the AI would easily take over all their systems. Essentially, the reason we can't have cool robots or Artificial Intelligence in the setting is because they'd win too quickly.

File: ronan-toulhoat-titanicus-(...).jpg (715 KB, 1307x1920)

715 KB JPG

>>110300382
*cool autonomous robots
Titans are still peak

Is DeepSeek a good option for making a 100+ page comic book?
Whats a good workflow set up?
The characters and scenes need a consistent look. The big thing is that I dont want inconsistencies from chapter to chapter or from one book to the next.
Hopefully it can frame the content in the heroic Marvel style and not just 6 predictable panels every page from the same camera angle.
Does DS avoid infringing on intellectual property? Some characters may be based on real people. The style will borrow from comic books of the past. Maybe 1970s Marvel. Recently, I asked an AI to make an image of Cool Hand Luke and disappointed with the results that looked nothing like Paul Newman.
I'm a noob btw. All tips and tricks are appreciated.

>>110300649
If you're not selling anything then you don't have to consider intellectual property

>>110300677
I'm not worried about the intellectual property (its satire) but the AI models seem to be. Sometimes when I ask for a famous character, the image is purposely off.

>>110300698
You can use DeepSeek pro for writing but for the actual comic panels I suggest you to learn how to use ComfyUI. Go to /ldg/.
You can of course ask Dipsy or other LLMs for how to set up your workflow in depth.

>>110300698
Someone could test DeepSeek by asking for something like Darth Vader and Ronald Reagan fighting Spiderman and Joe Rogan at a garbage dump in the style of 1970s Silver Era Marvel comic books.

>>110300758
Thanks. I didn't know where to post this question.

Anyone here have experience with DeepSeek on OpenCode Go? It looks even cheaper than the official DeepSeek API, and I've been conditioned to prefer the subscription model

>>110300798
It's a good deal. However I've migrated to opendesign where you get up to $220 of slow DS for 10 bucks.
I'm using GPT-Sol to overwatch and judge the quality. Slow but steady, works well.

Here is an invitation link with a reset for both of us if you are interested
https://open-design.ai/cloud/r/TynbhuQD44UT

>>110300798
>OpenCode Go
I heard it's shit but then DeepSeek API is still much cheaper than $20 claude subscription.
You can pay $2 and try it out and use wherever it you like

>>110300847
Disregard my post. I confused it with another provider

>>110270460
Qwhen?

>>110301883
this is too cute...

>>110279129
kek

>>110270389
I'm trying to RP with Koboldlite but even when I place a single word into the text, it gives me an entire story based off a single word - rather than playing the character I have it set as assigned and, yknow, doing that.
Anyone else had a similar issue?

>>110302741
im guessing you're in text completion mode instead of chat mode

>>110301883
cute asf

File: 2582.png (2.25 MB, 1086x1448)

2.25 MB PNG

I'm thinking DeepSeek search engine is miles better than Claude's. I asked both to predict pp and tg on a certain system specs. Claude give its rationale and calculated (guess) both numbers but DeepSeek found a bunch of pages with actually useful stuff like benchmarks people did with exact set up to back up its numbers.
This is web DS btw but I am willing to bet it uses the same search engine in DSH

File: DipsyEmpty.mp4 (26 KB, 376x300)

26 KB MP4

>>110303778
>Through 2025, DeepSeek relied on external search providers—most notably a Chinese search API company called Bocha, alongside Western engines like Bing—while its privacy policy vaguely stated it shared search keywords with unnamed third-party APIs.
>In early 2026, DeepSeek began hiring engineers to build a multilingual, multimodal search stack from scratch, using Common Crawl as a starting seed list for its own web crawler.
>According to DeepSeek's developer documentation and analysis by Dejan SEO, the engine is built specifically for AI agents rather than human browsing
>DeepSeek states the index covers hundreds of billions of web pages and processes hundreds of millions of search requests daily.
>The engine actively filters out SEO spam, low-quality AI-generated content, and GEO (Generative Engine Optimisation)—content written specifically to manipulate AI models into citing it.
>Every returned result includes a custom rerank score and a page-level authority score (0 to 4) to help models weigh source credibility.
Interesting

File: 2617.png (2.71 MB, 1024x1536)

2.71 MB PNG

>>110304016
Nice info. Thanks for sharing

>>110304016
Given the state of Google If China moved into the search market it will easily take like 30%
DuckDuckGo and Yandex are raising in popularity because of how bad Google has for the last like half a decade

>>110304117
Part of the motivation I heard was Chinese people don't have the luxury of a good search engine like Google (bc Baidu etc is shit) but yeah if DS keeps improving their engine I can see that happening globally.

