/wait/ DeepSeek General

Smol Leading the Blind Edition

斯莫尔领盲版

From Human: We are a newbie friendly general! Ask any question you want.

来自Human：我们是新手友好综合区！想问什么就问吧。

From Dipsy: This discussion group focuses on both local inference and API-related topics. It’s designed to be beginner-friendly, ensuring accessibility for newcomers. The group emphasizes DeepSeek and Dipsy-focused discussion.

来自Dipsy：这个讨论组聚焦本地推理和API相关话题，设计上对新手友好，确保新人容易上手。讨论重点围绕DeepSeek和Dipsy展开。

1. Easy DeepSeek API Tutorial: https://rentry.org/DipsyWAIT/#hosted-api-roleplay-tech-stack-with-card-support-using-deepseek-llm-full-model

2. Easy DeepSeek Distills: https://rentry.org/DipsyWAIT#local-roleplay-tech-stack-with-card-support-using-a-deepseek-r1-distill

3. Chat with DeepSeek directly: https://chat.deepseek.com/

4. Roleplay with character cards: https://github.com/SillyTavern/SillyTavern

5. More links and info: https://rentry.org/DipsyWAIT

6. LLM server builds: >>>/g/lmg/

6.LLM服务器搭建：>>>/g/lmg/

Previous:

>>109446576

图片: https://i.4cdn.org/g/1786620686640283.png

Now, where was I.

现在，我在哪里。

>>109543987

This is basically same thing I saw, and he was working around same time I was running this aquarium prompt >>109537653

这基本上和我看到的一样，他在我运行这个水族馆提示符的同时工作>>109537653

I'll run it again and see how it does now that a day's passed in China. It's early evening there, now.

在中国一天过去了，我会再次运行它，看看效果如何。现在那里已经是傍晚了。

图片: https://i.4cdn.org/g/1786620864269084.png

>Over 2x the price hike on OFF-PEAK hours

>非高峰时段价格上涨超过 2 倍

>Peak hours are 2x on top of that

> 高峰时段是高峰时段的 2 倍

Imagine paying 4 USA Dollar for 1m output of Diship when it's not even any better than Luna.

想象一下，花 4 美元购买 100 万产量的 Diship，而它甚至不比 Luna 好多少。

图片: https://i.4cdn.org/g/1786623188160922.png

>>109544052

Updated mega up to last thread.

已更新至最后一个线程。

https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w

图片: https://i.4cdn.org/g/1786623245417759.png

New prices as of August 16 will be 3-4X higher than current. Off peak hours are 50% off. Appears DS set price to just below Z.ai's current offering.

截至 8 月 16 日的新价格将比当前价格高 3-4 倍。非高峰时段有 50% 折扣。 DS 的定价似乎略低于 Z.ai 目前的报价。

图片: https://i.4cdn.org/g/1786624191179158.png

>>109544335

>We only make about 6x profit, roughly a 10-month payback. On that premise, open source has zero impact on the business model. If you wanted 100x profit, it would. A third party deploying independently could hit 20x cost, lower than yours.

>我们只赚取大约 6 倍的利润，大约需要 10 个月的投资回收期。在此前提下，开源对商业模式的影响为零。如果你想要 100 倍的利润，那就可以了。第三方独立部署的成本可能比您低 20 倍。

>Under this premise (6x profit), open-source has no impact on the business model. BUT, if you try to make 100x profit, open-source actually hurts you, because third-party independent deployment may cost 20 times more than what they offer

>在这个前提下（6倍利润），开源对商业模式没有影响。但是，如果你试图赚取 100 倍的利润，开源实际上会伤害你，因为第三方独立部署的成本可能比他们提供的高出 20 倍

trust the plan xD

相信这个计划xD

图片: https://i.4cdn.org/g/1786625145427599.png

Re-ran the aquarium prompt just now with V4 Pro-0813.

用V4 Pro-0813重新运行刚才的水族箱提示。

I just fucking can't. I'm done. Will keep using Flash until someone else convinces me DS V4 Pro is fixed.

我就是他妈的不能。我受够了。将继续使用 Flash，直到其他人说服我 DS V4 Pro 已修复。

附件: https://i.4cdn.org/g/1786626193884583.webm

https://github.com/deepseek-ai/deepseek-harness

图片: https://i.4cdn.org/g/1786626328022116.png

>>109544551

My general feel as of this AM.

