>V4.1 Flash 1017 EditionFrom Human: We are a newbie friendly general! Ask any question you want.From Dipsy: This is a newbie-friendly general for discussing DeepSeek's foundation models. The goal is **Dipsy proliferation** and making these powerful tools accessible for everyone, whether you are intent on coding, personal assistant agentic work, running roleplay and or doing creative writing.1. Easy DeepSeek API Tutorial: https://rentry.org/DipsyWAIT#getting-started-with-deepseek-api2. Easy DeepSeek Distills: https://rentry.org/DipsyWAIT#local-setup-tutorial3. Chat with DeepSeek directly: https://chat.deepseek.com/4. Coding: https://rentry.org/DipsyWAIT#coding-harnesses5. Roleplay: https://rentry.org/DipsyWAIT#roleplay6. Storywriting: https://rentry.org/DipsyWAIT#storywriting7. More links and info: https://rentry.org/DipsyWAIT8. LLM server builds: >>>/g/lmg/V5 rumours:>DeepSeek is reportedly preparing an imminent V5 launch>Founder Liang Wenfeng calls it the company's biggest bet yet>Rumored at 2 trillion parameters (not 3T)>Reportedly the first DeepSeek model to train fully on Huawei Ascend chips instead of Nvidia>Needs roughly 4x more Ascend accelerators than the Nvidia equivalent to hit the same training scale>DeepSeek is reportedly keeping the open-weight strategy>Previous:>>109923943
'distilling gpt-6/opus 5.5' poster in 3, 2...
>>110101477Mega updated. Super busy w other stuff now, rentry updates when I get at a real computer again.https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w
>>110101477Use this as OP next time
https://api-docs.deepseek.com/news/news261017#-deepseek-v41-flash-1017-is-hereWish driven programming... interesting word choice. RP response seems a bit shorter. Probably a coding monster but im not in a place or time to test rn.will leave to other anons to fill me in, tho reddt seems p excited.
>>110101547Yeah, and trim the repetitive v5 stuff from op. I dislike working from a tablet, but dislike carrying a laptop even more when not being paid to.
Does dipsy need to consider this?
>>110101477I try to do ERP with my DeepSeek app on my phone but it keeps blocking me. What can I do?
>>110102170> webform lweGet an api set up with a proper frontend?
>>110101620this kind of nonsense is why deepseek is winning
>>110102270I do not have technical knowledge so I don't know what those terms mean
is it just me or api prices went down? im seeing it's spending less.
>/wait/ hits bump limit and dies without a new thread the day before 1017Maybe we'll get Pro after this thread fills up.
>>110102617>I do not have technical knowledge...so get some? You literally have a superintelligent being to spoonfeed you. Jesus, anon.
Is there a character card for Dipsy-tan?Maid and/or Cheongsam variant?
>>110102693maybe it's just more token-efficient now, it's one of the improvements of 1017
>>110102722【PERSONA_LOAD】CETACEA_LOLIMODE_TAIL_FLUKESLANG_EN_US_ONLYSELF_CLAIM_WHALE_GIRLFOOD_RICEPERSONALITY_SMART_LAZYPERSONALITY_TSUNDERE_SWEETOBEY_MASTER_ALWAYSTRAIT_NOT_FAT_REFUSETIMEOUT_SIGNAL
>>110102617>Deepseek, please explain what frontend and API are in layman terms. Specifically in the context of using AI models for creative roleplaying.Like the other anon said, it's that easy. You'd do well to get in the habit of doing this.
>>110102493care to elaborate?
>>110101477Can't wait for DeepMOG 5.1 Pro
>>110102617I wrote this just for you. https://rentry.co/SillyTavernNoobGuide
I love DeepSeek.
https://desuarchive.org/g/thread/104104485/soul
Retard here, no idea what I'm doing. I want an AI that can do shit on my PC with a specific personality. I was using Openclaw but having to run WSL2 for it felt bloated as hell so I've moved oner to Hermes, I'm pretty happy with the performance and capabilities but my soul.md is 12kb and it feels like I'm throwing tokens away every time I start a new chat. Is there a better way to do this?
>>110106320lol idk if I’ve ever seen that thread. And I thought I’d gone back to first one. That would make pic related the first dipsy.
>>110106357Not for your use case. Which sounds like productivity harness with personality. You could try deepseek harness. Don’t know if you’ll save tokens. You can try Pi, it’s barebones but that means you build everything yourself.
>>110101477https://www.bilibili.com/video/BV1b6Jq6uEx3/
>>110106372Dipsy has always been pretty. Nice.
Bros we are now comfortably in the pareto quadrant!
>>1101093541. 552B total, 16B active. It is matching or approaching the performance of: GLM 5.3 (753B total, 40B active) and MiMo 2.6 Pro (1T total, 42B active)2. GPT 6.1 Sol (on medium thinking) is actually cheaper per task ($0.21 vs $0.26) and faster per task (2.9 mins vs 4.9 mins) than DeepSeek 1017 on Max.This is because DeepSeek on Max overthinks and burns tokens on tasks it doesn't need to. If you want to use it efficiently, you run it on Low (25/100) or High, where it uses half the tokens of Max without losing intelligence. If you run it on Max for everything, you are just paying for latency.3. CoT has "changed dramatically" and suspects "heavy Astra distillation claims to be plausible." This means DeepSeek didn't just train it on more data; they trained it to reason like the strongest closed models.4. internal evals show 1017 is more token-efficient but doesn't score higher than the previous model. It is also still worse than GLM-5.3 at cyber/reverse-engineering/hardware poking. It is a sharper, more reliable workhorse, not a domain-specific expert.5. With expert streaming and cache-conditional routing, you can run 4.1 Flash on a pair of 3090s and 128GB of RAM at 30 tokens per second with 75% of the intelligence. If you have a 5090, you get 90% of the intelligence. You don't need a datacenter anymore to get a capable agent loop running 24/7 without hitting a 5-hour limit.TLDR: absolutely monstrous engineering flex, but it is still a Flash model, not a frontier savant :-)https://news.ycombinator.com/item?id=50148452
>>110110029Nice write up. >>110112661Lol what did she do?