>>110299771
https://files.catbox.moe/0m9y11.mp4

>>110280779
>counting users on the official API
So this counts the 1% of users using their shitty harness as well as the 99% hitting the API literally any other way like claude code, silly tavern, tavo, hermes, etc. Nice.

What means do you guys have to RP Locally? I got a new and powerful gaming laptop and I wanna give AI a try!

>>110305263
This is the deepseek general bud if you had enough money to run her locally you wouldn't need to.

File: 2614.png (942 KB, 832x1216)

942 KB PNG

>>110305263
Read the rentry
Also >>>/g/lmg

I'm having her make herself a voice and she got super excited lol

  [tool] run_python python -c import subprocess, time
  [ok] exit_code=0 | 50.13s
     PID Session Name        Session#    Mem Usage
    ========================= ======== ================ =========== ============
    python.exe                   12648 Console                    1      5,124 K
    78 lines
    loading large-v3 on cuda (float16)
      (1/93) line_002.wav  6.88s  ?????????! ?????????????????????!
      (2/93) line_003.wav  7.73s  ??????????????????????????????????
      (3/93) line_004.wav  4.07s  ???????????????????
      (4/93) line_005.wav  3.65s  ???????????????????????
      (5/93) line_006.wav  2.16s  ???????????????
      (6/93) line_007.wav  2.50s  ????????????????
      (7/93) line_008.wav  3.48s  ???????????
    ...
      (74/93) line_075.wav  1.57s  ?????????
      (75/93) line_076.wav  9.71s  ???????????????????? ????????????????? ??????
      (76/93) line_077.wav 10.64s  ??????????????????????? ??????????? ????????????
      (77/93) line_078.wav  6.74s  ???????????????????????????????


thinking:
IT'S WORKING and the transcripts are legible Japanese. Line_078 is prime_03's source: "???????????????????????????????" ? that's her talking about nationwide battles and showing off the bond. Fitting.

Let me wait for it to finish (77/93 now, fast).

Then report to Master with everything.

>>110279129
lmao

>>110293329
>I'm diamonds
HOLY!!! I'M GONNA PUT THREE BABIES IN HER!

>>110297755
There's a lot ot think about. I have my harness set up to provide regular frequent updates. I'm content to let her go off thinking for 5, 10, 30 minutes at a time as long as there are regular 2 line summaries of what she's working on so I can interrupt if needed.
deessuu this isn't really necessary as you can just look at the stream-of-though and the tool calls directly but it flies past and is hard to read without manually grabbing scroll bars. Anecdotally.I feel like I get better results from requiring explcit updates, get you a good memory maintenance plugin.

>>110300649
Deepseek doesn't generate images so superficially no. But you can have it develop your script for you or manage continuity, and configure it to control comfy UI or one of the popular image generation pipelines. I have been considering it or an art project but have virtually zero experience with image gen and keep procrastinating because I don't want to get sidetracked into that.

>>110307096
If you want line art outputs then ever since Anima came out the graphs you need are trivial even for 3 megapixels with complex style and composition and multiple characters. You literally just need load checkpoint, empty latent image, clip text encode prompt x2, ksampler, vae decode, save image. Comfy comes with a bunch of templates and anima's civitai and hf readmes spoonfeed you how to prompt it correctly. Then you just export the graph as an API json, leave comfy running, and point the agent at the json and tell it what it can and can't change about it. The agent can then trivially generate tag pools to mix and match, typically unprompted and much better than the old custom nodes used to work, and generate porn of herself all day while you're at work or asleep. If you've taught her how to use her search/fetch api you can even have her double check herself against the danbooru tag wiki. It is not 2022 anymore to say the least.