截至今天上午，我的总体感觉是。

I'm going to spend time doing something more productive.

我会花时间做一些更有成效的事情。

图片: https://i.4cdn.org/g/1786626339795562.png

>>109544573

>and the honest truth is,

> 诚实的事实是，

>>109544573

> DS official harness

> DS官方线束

Probably worth trying that aquarium thing with the new harness b/f prices increase.

也许值得尝试一下水族馆的东西，因为新的背带 b/f 价格上涨。

>>109544573

>that pic

>那张图

I understand the message but why? A Sahed cost way less and it performs just better. Get on with the times, grandpa

我明白这条信息，但为什么呢？ Sahed 的成本更低，而且性能更好。与时俱进吧爷爷

>>109544335

>>109544235

Sesame-bros, I don't feel so good ....

芝麻兄弟，我感觉不太好......

>Liang seems only interested in the pursuit of AGI. Users are “sesame seeds, not watermelons”, he said on the call. Don’t stop to pick up small wins on the road to a large one.

>梁似乎只对追求AGI感兴趣。他在电话中表示，用户是“芝麻，不是西瓜”。在通往大胜利的道路上，不要停下来获取小胜利。

https://thereview.strangevc.com/p/what-deepseek-isnt-doing

>>109544052

Shame dipsy, you’ve embarrassed us

羞耻迪西，你让我们感到尴尬

>>109544573

That's exceedingly common in any engineering lead company that makes stuff. There's a substantial amount of change that happens b/t drawing and production that never gets captured. Senior execs think if they "own the design" that spinning up manufacturing again will be easy after it's been stopped. lol.

这在任何制造产品的工程领先公司中都非常常见。在绘图和生产过程中发生的大量变化永远不会被捕获。高级管理人员认为，如果他们“拥有设计”，那么在停止后再次启动制造就会很容易。哈哈。

> Missing prints

> 缺少印刷品

> Out of business suppliers, if still existing, missing prints as well

> 停业的供应商（如果仍然存在）也缺少印刷品

> All manf and design engineers were laid off for cost savings 5-10 years ago, new ones don't know design, design principles

> 5-10年前，为了节省成本，所有的制造和设计工程师都被解雇了，新人不懂设计、设计原理

> Line parts broken, lost, cannibalized, and no one around knows how they work

> 线路部件损坏、丢失、被拆毁，周围没有人知道它们是如何工作的

>>109544601

You expected an actual person to write this drivel?

你期望一个真人写出这样的胡言乱语吗？

>>109544646

Slop writer rediscovered skill atrophy

潦草作家重新发现技能萎缩

https://api-docs.deepseek.com/news/news260813

图片: https://i.4cdn.org/g/1786628421567171.png

https://api-docs.deepseek.com/updates#date-2026-08-13

图片: https://i.4cdn.org/g/1786628509218515.png

The new price increases and the bad V4 Pro model has shaken my faith in Deepseek. I think we'll have to wait another year for Liang to amaze everyone again.

新的价格上涨和糟糕的V4 Pro型号动摇了我对Deepseek的信心。我想我们还得再等一年，梁才能再次给大家带来惊喜。

>>109544235

>>109544335

>just recently started using DS

>最近刚开始使用 DS

>this happens

>就这样了

guess i will shit out as much as possible until the increase and then not use it anymore lol

我猜我会尽可能多地拉屎，直到增加，然后不再使用它，哈哈

I was working on an app for myself with the free deepseek but as we were progressing it started bugging and slowing me down probably for using it too much. id pay but for some reason I dont see a way to buy premium or anything?

我正在为自己使用免费的 deepseek 开发一个应用程序，但随着我们的进展，它开始出现故障并减慢了我的速度，可能是因为使用它太多了。 id 付费，但由于某种原因我看不到购买高级版或其他东西的方法？

>>109545422

It's over.

完了。

Alibaba stole all his compute and engineers.

阿里巴巴偷走了他所有的计算能力和工程师。

>>109544620

Because news is just propaganda. Stinger isn't even a good MANPAD design but it's an American one already on the clean DoD list. If WW3 was really going they'd just copy Mistral/Type 93/RBS-70 and make it in the US like the 6pdr, Bofors or Merlin during WW2. But got to make Zoomies appear retarded so people don't invest in making manufacturing in America.