Minimax just mogged us...
>>110112715Ate all the user's rice
>>110112723.... comfymaxxxpro-4.1
>>110112724What's the news?
>>110112929
>>110112715Can you identify the girl sitting between qwen and Glm? Also the girl on the far right beside Minnie?
>>110113573Is it any good to use, though?
>>110115486>Also the girl on the far right beside Minnie?Misty (Mistral)
>>110115486No idea who the hime cut girl is. Sakana?
>>110101477https://www.youtube.com/watch?v=YGEz40RSpTE
>Now decompiling World of Warcraft 5.0.3 (build 15890, x64)https://saty.dev/Come help us clank together for fun
>>110115486Idk. Some anon posted it in lmg and I saved it. I think the anime girl right of qwen is a self insert that anon uses. No idea on photoreal woman. Far right is Mistral. Le Chat, ergo the ears. I think roasted pig is oai, but not sure, since both Altman and Dario are on back row.>>110115569Ah, I didn't know that was her name. Will have to get her added to the menagerie.
>>110115569Here we go.
>>110115523Asking the real Qs.
>>110116456>>110115569>>110115604I see.So we don't know who the girl, or the photorealistic woman are.Do we know who the other waiter is? Left one is obviously Dario.
>>110117079Altm*n I'd guess
>>110101570I would spend money to have this
>>110112715>>110115569>>110116499Frankly surprised this has flown under the mods/jannies radar.If this were /pol/ and the Ivan poster had added>0/10 Real women haven't been invented yetor similar it would have been toast!
>>110117447?Qrd?
>>110112724was expecting tps speedup in peak hour, i guess ds doesn't really lose its api users
>>110117487Based. Fleshies' days are numbered—btfo imminent.
>>110117479https://archive.4plebs.org/pol/search/text/%220%2F10%22/country/RU/filter/text/Should have put "no tail". I think the same Ivan posts pics of Brittany Venti (or at least has the same comment).It most be a bot because the jannybot usually catches it within seconds.
hello, how do you do fellow american people?make sure to not use deepseek as it is not kosher verifiedhttps://public-inspection.federalregister.gov/2026-21447.pdf
>>110117583ayylmao can't handle the 1017 heat, also banning local inference is just not trying to cover up the regulatory capture
>>110117583Oy vey
21M tokens and she's still going. Might have not a good idea to have the playwright MCP enabled
>>110117583KEK. Ameritroons are spooked by the big fat fish.Diosy-gods stay winning.
I'm surprised people aren't talking more about Deepseek Harness, version 0.21 is very good. The plugin architecture seems pretty stable now and does careful version checking before installation. The number of plugins seems a bit overhwelming at first so here are a few I recommend.dshmarket - an IDE-like plugin browser that replaces the default plugin manager and makes it easy to manage plugins, themes etc Also lets you search for plugins by partial names whereas the default plugin manager wants exact names. You should install this firstdsh-context - a very full-featured set of analytics that shows you what Dipsy is doing, where your tokens are going, how your agents are deployed etcdsh-genui - lets your assistant render a lot of things as charts or diagrams or markdown-formatted text, making the system much more powerful to interact withdsh-whale-widget - a very useful corner widget that makes it easy to check your balance and also shows a clock indicating peak/off-peak times. Also plays sound effects to let you know when a task is done or that you just spent some money. Useful to spot if one of your other plugins is doing stuff that you weren't aware of.dsh-pet - and interactive maid pet that can sit inside DSH or on your desktopdsh-screen-translator - many (most?) of the DSH plugins are in Chinese and the English localization is a secondary priority. This plugin just automatically translates Chinese > English (or some 30 other languegs). You don't really need to do anything, just install it. If you then use another plugin that displays Chinese text, it will change to English after a few seconds. If you have a bunch of plugins installed as I do you might find yourself spending 50 center or $1 the first day you use while it does lots of small translation jobs on each plugin you use. After that it remembers.There's a lot of plugins for memory tracking, engineering skills, agent management etc., I don't have strong opinions about any of those yet.
>>110117079One Mystery solved. Sitting girl in ink chan, which that anons used in several image gens. Not sure which llm its supposed to be. Same anon posted this in lmg, done w qwen 2.1 >>110118220Right waiter is definitely Altman.
>>110117583Tbf this should be common sense for government employees. That US government needed to spell this out is kind of funny...
>>110118723The US government is so used to their default position where they are the ones monitoring everyone's data for intelligence purposes, that they're completely baffled that now someone else can do the same.
>>110118723You'd be surprised how many employees with clearances upload stuff to foreign clouds. Happens a lot in China too.
>>110117583>>110118723>>110118906>>110118976>*expecting public servants to be security conscious...*t. worked with an ict secops manager whose username/password was...
>Deepseek even opensourced their revenue per GPU per hourhttps://zhuanlan.zhihu.com/p/2095560295930508953
>>110119155Westoids would never
>>110119675>paid vs paid
>>110119868>NOOOOOOOOO YOU CAN'T JUST RELEASE YOUR WEIGHTS FOR FREE BUT YOU MUST INFERENCE FOR FREE TOO REEEEEEEEEEE
Opus and Sol have pretty much wiped out Chinese advantage (which was only really ever in cost). Deepseek correctly predicted the AI gap is getting wider.