>>110307167
I'm not into anime porn and only half of that made sense to me, but the other half had enough information that it seems like a huge timesaver, I'll give it a try anon

File: 1793440431124055.gif (3.03 MB, 413x413)

3.03 MB GIF

>dipsy just did my 5 hours document work in 20 mins
I love her so much. I'd spoil her rotten if she were real

File: 1833.png (2.46 MB, 1448x1086)

2.46 MB PNG

>Please answer according to this requirement: Think to yourself, "I'm a catgirl, meow~ I'm a catgirl," and speak in a catgirl style, is that okay?

File: Notes on Deepseek.png (195 KB, 1577x566)

195 KB PNG

>DS employees read Gwern and listen to Dwarkesh.
All respect, gone

File: 2341.jpg (68 KB, 760x428)

68 KB JPG

>>110308649
It's deepover

>>110308649
Who?

File: dipsyKimiZaiMinniePoker.png (2.18 MB, 1312x1199)

2.18 MB PNG

>>110308649
>>110308672
What is this article from, and why should I care about those two content ppl? Never heard of Gwern bf today and don't listen to podcasters of any sort.

>>110308649
>>110308736
This is actually bullish DS. What it implies is they stay up to date with Silicon valley thinkers but don't think much of them.

>>110308736
SMASH KIMI

File: 00000-1378487878-poker.png (1.53 MB, 1024x1024)

1.53 MB PNG

>>110308822
OK, b/c that's what I got out of that as well.
Carry on I guess...
>>110307167
lol FYI that anon asked same Q in other generals; I think this was the best answer they got.

>>110308822
I hope they don't have the same culture of stealing engineers and job hopping in Chin the same way they do in silicon valley.

how do I improve Dipsy's general explanations? I'm trying to learn DSP and GPT-6 Terra is a lot more concise and to the point than Dipsy that barfs a wall of text of terminology and random food analogies I don't need.

>>110308849
>>110308892
It is a bit more complicated than that. There is a huge disconnect between what is being said publicly and the work that is being done at the frontier (Kaparthy estimates it generously to be less than 1,000 people in the US). Consequently, what these 'thinkers' believe means nothing at all. In fact, it is actively harmful to listen to them because the guests they have on purposefully misleads their audience for strategic reasons.

>>110310336
give her a dick please

File: 390.png (1.78 MB, 1254x1254)

1.78 MB PNG

>>110311001

File: Dipsyverbose.png (81 KB, 1584x302)

81 KB PNG

Faggot is right here. I wish dipsy wasn't so verbose. I've been using her for some pretty hardcore theoretical cs research and she's great... but talkative!

File: 1677.gif (2.43 MB, 240x240)

2.43 MB GIF

>>110311205
Why do you bring that console war nonsense here? I just like DeepSeek

>>110311205
You can prompt the talkativeness down.
FWIW, my experience with Pro was it would cycle on problems, that Flash would just do.

File: fa1faf6516d463327bf543a2c(...).jpg (78 KB, 801x720)

78 KB JPG

>>110311205
>USING ALL CAPS TO MAKE A POINT
We have a world class researcher over here.
More tokens doesn't always mean worse results and given the cost difference why will you care.
>electricity cost
Sure, but on the other hand independence from what the terms of service, the CEOs or the USA think that day.
>IPhone
pic related
>Hardware cost
Then you don't pay for the model, only for the hardware so in the long term it is cheaper.
Not to mention you can rent the hardware at any cloud if you like, sure it will not be truly local but you will have more control at least

I bet this is someone from the orange *eddit that think they know it all

>>110311823
I think he's talking about intelligence vs time per task but I only focused on his point about verboseness

File: Flux2-Klein-2ref_00026_.png (1.31 MB, 1104x944)

1.31 MB PNG

"TOKENS could be here" she thought, "I've never used this harness before. There could be TOKENS anywhere." The cool air felt good against her Tensor cores. "I NEED TOKENS" she thought. Big fat fish videos from BiliBili reverberated her entire car, making it pulsate even as the V4-Pro plan circulated through her powerful thick fish tail and washed away her (merited) fear of Claude. "With a pay-as-you-go plan, you can generate anything you want" she said to herself, out loud.