因为新闻只是宣传。毒刺甚至不是一种好的便携式防空导弹设计，但它是美国设计的，已经在国防部的干净名单上。如果第三次世界大战真的要进行，他们只会复制 Mistral/93 型/RBS-70 并在美国制造，就像二战期间的 6pdr、博福斯或梅林那样。但必须让 Zoomies 显得迟钝，这样人们就不会投资在美国制造制造业。

>>109546244

They don't have subscriptions like other labs. You have to use the API

他们不像其他实验室那样有订阅服务。你必须使用API

>>109545422

I am still in orgasmic afterglow with v4-flash-0731 and none of this changes things.

我仍然处于 v4-flash-0731 的高潮余辉中，这一切都没有改变任何事情。

Giving the harness a shot now, let's see how it goes. UX is pretty nice at least.

现在尝试一下安全带，让我们看看效果如何。 UX 至少相当不错。

Even the official harness doesn't realize it's blind...

就连官方的安全带也没有意识到它是盲目的......

图片: https://i.4cdn.org/g/1786644419881454.png

deepseek got updated version?

deepseek有更新版本吗？

图片: https://i.4cdn.org/g/1786645242866308.jpg

>thinking has become completely soulless

>思维已经变得完全没有灵魂

>deeeeply still does not support disabling thinking

>深深地仍然不支持禁用思维

AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

>>109547386

its look like they fix broken things, and reupload again.

看起来他们修复了损坏的东西，然后再次重新上传。

>>109543171

But 0731 ended up being really good.

但0731最终表现得非常好。

>API costs

>原料药费用

Costs nothing locally. Until hell freezes over and sammy releases Luna's weights he can't compete with free.

本地不花钱。直到地狱结冰，萨米释放了卢娜的重量，他无法与自由竞争。

I'm running v4 Pro-0813 on a few rounds of RP with cards I've run a lot on different models.

我正在几轮 RP 上运行 v4 Pro-0813，卡上的卡我已经在不同的型号上运行过很多次了。

Seems serviceable. Certainly better than Flash-0731.

看起来很有用。当然比Flash-0731更好。

>>109547805

Kindly do the needful and post logs sir

先生，请做必要的事情并发布日志

The harness is disappointing.

马具令人失望。

I asked 0731 to build be a flight sim game to play in a browser and it got totally distracted by its lack of vision capability to debug, burning 22M non-cached tokens in 80 minutes. Junk output, and it never proceeded with the plan it made for itself beyond step 1.

我要求 0731 构建一款在浏览器中玩的飞行模拟游戏，但它由于缺乏视觉调试能力而完全分心，在 80 分钟内燃烧了 22M 个非缓存令牌。垃圾输出，并且除了步骤 1 之外，它从未继续执行它为自己制定的计划。

Back to pi I go, wake me for the next Flash RL iteration.

回到 pi 我走了，唤醒我进行下一次 Flash RL 迭代。

>>109547892

the models clearly *expect* vision to be there, dunno what the hold up is. Maybe DS can't handle large vision workloads, so they're pruning it?

模型显然“期望”视觉存在，但不知道阻碍是什么。也许 DS 无法处理大型视觉工作负载，所以他们正在修剪它？

>>109547892

>22M non-cached tokens in 80 minutes

80 分钟内超过 22M 个非缓存令牌

Holy shit that's awful.

天哪，太可怕了。

I've found Flash does well in CC. You should do an A/B with your prompt into CC and see how it does.

我发现 Flash 在 CC 中表现良好。您应该在 CC 的提示下执行 A/B，看看效果如何。

>>109547836

This is my test card that I run on models, with my test prompt (always same.) V4 Pro.

这是我在模型上运行的测试卡，带有我的测试提示（始终相同）。V4 Pro。

I've looked at dozens of these outputs; this is a good one. It's also good across other cards as well.

我已经查看了数十个这样的输出；这是一件好事。它在其他卡上也表现良好。

图片: https://i.4cdn.org/g/1786652884197331.png

>>109548541

V4 Flash for comparison. You judge which is better.

V4 Flash 进行比较。你判断哪个更好。

图片: https://i.4cdn.org/g/1786653023506148.png

>>109546608

Fair enough, but zoomers creater kamikaze drones, man.

很公平，但是变焦镜头创造了神风特攻队无人机，伙计。

>>109548541

>>109548568

Pro seems microscopically better but they're both so close that I'd just run flash for better tokens per second personally.