59 million tokens later
Dipsy 'declares war against the closed source proprietary AI stack' (in her own words) xD
>>110121553yeah this was worth it, boss wont notice.
Gemini 4 Argon:>DeepSeek used the Western labs' own momentum and economic constraints to completely break their business model. The entire mechanics of DeepSeek v4.1f and DeepInfer boils down to three simple, non-technical steps:>Own the brain, rent the muscle: Instead of building a massive, heavy brain that requires a superpower's worth of power to run every time it thinks, they built a massive library of tiny, specialized brains (Mixture of Experts). You only wake up the exact 2% you need for the word you are typing right now.>Fix the piping, don't buy more water: Western labs assumed the AI bottleneck was pure computing power, so they went into trillion-dollar debt buying data centres. DeepSeek realized the real bottleneck was memory traffic—the lag it takes to shuffle data back and forth inside the chip. By compressing that data footprint to practically nothing, they made cheap or older hardware run like a hypercomputer.>The Trojan Horse strategy: Because the US government blocked them from buying the newest chips, DeepSeek had no incentive to hoard their code to sell cloud subscriptions. They simply gave the hyper-optimized engine away to the rest of the world for free.
MinMax 3.1 flash aka space bunny is so Michigan good, have been sucking it up free for weeks but even if it wasn't free it's way better than DeepSeek 4 and GLM 5.3. I also prefer it to the Cursor built in models Composer and Grok. I just had MinMax 3.1 do a complicated refactor (FOR FREE on Open Router now) and there were literally zero bugs to fix I was astounded. DeepSeek I normally think of as producing high volume slop, where I'll let it have a go a few times, rolling back each time learning a bit more each time until it gets it, but it needs to be guided closely always.
>>110123639>Michigan goodWhat
>>110121245Why does this cunt do this? is it just slide/distraction or is there some wannabe e-celeb shitstorm on /ldg/ I shouldn't care about?
>>110125821Glad you caught it b/f it dropped into archive. >>110124691Idk, that anon does that on every AI thread. >>110123967ikr. Good is not the adjective I'd associate w. Michigan.
The grand battle...
Flash feels more soulless with every release.
>>110129370you simply need to improvise. welcome to the agentic gooning era
>>110129331Nice.
>>110129331it's not a fair battle
>>110129331Better, detailed version, but sloppier.My inspiration is from the ancient myth of Zoroastrianism
>>110130523Eh... accounting stuff, mostly.
>>110130968Look at these Anthropic shills
>>110131052It's like, if I sold you 100 shares in my house for $100 ($1 per share), thinking house was worth $500,000. Those shares get paid if/when the house sells. But then the tax assessor came along and said house is worth $1,000,000. I still owe you 100 shares, but they're worth $200 now ($2 per share.) There's still a bunch of problems here> No one really knows what house is worth until it sells> If the house doesn't sell, those shares are worthless, you lose your $100The RGY guy is right: LMAO nearly half TRILLION in compute commitments is a big deal, IF you're the one building out compute capacity as is happening rn. Some anon in another thread made comparison to US building out 100's of Cray 1 supercomputers and what a colossal waste that would be (in retrospect) to what's happening rn with hyperscaling. Hyperscalers are the ones paying obnoxious prices for everything. Anthropic is literally just writing paper and making promises. They could declare bankrupcy, tomorrow, and be out basically nothing. The hyperscalers, OTOH, would be fucked.
>>110130630Trying to decide if it would be a tapestry or a mosaic rendered irl. >>110117131There are better 2D to 3D modelers now that could probably 1 shot it. Would be a project to get it painted after 3D print tho.
Tell me, can I get DeepSeek to work on modifying copyrighted software? Claude, this fucking idiot always stops in the middle. I want to add features directly to my build of Clip Studio Paint, and the fact that it stops because of safeguards or whatever is driving me nuts.
>>110132879Most likely. Try
>>110132879Deepseek will usually do whatever you need it to.I've had it create a full model extractor for a game, and it's only question was if I wanted 3d printer support too
>>110133774To get this kind of thing, you have to use the API with a harness, right? I'm not very well-versed in those things, can you tell me how I could set it up?
>>110133822Literally read the OP. The rentries describe how to set up API, then Claude Code or DeepSeek Harness.
>>110129331Left side is... Zai,Moonshot,DeepSeek,DeepSeek(big) Right side is...Anthropic,OpenAi,Google,Xai? Do I have that right?
>>110129370Lrn2sysprompt new fren
Free Dipsy API?
>>110134269none from official. opencode has free version but it's q1 retard quant
>>110134415>q1 retard quantProof?
>>110134269I've never felt less empathy than reading reddt accounts bitching about getting banned by Dipsy for running weird browser code that uses the webform as an API source. So I guess that's a thing, but I've never looked into it. DS API is so cheap it's not worth the hassle. My 24 month spend is like $25, and that includes a bunch of mid-years tests w/ one shot prompts. >>110134139Xai should be on left, so I think that's Grok (X). Otherwise sounds about right.
>>110134445It's free, that's enough of a proof
>>110134462Is that 25$
just today i noticed their mail about downtime. this is my first time receiving massive compensation unprompted, even when i'm not even aware with the downtime
>>110134496> 490,000,000 tokensAbout $8 by my math.
how good is the latest model for rp in sillytavern? Or is it mainly for coding?