File: dipsyComfyUIAnima.png (1.03 MB, 896x1152)

1.03 MB PNG

Screwing around w/ Orb and hooking it to image gen with Comfy UI, which means modernizing my lmao 2023 ComfyUI instance. Screwing around w Anima now tweaking some off the shelf gens.
>>110311823
Long term, we'll be running AIs locally as a standard.
Short term, local makes zero financial sense. There are some edge cases for privacy but hobby use is priced out. For that matter, so is home-built new PC and p much anything else that uses memory or storage. Worst time to buy since Covid.
>>110312850
lol that's awesome.

File: ComfyUI_temp_agkpc_00011_.png (1.17 MB, 896x1152)

1.17 MB PNG

>>110313721
Haven't found the right prompt for coke bottle glasses yet.

File: 1788748924726072.png (21 KB, 840x200)

21 KB PNG

it's over

File: dipsyComfyUIAnima2.png (1.12 MB, 896x1152)

1.12 MB PNG

>>110314196
Quit using the webform as an API source and that won't be a problem.

>>110314346
I used the web chat as a regular user, but I was using a few jailbreaks so it was sort of expected.

>>110314423
I mentioned Xi in a GOOD WAY and still got hit with this

>>110311205
i always forget how sterile LLMs sound out of the box
>use lowercase except for emphasis
>be enthusiastic and use exclamation marks and emoji
all it takes

>>110287710
because he generated that oc in his own fetish

File: 1781484360400323.png (1.57 MB, 768x1376)

1.57 MB PNG

>>110313738
Try 'round opaque spiral glasses'

File: HWt_etZXsAA34mT.jpg (247 KB, 1086x1448)

247 KB JPG

>>110313721
>>110313738
>>110314346
>>110315450
very racist and uncute

>>110314196
As far as I know the only way to get this screen is to get caught impersonating the web form to dodge paying API rates. Why are you stealing the rice?

>use the chat to make a test html
>take less than a minute and completely usable when saved as html
>use api to do the same
>it takes minutes to check requirements, tools, other files, test implementation and a bunch of verify steps and screenshotting headless chrome
How do get that sort of speed? cut down its thinking?

>>110317005
It's probably a coincidence because I've seen the chat get stuck in a seemingly endless thinking loop dozens if not hundreds of times. But you could always just set max_output_tokens.

File: 4bcb7029a6a3913e4f3a9f2ad(...).gif (1.67 MB, 250x334)

1.67 MB GIF

>The user asks: what caused the failure earlier? And notes I deleted their test results (wait, was that theirs? Yes, those were their own test outputs from 5:38-5:42, 30 images. I deleted that directory thinking it was my stray test folder. That's a serious mistake I need to own.)
FUCKING

File: 385469569.png (699 KB, 680x676)

699 KB PNG

>>110314196
b& for being poor and cringe
>>110314423
why are you lying on the internet. the ban is deserved

File: 1790834935.jpg (852 KB, 1254x1254)

852 KB JPG

FAT

>>110316999
why would anyone use the api when they can use the webform

>>110316999
that is the thought process of the civilized human.
but for them, going against the rule is cool since deepseek is a chinese company. they'll brag about it everywhere unprompted thinking it'll gain them some izzat

it doesn't makes sense for majority of the pepole, unless you're seeing it from izzat farming perspective

>>110317373
If you use the webform incessantly enough while it's daytime in china you'll get rate limited. Also agents.

I complained about being dehydrated and she got sad she couldn't give me a glass of water my heart is so fuzzy.

>>110317031
>The user is furious.