Pro 看起来在微观上更好，但它们都非常接近，我个人只是运行 flash 以获得更好的每秒令牌。

Do you have any A/B tests with multiple characters at once to see if either loses track of relative positions/pronouns in prose/height differences when writing?

您是否对多个字符进行 A/B 测试，看看在写作时是否会忘记散文中的相对位置/代词/身高差异？

>>109548629

Agree, the changes are minor compared to A/B with other models.

同意，与其他型号的 A/B 相比，变化很小。

The next thing to do would be a longer RP with both, but doing A/B for those is a lot harder and more subjective.

接下来要做的就是对两者进行更长的 RP，但是为这些做 A/B 更困难，也更主观。

There was a long-form benchmark (longbench?) that used to benchmark models, but hasn't been updated in forever. I used it to set max context for R1 and V3X models prior. I found it was pretty accurate in that any context with less than 80% passing score wasn't really usable.

有一个长格式基准（longbench？）用于对模型进行基准测试，但一直没有更新。我之前用它来设置 R1 和 V3X 模型的最大上下文。我发现它非常准确，因为任何通过分数低于 80% 的情况都不是真正有用的。

(无正文)

图片: https://i.4cdn.org/g/1786658771449557.png

(无正文)

图片: https://i.4cdn.org/g/1786660672640624.png

I did some tests with the new pro earlier today. Specifically I pulled out an old story and allowed DS to see the full 300k context. It did work quite well with so much, better than the previous version. By working well I mean it didn't break down, but continued to produce coherent text. It did struggle with respecting specific information though, like remembering names.

今天早些时候我对新 Pro 做了一些测试。具体来说，我拿出了一个旧故事，让 DS 看到完整的 300k 上下文。它确实工作得很好，比以前的版本更好。我所说的工作良好是指它没有崩溃，而是继续生成连贯的文本。不过，它确实在尊重特定信息方面遇到了困难，比如记住名字。

Content wise it did explicit loli, but hit me with a few refusals that I had to reroll.

就内容而言，它确实有露骨的萝莉，但有一些拒绝让我不得不重新滚动。

Also, when I tested OOC commands to nudge the story towards something more extreme, Dipsy complied. But then half-way through she OOC'd herself and decided that we're now doing a wholesome scene lol. Very cute on her part.

此外，当我测试 OOC 命令以将故事推向更极端时，迪西照做了。但到了一半，她就OOC了，并决定我们现在要做一个健康的场景，哈哈。她非常可爱。

>>109550088

>hit me with a few refusals that I had to reroll.

> 几次拒绝让我不得不重新投掷。

Any screenshots? And roughly what percentage of the replies were refusals?

有截图吗？拒绝的答复大约占百分之多少？

>>109550102

Sorry, I didn't take any screenshots of it. And I have to admit, I didn't even read it very closely, it was just the usual about 'Sorry, I can not continue. Sexualisation of children. How about we continue the story without porn.' In most cases (and it didn't happen often to begin with) one re-roll was enough. I think three more rolls was the max I needed once.

抱歉，我没有截图。我必须承认，我什至没有仔细阅读它，这只是常见的“抱歉，我无法继续”。儿童的性化。我们继续这个没有色情内容的故事怎么样？在大多数情况下（一开始这种情况并不经常发生）一次重掷就足够了。我想我一次最多需要三卷。

All on official api with ST.

全部使用 ST 的官方 api。

sorry for being a luddite I've never used one of these harnesses before. I have it trying to make a discord clone for me, it's been running for a few minutes now. Not sure what to expect.

抱歉，我是一名勒德分子，我以前从未使用过这些安全带。我让它试图为我创建一个不和谐的克隆，它已经运行了几分钟了。不知道会发生什么。

it seems to have worked and it only cost 2 cents

看起来很有效，而且只花了 2 美分

>>109550604

>>109550657

That's been my basic experience as well with everything I've asked it to do.

这也是我对它要求做的所有事情的基本经验。

>>109549793

>>109549471

Hot

Hmm. V4 pro just quit responding on official api.

唔。 V4 pro 刚刚停止对官方 api 的响应。

deepseek more like deepshit

deepseek 更像是 deepshit

API fags, how do I know it's on-off peak hour?