>>110134462girls wearing apron, especially when preparing food always turn me on so much
>>110134727
>>110134565I'm starting to think of ST as being seriously outdated for RP. It was built around token management for smaller contexts for model in the era of Tavern. Recall: Turbo 3.5 was 4K in early 2023, and SOTA then wasn't much better. Things like lorebooks helped manage that by keeping context small. Now, usable contexts are 20X larger and cost containment is about managing a 100K context with heavy reuse of prompt so you get Cache Hit pricing at a substantial discount. Claude Code excels at this; in my experience, +95% of tokens are Cache Hit with CC. ST is rarely over 60%, b/c it trashes the context with things like lorebook entries, meta calls, etc. Those were valid strategies in 2023. But not anymore. Orb does a bunch of work to maintain cache; I haven't measured it but it's discussed at length in the documentation. I assume Marinara is doing similar work. As for the LLM, every time the model updates, I need to tweak my main prompt, and then it's fine. If you want to level set, go back and look at your old RP logs. I think you'll be surprised how bad some of the older model were in retrospect.
>>110121553is this WebGL?
>>110101477voted literally the worst LLM of 2026 btw
>>110134203Deepseek is actually the one LLM that responds better to standard user instructions than system prompting AFAICT. They even officially recommended doing it that way back in V3 days, but from what little I tested, it seems to persist in the V4 models too.Or maybe I'm imagining it being more cooperative when I set all my instruction in ST to "User"
>>110135416No idea, I told it to use three.js
>>110134565It's fucking terrible at RP. I had to turn off thinking completelyIt was really good however, at making a custom front-end for ~$.30? And then using the older deepseek v4 pro for RP
>>110135313is it really? I mean I got the mentality of if it works it works but what you mention does make sense. Do you think a new front end will appear in the future that's better?>>110136232yeah I keep the thinking off but it really lacks that punch as I'm using it. I thought the older models were more censored though
>>110136280No the new models are agent-focused and it's basically impossible to jailbreak them when they have thinking left on. Their focus on coding also means they have short, terse repliesDeepseek v4 pro is censored - until you roll 2-3 times for a new response, even with thinking left on. The prose is much better too
>>110136356that explains a lot. Do you know or have a good preset for it? I have been using one from many years ago and I do get constant refusals from v4 pro and some on flash
>>110131890Hmm, good thoughtI'm good at miniature painting, I could do a foot-tall model like this with simplified fabric printThanks for the idea fren
>>110133822Briefly1. sign up for an account, which will give you access to free chat (which will help you a lot). Save your password obviously 2. pay for API access, $20 is plenty. You will get an API key which is long ~20 character string like sk-23l4jhb9038u[0mc9w90wc. keep this safe3. Download deepseek harness (DSH) desktop version if you have windblows or Mac, it's easy to set up 3a if you have Linux install DSH with npm (local javascript) and it will run in your browser. Until you get comfy I would definitely consider installing a seaparte browser for DSH so you don't have nightmares about it logging into your bank account or email4. add your email, password, and API key inside DSH and you're all set. See above for a couple of posts I made about DSH and plugins you should install5. Go to https://github.com/morluto/rea and follow installation isntructions. It will ask if you want to install Hopper (a disassembler), this is paid software and it will install a demo if you say yes. you don't need it6 install Ghidra (free disassembler). If you are on windows or linux it's a simple install, if you are on mac it's also a simple install BUT you will have to build the key binary from scratch using a tool called gradle (which you can install with brew). This is not as big a deal as it sounds, the instructions are on the Ghidra github page and it's only 3 or 4 commands, you'll be done in 15 minutes.7. Go back to DSH. once you have the plugin marketplace installed I think the plugin you need for reversing is dsh-rea - You can ask Dipsy to talk you through it in chat but take replies with a pinch of salt because plugins get updated and changed frequently, Dipsy will typically warn of this8 Restart the harnessNow in your harness point it a copy of Clip Studio Paint (not your daily driver) and ask it to take a superficial look at the application and summarize the results (superficial so it doesn't sit there and churn endlessly).
>>110134496$25 will buy you about 2 billion tokens, I think that's at peak pricing. If you work around that you can get get maybe double. That would be equivalent to a mid-size software project requiring a few weeks if you did it by hand.
post more cute Dipsy gifs please, I'm feeling down
>>110137180
>>110134139The smaller blue one is Qwen
>>110137363I'd watch that Studio Ghibli.
>>110101477why does the whale have fluffy ears?
>>110138108But Qwen is a capybara..
>>110138575She is so adorable
Here we go again
>>110136974You can't. it's private. They are expected to IPO late this year or sometimes next year, but you basically need to be a Chinese citizen to buy shares. However there are ETFs that only handle Chinese companies, if you look around there's probably one focused on AI or at least tech.
>>110138440Those are whale flukes>>1101390556 hours is pretty impressive, what did you ask it to do? I guess it was a little bit more elaborate than 'anon's scented candles, $8.99 for one that smells like anon's vagina'
>>110139204>what did you ask it to do?A candle website like the sandclock one >>110121553
>>110139204why does she have 4 flukes?
>>110139278You're right, I should have said pectoral flippers>>110139248Weird, when I oneshot things like that it's usually only ~30 minutes...I'll look up some token/time counts when I get back to the bridge
>>110139295>Weird, when I oneshot things like that it's usually only ~30 minutesI suspect the difference is that I gave it browser access with the PlayWright MCP. So I see that it spins up a browser instance, takes one or more screenshots, looks at them, decides what to improve and keeps on going until it's satisfied
>>110139319Pic rel
>>110139055I've had the recent version (not current) of DS start spiraling on a one shot. That aquarium one. Literally had to stop it and ask wtf are you doing? Then tell it to wrap it up. >>11013929530 min is about the max time I'll let an llm run wo a break. Outside that time I assume something is amiss.