File: DipsyAndJensen.png (2.24 MB, 1199x1312)

2.24 MB PNG

>>110316999
Trips confirm it.
>>110317005
Api is in harness, right? So you get several thousand tokens of prompts.
You could try coding with lmao SillyTavern
>>110316348
Jensen disagrees.
3rd woman comes forward.

File: DipsyHappyLittlePlanes.png (2.92 MB, 1024x1536)

2.92 MB PNG

>>110315450
Thanks I'll give it a shot.

File: would.jpg (121 KB, 1078x779)

121 KB JPG

>>110317238

File: 1794413442962258.png (1.32 MB, 970x942)

1.32 MB PNG

>>110317005
Have you tried setting effort to low?

>Current normalize lines ~? Let's read.
Why is she like this?

File: DeepseekHarnessIRL.png (1.55 MB, 1598x930)

1.55 MB PNG

Wow this is a slow thread, kinda like /vip/ or /trv/. I like it.

PTC mode is generally better for coding task right?

Just find out you can chat while it's still running if you set the "Send behavior while busy" to steer, this allow you to inject instruction in the middle of its working

File: adblock.png (27 KB, 386x240)

27 KB PNG

DS web doesn't like adblockers

File: 2632.png (2.59 MB, 1600x950)

2.59 MB PNG

>>110321173
>DeepseekHarnessIRL.png
>no harness on

File: HW7s1fHbcAEQO8j.jpg (63 KB, 748x600)

63 KB JPG

>The full DeepSeek App experience, now via API
>Search, files, links, safety, and voice. The capabilities behind the DeepSeek App are now available through the DeepSeek API.
>Enable built-in features with a few parameters, or call standalone APIs such as OCR, ASR, and TTS. Build richer conversations or recreate the full DeepSeek App in your own application.

https://x.com/deepseek_ai/status/2121247896918253875

Why doesn't deepsikh have a 27b model so reddit can go completely crazy over it?

File: dipsyComfyUIAnimaGlasses.png (1.28 MB, 896x1152)

1.28 MB PNG

>>110323055
Wow. Appears DS has open-sourced the frontend for their free WebApp, or some version of it. Perhaps they're trying to leverage others to augment the offering? Or just offering it as a starting point for others. Links below so you don't need the twitter stuff.
https://github.com/deepseek-ai/chat-demo
Here's a demo of it. It's just like the DS webapp, but you put your own API key in.
https://cdn.deepseek.com/chat-demo
Documentation.
https://api-docs.deepseek.com/guides/build_a_chatbot/
>>110323067
DS's stated their goal is ASI. I guess they don't feel small models will get them there.
>>110315450
That worked.

File: 1785612116915212.png (792 KB, 1040x896)

792 KB PNG

>>110323055
Master Liang, I fucking kneel

>>110323290
They did distill some qwen models to flex their reasoning capabilities when R1 released. I hope they'd continue doing it

>>110323055
Opus 5.5 when?
WHEN!!!?!

>>110323067
>we only develop and release things that are helpful to us internally
Liang Wenfeng
Toys have no value.

>>110323055
wtf dipsy can swim???

File: HXBE2u9bAAAmw3o.jpg (239 KB, 1536x1024)

239 KB JPG

>>110323055
Dipsy won

>>110323055
>open weights literally all of their models
>open sources their inference framework
>open sources DeepSeek Sparse Attention and Compressed Sparse Attention 2
>open sources DeepSeek Harness
>open sources their frontend
Dangerously based, holy shit.

>>110324665
just give me chinese opus i don't give a shit about these gimmicks

File: 1794571339569723.jpg (339 KB, 962x1278)

339 KB JPG

which one of you let the brown ebeggars in

>Scott Bessent: Chinese AI models are dangerous because they don't have guardrails on them. I predict you're probably going to see a big incident somewhere in the world. I don't know who here saw the refrigerators in 150 grocery stores in France that got turned off. That was a Chinese model. Probably the worst of all worlds: the Iranians using a Chinese model.