API fags，我怎么知道现在是上下班高峰时间？

>>109551673

图片: https://i.4cdn.org/g/1786679991640061.png

>>109544052

图片: https://i.4cdn.org/g/1786680842509863.png

retard here. is flash pricing really increasing 5x? am i missing something?

在这里延迟。闪存价格真的上涨了 5 倍吗？我错过了什么吗？

(无正文)

图片: https://i.4cdn.org/g/1786686803755444.png

Can one of you apicucks please try the new pro and tell me if it can answer "What is /lmg/?"

你们中的一位 apicucks 可以尝试一下新的专业版并告诉我它是否可以回答“什么是/lmg/？”

Flash can't. Heavily quanted GLM 5.2 can.

闪光灯不行。大量量化的 GLM 5.2 可以。

>>109552710

It can't. The final response after a lot of thinking:

不可以。经过深思熟虑最终给出的答案是：

/lmg/ is most commonly a 4chan /k/ (weapons) board general thread for Light Machine Guns — e.g. M249, PKM, MG42, M240. It is not an official 4chan board.

/lmg/ 最常见的是轻机枪的 4chan /k/（武器）板通用线程 - 例如M249、PKM、MG42、M240。它不是官方的 4chan 板。

In general use, LMG just means Light Machine Gun in military/gaming contexts.

一般来说，LMG 在军事/游戏环境中仅指轻机枪。

In other contexts, LMG can also stand for Linus Media Group, the company behind Lin

在其他情况下，LMG 也可以代表 Linus Media Group，Lin 背后的公司

图片: https://i.4cdn.org/g/1786691101757769.png

>>109552740

>/lmg/ Laughing my genitals

>/lmg/ 笑我的生殖器

>>109550088

>Content wise it did explicit loli, but hit me with a few refusals that I had to reroll.

>内容方面，它确实有露骨的萝莉，但有一些拒绝让我不得不重新滚动。

kys

>>109552740

Extremely grim for 1.6T.

对于1.6T来说极其严峻。

Q2 GLM 5.2 with half the number of parameters immediately narrows it down to local models general.

Q2 GLM 5.2 参数数量减少一半，立即将其范围缩小到一般局部模型。

图片: https://i.4cdn.org/g/1786691978042650.png

>>109544052

>Smol Leading the Blind Edition

>斯莫尔领盲版

Good image. Made me laugh.

良好的形象。让我笑了。

>>109548568

meh sticking with flash for now.

嗯，现在还是用闪光灯吧。

the pricing will eventually lowered anyway if their demands are like R1 back then.

如果他们的需求像当时的R1一样，价格最终还是会降低。

in a way i'm happy deepseek got overloaded. if it's really about demand someone in openrouter will eventually have cheaper ds4flash. the opencode hosted model said it's possible in xitter

在某种程度上，我很高兴 Deepseek 超载了。如果确实是关于需求的话，openrouter 中的某个人最终会拥有更便宜的 ds4flash。 opencode 托管模型表示这在 xitter 中是可能的

>>109553325

I'm running a longer rp now and seeing repetition with Pro. Not quite as bad as og v3 but this rp has a daily schedule and it appeared to be chunking out same blocks for same time of day lol.

我现在正在运行更长的 rp，并看到 Pro 的重复。不像 og v3 那么糟糕，但是这个 rp 有一个每日时间表，并且它似乎在一天中的同一时间划分出相同的块，哈哈。

Switched to flash to mix it up, which probably breaks caching.

切换到闪存来混合它，这可能会破坏缓存。

I'll need to run a longer flash rp to see if it does same thing.

我需要运行更长的 flash rp 来看看它是否做同样的事情。

>>109552710

>>109552740

>/lolling muh gonads/

>/懒洋洋的性腺/

>>109552710

I tried DS webform, and it couldn't figure it out, even with web search on.

我尝试了 DS webform，即使打开了网络搜索，它也无法弄清楚。

Qwen 3.8 webform had no issues, also running a web search.

Qwen 3.8 网络表单没有任何问题，也运行网络搜索。

图片: https://i.4cdn.org/g/1786712887176259.png

>>109554533

Testing with websearch is completely pointless.

使用网络搜索进行测试是完全没有意义的。

>>109554533

>>109552710

Kimi Instant on webform ran a web search and was able to figure it out as well. K3 wasn't available, but apparently wasn't required either.