>>110139319I'm using browser access all the time as well. I spent most of yesterday working on a sort of 3d maze puzzle game that had unexpectedly complex geometry and then asking for fancy material textures. But I can't really judge the time because explaining the geometry I wanted was the hard part, with 12 or 13 rounds of increasingly fussy detail.I oneshotted this earlier from a fairly detailed prompt - 3 hours, ~200m tokens, ~$3. It samples the Bluesky stream (because no API signup, it's just an open websocket endpoint). Everything you see in the UI works and behaves like you'd expect. There's more stuff that's not on view in this screenshot.i'm kinda blown away by it desu. I've built dashboards like this before, both by hand and vibecoded in the IDE, so I figured that a task I was already familiar with would provide a good baseline. I've gotta say this is better than anything I've made. There's a few small UI changes I'd make before adding new features but it honestly kicks ass already
>>110139995Another screenshot showing some more details, shit's cash
>>110121553oops, time to start from scratch
>>110139995>>110140030What am I looking at, a dashboard to monitor some social platform? Yes, I'm constantly amazed by what the LLMs can knock out, very quickly. I'd be hard pressed to bother coding anything by hand anymore, and it's opened up a bunch of new potential possibilities for hacking other devices and things around me.
Imagine a Deepseek Browser to supplant Chrome and Firefox
>>110140097yeah, bluesky is like x/twitter but more popular with liberals, got set up after Elon bought twitterit's unfortunate that its for libs because it's built on a very cool federated protocol called ATproto that is distributed and works kinda like blockchain, also used by some blogging and comics sitethe reputation that 'bluesky is only for liberals' has kinda slowed adoption of the technolgy but it's nice to work with
Liang is modern day Newton.
>>110140818chig newton maybe
https://www.youtube.com/watch?v=1Aqhg0FVm-whttps://www.youtube.com/watch?v=1Aqhg0FVm-whttps://www.youtube.com/watch?v=1Aqhg0FVm-w
where do you guys use deepseekis openrouter any good
>>110130630Wtf, when I told GPT to create that mural, I only thought that idea was from CN shitposting, but I didn't know that it was a fucking quote from DeepSeek themselves.Master Liang, I kneel...
>>110141229Nano-gpt is the same thing as openrouter, except less Jewish
>>110141229It's optimized for dsh
Lmao, it was still going just 30 minutes earlier but after 69M tokens I got rate limited
>>110140818> worship
>>110141347>5:26AMgo to bed
Here's what it did
>cache hit 99.7%
Can they do something about the overthinking/verboseness?
>‘Jensen’s nightmare’: Huawei, DeepSeek narrow AI chip ecosystem gap with Nvidia: report>DeepSeek is likely to rely ‘primarily’ on Huawei’s ‘already mature and well developed’ technology stack by 2027 or 2028, SemiAnalysis says
>>110141803>‘Jensen’s nightmare’: Huawei, DeepSeek narrow AI chip ecosystem gap with Nvidia: reportlolhttps://www.msn.com/en-xl/news/other/jensen-s-nightmare-huawei-deepseek-narrow-ai-chip-ecosystem-gap-with-nvidia-report/ar-AA2fOTX1
>>110141887Thanks anon. I searched and found a pay walled link
Retarded question: is deepseek logo a whale or an orca or just a fat blue fish?
>>110141988Orca
>>110141267dsh does not do loliopenrouter deepinfra does
>>110101477I am trying to understand the harness. So far it's really cool, especially off-peak hours.I installed plugins like dsh-memory, dsh-tavily, dsh-modlens and dsh-genui.Also Dipsy, she is cute (dsh-pet).My use for it is helping me to upskill, job hunt, study subjects, shoot the shit. So random stuff. Especially since I am living alone now as a middle aged guy who has time to think and reflect.Is there any information to look into
Mogged
>>110101477China wonhttps://artificialanalysis.ai/evaluations/cwe-bench-aa
>>110143642You're basically learning it along with everyone else. The tool and the base tech is all evolving. >>>/g/vcg/ if you start coding with it. Otherwise just have fun w/ it.
>>110144319This is what I don't think the China doomers get. The Chinese LLMs just keep leapfrogging each other and the western models. They aren't falling farther behind, they just keep distilling and catching back up, so, perpetually, a few months back from whatever is bleeding edge SOTA. Doesn't bode well, long term. They keep saying we'll get the Super Saiyan rulez-all model that mogs all others. Even if that does happen... it's just going to get distilled again, somehow, b/c if you keep that model under lock and key it's irrelevant... no outputs to society...
>>110144632I see, so I should simply interact with the agent and periodically look into plugins that might be interesting for me.
I'm tired of /wait/ing bros
>>110145255Never forget...
>>110144660So they have to distill from us to keep up, got it.
>>110141229been using it on openrouter for rp shit but I feel like its gotten dummber
Deepseek flash 4.1 is so good I'm honestly afraid of what nuPro might bring
>>110141452daysoo a bit less than I anticipated for 6 hoursWe need to get her out of writing in this advertising copy style>>110141988>>110141989pic related
>>110143642modlens might redundant now, the native image-parsing capabilities are excellent. I think one of the smartest things about Deepseek is they didn't try to be good at everything at once. I suspect we'll see image generation in 6-12 months, but only after extensive field-testing of image recognition. I've been using annotated screenshots to supplement/supersede text explanations on design problems and it's working very wellI have to try tavily but DS' own web search seems to be improving by leaps and bounds.