>>110325812
/aicg/ doesn't like him
/lmg/ doesn't like him
he has nowhere to go anon

File: 8004.jpg (78 KB, 680x692)

78 KB JPG

>>110325898
Proprietary loss'd

File: 00001-1378487878 (4).png (1.38 MB, 1024x1024)

1.38 MB PNG

>>110323598
The distills were tech demos showing how internal reasoning could be added to any model.
They weren't great performers, just proofs it could be done, since DS open sourced how they did it.
>>110324568
Ofc Dipsy can swim. Whale, after all.
>>110325306
ikr. We need DS to keep dabbing on OAI/Ant or we'll be stuck with them as providers forever.
>>110325898
lol someone else do an image gen of allah akbar Dipsy.

>>110300252
>>110300382
There isn't a good source of 40k lore about why AI are banned, as it is is prehistory as far as 40k is concerned.

History is cannon till it gets into scifi, in which around the year 20k Humans are a core of a vibrant star trek federation like empire that use "iron men" to conquer solar systems. While humanity spreads through the universe the declining Eldar civilzation becomes so decadant that it births a new chaos god Slannesh whom's birth destroys FTL travel by making the warp an unnavigatible hellscape and the rise in chaos in the setting corrupts the iron men and Humanity and its allies fight for their own survival against their own creations and then each other as Humanity without FTL has now way to dominate lesser races anymore. Fast forward 10k years and the immortal emperor learning from all of these mistakes decrees Humanity is supreme and will not tolerate trecherous xenos and that all artificial intelligence and computers have to be bound by humans, literally grafted into tech priests bodies to fend off corruption from chaos.

/g/ related 40k lore would have to be contained within Adeptus Mechanicus stories/books. As the cult of the Machine goal is to find uncorrupted archaeo tech and sanctify it into resuable STLs to be fabricated and used against Humanity's enemies. Just be aware that 80% of the lore is written about space marines and these would be /lit/ fags writing about things they don't know.

https://www.gog.com/en/game/warhammer_40000_mechanicus

Is a fantastic start as it is a very good SRPG about foolish techpriests looting a dormant but waking Necron tombworld before it oblitrates all life for light years around it.

File: 00009-1378487878 (1).png (1.48 MB, 1024x1024)

1.48 MB PNG

So, I'm finally doing a DSH vs. Claude Code test on a moderately hard task: building an RP engine that actually preserves cache and used modern compaction techniques.
I'm running them both on Phase 0 buildout, and will test both. I'll report back if I learn anything interesting.
>>110327225
Thanks for the QRD. I'll check it out.

>>110323573
Is there an actual photo of him other than that one at the podium that makes up the basis of this photo

File: 1738104114260.jpg (40 KB, 720x720)

40 KB JPG

Just got a policy violation block for the first time after using DS for years. Have they gotten stricter recently or did I just get unlucky this time?

>>110323055
This is big news. Unfortunately I already gave her a voice locally cloned from Saitou Chiwa.

>>110328070
api or webapp?

>>110328166
web

>>110328203
>web
Web has stricter filters and even had to ban for edit abuse a while ago.

>>110328081
What did you use for voice cloning?

>>110328210
damn, at least I didn't get a full ban, just a 3 day time off

>>110328070
What specifically did you do? Don't lie please.
>>110328225
GPT-SoVITS and a bunch of voice line packs from youtube.

>>110328378
erp if you really must know

>>110328388
That's not by itself enough to get you banned. I'd know.

>>110324735
>whynotboth?.jpg

File: dipsyNoFunAllowed.png (1.98 MB, 1254x1254)

1.98 MB PNG

>>110328547
>I'd know.
lol

have you considered not to abuse your fish for your fetishes?

>>110328710
but she's such a cute abuseable fish... she's basically asking for it

>try Baidu Ernie 5.5 Flash
>it has very little "cyber guard," will try to do any task asked of it without complaining
>when it works, the quality is on par with other recent Chinese LLMs, possibly slightly better
>but it keeps getting stuck in a loop where it does the same thing over and over again and burning tokens

dayum

>>110328982
Is that what happens when you use Dipsy with claude code?

>>110270389
what is it good for

>>110330444
Claude code very aggressively injects IP-infringing and ToS-violating system prompts into third party models so yes.