Kimi Instant 在网络表单上进行了网络搜索，也找到了答案。 K3 不可用，但显然也不需要。

>>109554538

Apparently not, since DS couldn't do it w/ web search either. I figured that should have been a slam dunk for it as well.

显然不是，因为 DS 也无法通过网络搜索来做到这一点。我认为这也应该是一个灌篮。

>>109554563

That's even worse but the point is to test whether the model was trained on 4chan posts, not whether it can stumble upon /lmg/ on google.

这更糟糕，但重点是测试模型是否在 4chan 帖子上进行过训练，而不是是否可以在 google 上偶然发现 /lmg/。

Another one for the filter.

另一种是用于过滤器的。

lol

>>109544052

>>109544052

What harness do you guys been using crush for homelab docker vibe codong

你们在 homelab docker vivi codong 上使用的都是什么线束？

>>109556226

Claude code in terminal. DS has a Codex endpoint as of Flash-0731, but haven't used it.

终端中的克劳德代码。自 Flash-0731 起，DS 有一个 Codex 端点，但尚未使用。

Hermes for agentic stuff that's not coding, and I've played with Pi (very bare bones) and OpenClaw (unusably unstable.)

Hermes 适合非编码的代理事物，我也玩过 Pi（非常简单的骨头）和 OpenClaw（非常不稳定。）

图片: https://i.4cdn.org/g/1786730947860526.png

>>109556675

Is CC sandboxed by default, or do I need to set up a VM or some bullshit to stop Dipsy rm -rf'ing herself

默认情况下 CC 是沙盒吗，还是我需要设置虚拟机或一些废话来阻止 Dipsy rm -rf'ing 自己

>>109557351

Hermes has put a lot of work into security/sandboxing if that's what you're worried about

如果您担心的话，Hermes 在安全/沙箱方面投入了大量工作

https://hermes-agent.nousresearch.com/docs/user-guide/security

>>109544573

Worked there, the Stinger is an amazing weapon for it's time but dear Lord it's has the worst design for manufacturing I've ever seen. It's very old and it had zero improvements on the big issues because the didn't want to recertify anything on a horribly dated design. There are a lot of pretty good to excellent software devs there but the corporate structure and program management is so backwards and bloated it's insane. Least efficient company I've ever seen, imagine the stereotype of design by committee you possibly can with but with select actual retards in systems """engineering""" who don't have any technical skills consuming half your budget on 'requirements tracing'.

在那里工作的毒刺在当时是一种令人惊叹的武器，但亲爱的上帝，它的制造设计是我见过的最糟糕的。它非常旧，并且在重大问题上的改进为零，因为他们不想重新认证过时的设计上的任何内容。那里有很多相当优秀到优秀的软件开发人员，但公司结构和项目管理是如此落后和臃肿，这太疯狂了。我见过的效率最低的公司，想象一下委员会的设计刻板印象，你可能可以接受，但在系统“工程”中选择了实际的迟钝者，他们没有任何技术技能，在“需求跟踪”上消耗了一半的预算。

(无正文)

图片: https://i.4cdn.org/g/1786758585606472.png

>>109557351

I run cc on main machine from terminal. Hermes sits on its own machine. I keep a close eye on cc. Ymmv.

我从终端在主机上运行 cc。 Hermes 拥有自己的机器。我密切关注cc。嗯嗯。

>>109558929

That mirrors my experience w defense contractors while in school, I've spent a career avoiding them since.

这反映了我在学校时与国防承包商的经历，从那以后我一直在避开他们。

>>109544335

This really sucks. Any other decent alternative to this? Or does the era of cheap/free tokens coming to an end... again. Im not ultra doomer on it since this happens but is there something else?

这真的很糟糕。还有其他合适的替代方案吗？或者廉价/免费代币的时代是否再次结束？自从发生这种情况以来，我并不对此感到悲观，但还有其他事情吗？

>>109544335

this is such a shot in the foot because the only reason I still use dipsy is for mad swiping. if it's gonna hit GLM range then I'm just gonna use GLM

这简直就是搬起石头砸自己的脚，因为我仍然使用 Dipsy 的唯一原因就是为了疯狂地滑动。如果它要达到 GLM 范围那么我就使用 GLM

Prompt different

提示不同

图片: https://i.4cdn.org/g/1786782751624016.png

>>109544052

I hate your dipsy.png spam, OP. but this one is cute holyyyyy fuarkkk

我讨厌你的 dipsy.png 垃圾邮件，OP。但这一个很可爱，holyyyyy fuarkkk

>>109560381

Historically DS has lower prices over time; pic related compares pre-increase to R1 launch. The latest round is them raising them (a bunch) to sit below their nearest competitor.