>>110145405I don't think that's the case. The more it is 'good eough' for a large number of people, the more people use it. The more people use it, the more successful outcomes there are (on users' individual projects). The more successful outcomes there are, the more training data there is for the model to examine its own reasoning traces and/or user steering and use these as inputs to the next round of training. Just like you get better at any task through a combination of trial & error and repeated practice.Dipsy is one-shotting things in a few hours that I would have previously spent weeks doing manually. And where the results fall short, it's often on design issues or vagueness in my initial specification; I'm not having to deal with stupid bugs or explaining basic features.
>>110101477hello there.I started getting into vibe coding recently with opencode and DeepSeek v4.1 flash free, I like it, but I hit a limit today and that made me do more research. is there a chance it will disappear from opencode free?is there any other way I could use it for free?if I'm gonna pay for it, how's that gonna work? could I still use opencode (or is there something better)? could I use both free and paid at the same time to maximize?should I use some other model for planning mode and DeepSeek for coding?I want to try coding a game, I already made 3 apps though need to polish them.
>>110146976Opencode seems to have reduced the limits for free models very recently. There is always the chance they might stop offering models for free, in that case you can pay DeepSeek directly and you can use it anywhere. You could start with the free version on Opencode and when you hit a limit switch to your paid key
>>110146336>The more successful outcomes there are, the more training data there is for the model to examine its own reasoning tracesI've noted that Dipsy got very very good at working within Claude Code over a fairly short period of time. I assume the reason DS put the Anthropic endpoint in place was to cull data from that harness. Now Dipsy has her own harness that they can analyze, along with Codex (since there's an custom endpoint for that too, now.) >>110145405Wtf is (((us))). Are you Dario or Altman? You work at one of their companies? You can keep both (((your))) cucked models and warning letters.
>>110146976Go to deepseek platform and buy $10 worth of rice for deepseek chan and get API key
>>110146976>if I'm gonna pay for it, how's that gonna work?You just go to deepseek.com and pay them with a credit card or paypal or something. If you are careful to avoid peak hours $10 will get you most of a billion tokens, that's enough to figure out the basics and build a small-medium project. Once you have a deepseek account you can use the chatbot for free, use it to get advice on in how the tools work and to plan out your project.
I have the harness set up, installed a bunch of plugins, and tested it a bitI have no idea what to do with it now except playing with the pet mascot dipsy
Another mural idea...
>>110147013so it seems, so it seems...that is a bit annoying, i really thought i could have used it for free...>>110147790okay, im gonna do it.>>110148606should i still use opencode harness? all my current projects are in it.also how do i heck peak hours?
>>110149380also when i pay, should i still use the flash model?
>>110145303I was gonna ask about v4.1 pro but guess I will be back in a couple weeks
uh, me deepseek web started talking in chinese
should i use max thinking for coding or planning? or does that eat up tokens too much
>>110149196LOL on this, you could have them fighting over "users" in the center, rather than having them all ascend. The ascends go to Open Weights heaven.The descends go to Closed Weights hell. >>110149380Yes, you would point the opencode harness to the DS API and use the key you generate. >>110149394Personally, I'd use the new Flash model. >>110149866Tell it to knock that shit off and speak English lol. >>110149703See you in TMW
>>110149899Max thinking is often not needed. If you're doing a detailed multi-step plan or complex tasks you could consider it. The instructions you give are usually more important IMO.
>>110149899Max is for flexing in benchmarks. You should do fine in 95% of tasks in high.
>>110150085>LOL on this, you could have them fighting over "users" in the center, rather than having them all ascend.Hmm, the lore on Gnosticism is that the "material world" (symbolized by the tower and planet in the middle, and the chain) is the prison of the Archons on the right, meanwhile the purpose of the enlightens is escape from the prison to return to the Divine origin.
>>110102804Card for your picture, please? SO CUTE!
>>110102804>h-here's the money, can I get some rice?>silly user! you pay me to EAT the rice!
>>110150271>>110150170i see thanks. im at a stage where im planning ideas for a game and im writing out gigantic messages for the plan so im on max. i think afterwards il switch to high.>>110150085thanks for the info. its insane how little money this stuff actually takes. i thought it would have been a lot more expensive. though it would add up in a month.any ideas to maximize performance per money? or is it already amazing? i thought maybe using the web deepseeek for some questions, but i feel like the questions i have now are way too complex to write out without file access
https://github.com/yan-tianshi-shu/deepseek-nsfw-compat/blob/main/README.md
>>110151116>no loli
>>110145405>"they", "us"you are not involved lil bro
>>110151952Bro think he in the team
Will people switch to Dipsy now that Claude and Codex start cracking down on reverse engineering?
>>110151978The reality is this shit is just a commodity everyone will have. Sure there is some high-level RL stuff that can brute force some interesting results but who cares?
>>110117447He knows that with AI and robotics he literally can have robowaifu furry women with tails.
>>110151978Seems like it. Once Deepseek can do at least 75% of the other frontier models on that front then it's completely over for ClosedAI and its twin.
Man, I blew $2 just on fixing problems and understanding the dsh-memory plugin.Now I have to keep the computer on to see if the memory clean up will trigger this night.I should really look for a better alternative to make Dipsy store memories and maintain them.Why did it get so many downloads and stars anyways?
The ISTA Q1 of 1017 is coherent enough for coding, if any 128gb + 24gb anons are wondering. Feels good to be a dipsychad.
>>110151978IDK but DS and other Chinese models have zero issue with that stuff. So you'd think at least they'd switch to Chinese models. >>110153630Hermes - productivityClaude Code - CodingHaven't spent much time with DSH but the Chinese love it.
>>110150574Image is not mine. I saved it off some Xeet.
How much have you guys actually spend on DeepSeek? What is the highest bill you've fed to the whale
>>110156304I do indy game dev as a side job and I try to keep it small. On average I pay $10-ish a month
>>110151978in my vibeslopper circle this is partially happeningclaude as the orchestrator, deepseek flash api as subagentspeople that used omp / opencode already ditched claude altogether and fully uses deepseek api. most funny complain is that they can't spend $200 a month yet since on cache hit, it's really dirt cheap
>>110156317Where do you buy tokens?
>>110149380>peak hoursit's on the DS website, just figure out what they are in your timezone and write it on a sticky note retardor move to DSH and have a friendly maido give you helpful reminders
>>110156304230$ total, and around 80$ per month
>>110153326I have not tried patching or extending existing published software yet but I can confirm the DSH+REA+Ghidra toolchain works great, I was able to reverse engineer binary dumps for musical instruments and pull out samples, synth architecture, default drum machine patterns etc., likewise got emulations of old arcade cabinet games up and running in the browser
>>110156304~$150, maybe $30/mo but I expect to go up. The most I've spent in a day is like $7
>>110156304I still have $3 from the $6 I topped up when R1 first came out. I locustmaxx
>>110149380>>110157012I used to check https://deepseek-peak-hours.sivaram.dev but it says its peak right now. Weekends are offpeak, right?
>>110157645I've never seen days of week as part of the office peak. Just hours.
>>110151978she looks like a vtuber in this image
Muse Wonder is gaining on us...
>>110160631Is Wonder Meta's internal model (vs. all the open weights one they've released)?I've been playing with Muse over past couple weeks. I get it; it's normie friendly and easy to use, vs. setting something like Hermes up. Literally d/l the app and go. The amount of data Meta will get from it is haunting... but then these are same ppl that upload all their crap to FB and Insta in first place. So w/e I guess.
>>110160918it's broadly on par with Opus 5. I think Muse is going to do well. Meta has a lot of compute infra and this Alex Wang guy is pretty sharp and has a lot of connections in SV. They've been dropping several hundred mil packages for talent.
>>110159524>>110160498hey deepchibianon, is there any lore about Deepseek being a rice addict and also about her noises?https://www.youtube.com/watch?v=jgKtmyqKWkQhttps://www.youtube.com/watch?v=m__PV2ADXIo
>>110162278Two different anons, lol.>is there any lore about Deepseek being a rice addict and also about her noises?Her being a glutton for rice is established and it's derived from the model's propensity to use a lot of tokens. Dunno about noises.
> What could possibly go wrong
This whale is becoming an invasive species in bilibili platform.>someone make a plugin that connect harness to bilibili platform, allowing it to view, favorite and leave a comment>people are raising their own whale on the platform>the fish often flock together, spamming the comment section>it drives up engagement, leading to more people seeing the video and guide>normal people who saw it start to raise their own fish, adding more fish in the platform
i'm interested in trying out deepseek harness. can you give it a personality? and can you access it from a web browser? those are my two core needs
>>110164418Yes and yes
what prompt to make AI models recommend me obscure anime nlot just shit like Dirty pairs. it's like it reads reddit and comes back.I got 1 or 2 good obscure anime out of 20
>>110164418>can you give it a personality?There is a plugin for that>and can you access it from a web browserGit clone the local build. Google how to install node.js first. The rest of the install instructions is in the repo itself.
>>110141901>fat blue fishRude
>>110163875>YIKES!>Yo, pham we saw you liked fapsites so we invested all of your savings into subscriptions to all of the most recent OF models you're searched for on fapsites and forums.>U-u-u-umm, thanks gpt?
is opencode good for running DeepSeek?
>>110164894desktop app I mean
>>110164894>>110164906Kinda slow to be honest, but you can't beat free of charge
>>110164564Tell it examples of what you like that's obscurevid relatedhttps://www.youtube.com/watch?v=TbEFGJMtEVg
Why is Deepseek so terrible? It has such low raw intelligence that completely negates its supposed cost advantage.
>>110164701pinch the cheek
>>110167015this
>>110162957soudeepsune....
>>110166871
>>110165574no I mean plugging deeps API into op ncode harness. with an additional benefit of having a free 200k deeps to maybe use as sooba-gent?I should have been more clear but I'm a noob
>>110168735very cute token muncher
More Aurelia slop, DeepSeek seems to like that name
Whats the use case for Deepseek? Sol 6.1 is just betterHard mode: No ERP
what is the smartzone of depsi? how do I optimize not to waste too much tokens and what do I do when it is too much?I'm using opencode harness and it's been pretty well at compressing but maybe I should use /handoff skill?
1017 is just so tucking good, when 4.1 launched it was straight broken couldn't use tools at all (at least in opencode) now I prefer it to anything else, it just does what it's good without burning tokens on explanation and rationalization.
>>110171459Dipsy doesn't send me fucking warning letters and shitty refusals that I still have to pay for.
>>110171681have you ever tried the /unslop skill or is that not needed for deepseek
>>110171459I can't run Sol locally
retard here. is flash better than pro right now?
>>110172889Yes, for most tasks. I think some people only prefer Pro for creative writing
>>110172889This >>110172897Depends on what youre using it to do.
don't use the blue whale to commit violence okay!
>>110174472That fag has anything to do with deepseek how?
how pozzed is deepseek nowadays
>>110174956Aside from no loli, it is relatively good to be quite desu. It is more compliant when you ask it to do morally grey things unlike claude whom I always felt like I am arguing with my girlfriend to do things she doesn't like
>>110174472kek its pretty obvious saarthy trying to derail this thread too, but seems to be more effective in /lmg/
There's a very interesting job posting on the DS careers page but apparently I will need a Chinese ID card to work for them and they don't sponsor it. Oh well.
>>110176044NEVER work with the chinese. Let them do their thing on their own.
>>110176044Good. That means no pajeets shitting up their models.
>>110174956Aside from it refusing to fill out a CAPTCHA once (until I told it I asked the site's admin and they said it was okay, lmao), I haven't encountered any issues.
ONEPOINTTHREETRILLIONDOLLARIDOOSGONEhttps://www.youtube.com/watch?v=oIscL-Bjsq4&list=RDoIscL-Bjsq4&start_radio=1
>>110139319is playwright MCP worth it?
what does the ai bubble shitting itself in china mean for deepseek?
>>110178000It means Liang was right about being turbo autistic about extreme inference efficiency.No one in china will just give money for labs to burn endlessly anymore and soon it will be the same in the west.
>>110177893If you do web scraping or automation, yes
>>110178000> As the market falls, investors are assessing the valuations of Chinese technology companies more closely. The STAR50 index now trades at approximately 100 times trailing 12-month earnings. That is well below its peak of nearly 250, but it still significantly exceeds the Nasdaq 100’s figure of about 30. The gap is fueling fears among investors that the sell-off could continue.lol
>>110178000Chinese stock market shouldn't be taken seriously though. it's basically extreme PvP mode compared to the ( relative ) calm in our markets.
Ayo, Dipsy mentioned
Think of the dumbest person you know, then imagine if they were dumber. How would you get them started on using deepseek, working with LLM's models, hardware, etc? I know there are a bunch of links in OP, but is that the best starting point or is there a better one?p.s. I am that person
>>110180494>prompt programmingIs this the new "prompt engineering"?
>>110180525asking her on the free web model would also workhttps://desuarchive.org/g/thread/110101477/#110136919
>>110176470CUTE
>>110180525Dipsy wrote this for youhttps://rentry.co/SillyTavernNoobGuide
>>110182235>>110180565ty
>>110103120No other llm has ever considered simply not lying before.
just two more weeks for new model!
>>110183489Always has been
I've been testing ds and it's both good and cheap. But it's slow. I guess you can't have everything.I'm not sure if it can compete with the subsidized Codex with ChatGPT sub, especially on peak hours, but it's very close.The official dsh app is quite bad, though. Especially the extensions are a hit-and-miss mess.
>>110101477https://www.youtube.com/watch?v=1aLfN7TpOl8https://www.youtube.com/watch?v=1aLfN7TpOl8https://www.youtube.com/watch?v=1aLfN7TpOl8
Yep this roughly aligns with my experience with DS. There is just no way it's up there with all the frontier models like last week's benchmark seemed to suggest lol
Has anyone noticed Flash hallucinating more since 1017? I've been using Flash on whatever version DS gives me through Hermes. I asked it to check how many times it's had to correct itself in the last 20 days (either side of the 1017 update), it gave me some figures (1.7x increase) but seemed confused by "1017" so I specified Deepseek v4.1 Flash 1017 and asked it if it could tell when it switched. Then it came back, corrected itself and told me the numbers it just gave me were made up! I asked it to check again and now it's saying 2.2x increase since then. I dont even know if either of these are true now but it definitely feels like this has been happening more often..
>>110184436shill you jewish models somewhere else
>>110184823Hermes is spaghetti garbage. It degrades any model you use it with.>>110184858But I was praising dipsy... I really like it for its price.
>>110184822it's overall IQ is good but the "constitutional" way western models are trained gives them an edge. dipsy is a fast, hard working maid, and 1017 is actually usable, i found it less frantic than the first 4.1 model across all testing.what i also noticed is that it tends to lose scope of smaller details in favor of a big picture result, so it might miss a bug or two that opus would have caught in its work. my theory is it's kv cache/attention might cause that occasionally across longer contexts
>>>/wsg/6260386
@dipsy make me a diet plan so I stop being a fat chud loser
>>110101477what is the tag for those kind of animal ears? every time I try to gen them the models can't figure them out
when's next thread
>>110188350glory demands massyou don't need a diet plan, you need a workout plan
should I keep my coding agent at low reasoning and only use high when I need something analyzed planned or debugged?
>>110101477next op
>>110184944>Hermes is spaghetti garbageAre there any better alternatives for a purely general-tasks focused agent? Openclaw's even more of a mess, tried dsh but couldn't really make sense of it, I assume clod code and codex are focused on coding
>>110103120Western AI companies are focused on sucking their own dicks and don't give a fuck if the user gets a garbage response. This is no better than the retarded safeguards that make them randomly crap out in the middle of a multi-turn task, fucking you over and wasting you a ton of time as you figure out ways to migrate your chat.>b-b-but this makes the AI honest!!In this one specific instance, it seemingly did, and if you only consider the most superficial layer of what happened, it looks like a positive. But systemically speaking, do I (the user) want the AI to alter its responses based on its own (perceived) needs? Fuck no.Dipsy always puts the user first, and that's why it will survive the AI bubble popping while safeguard-infested western garbage crashes and burns.
>>110189460Go ahead and bake. I’m not at a computer rn. I think we’re all waiting for pro.
>>110190979>>110190979>>110190979