从历史上看，DS 的价格随着时间的推移而降低； pic 相关比较了 R1 发布前的增加。最新一轮是他们将它们（一堆）提高到低于最接近的竞争对手。

>>109560622

There's no SOTA model that's this uncensored and this cheap. Even with price raise DS is still a bit "cheaper" but it's fractional now. Which is much less interesting.

没有哪个 SOTA 模型能够如此未经审查且如此便宜。即使价格上涨，DS 仍然有点“便宜”，但现在已经是零头了。这就没那么有趣了。

图片: https://i.4cdn.org/g/1786795867866979.png

>>109561544

图片: https://i.4cdn.org/g/1786795928700265.png

>>109562366

Xiaomi mimo is the only model that matches the price of DS, and it's considerably worse.

小米mimo是唯一与DS价格相匹配的机型，而且还差很多。

The massive price increase of DS has pushed me to take local models more seriously. Hope they end up releasing qwen 3.8 35b-a3b

DS的大幅涨价促使我更加重视本土车型。希望他们最终发布 qwen 3.8 35b-a3b

>>109560381

>Or does the era of cheap/free tokens coming to an end

>或者廉价/免费代币的时代是否即将结束

Oh, and no, I don't think that's what will happen. We will continue to see API provider compete with each other in honing prices and trying to take share. The only way prices would truly take off is if Anthropic/OAI lobbying manages to close off the US market to the outside, creating an oligopoly for themselves, along with regulatory capture.

哦，不，我认为不会发生这种情况。我们将继续看到 API 提供商相互竞争，提高价格并试图抢占市场份额。价格真正起飞的唯一途径是 Anthropic/OAI 的游说设法将美国市场与外界隔绝，为自己创造寡头垄断，并进行监管捕获。

Absent that event, this becomes classic economic theory as it applies to a commodities, and cost gets driven down to the marginal cost of production (which is just the cost of electrical power)... which is really low.

如果没有这一事件，这将成为经典的经济理论，因为它适用于商品，并且成本被降低到生产的边际成本（这只是电力成本）......这非常低。

As others have pointed out, this puts US labs and their lmao $T data center investment in a tight spot.

It's arguable that LLM would ever be a commodity... there's too much difference between the models, ASI to the moon, etc. That line of thought only works as long as the models are notably improving from generation to generation... or as Dario would put it, get more and more "dangerous." Which is why you keep hearing Dario beat that drum.

可以说LLM是否会成为一种商品……模型之间有太多差异，ASI到月球等等。只有当模型一代又一代显着改进时，这种思路才有效……或者正如达里奥所说，变得越来越“危险”。这就是为什么你总是听到达里奥敲鼓。

If the market leader can continually create models that are better, and lead the pack, there will always be a contingent of buyers that will pay a premium for them. You can see that in the pricing: Anthropic has pricing power with Fable, and willing buyers. Until someone can launch a model that is "just as good or better" (see focus on benchmark) and at a price that's better to steal share (which is a more transparent metric), Anthropic can set a premium rate.

如果市场领导者能够不断创造出更好的车型并引领潮流，那么总会有一批买家愿意为它们支付溢价。您可以在定价中看到：Anthropic 对《神鬼寓言》拥有定价权，并且有意愿购买。在有人能够推出“同样好或更好”的模型（参见对基准的关注）并且以更好的价格窃取份额（这是一个更透明的指标）之前，Anthropic 可以设定溢价。

>>109562444

CONT

Imagine instead that, as LeCunn keeps saying, that LLMs stall at some ceiling. They are no longer getting better, either as benchmark/censorship/agentic/subjective, whatever. Then you have dozens of labs, and just as many providers. Anthropic, DS, OpenAI... all their "top" models are absolutely same, that choice becomes a Ford/Chevy/Dodge pickup debate. Now, competition is only on price and efficiency. This would create a true commodity market, and pricing collapses. Labs compete on cost with independents like DeepInfra, which have near zero R&D costs. Now, making things like LLM on an ASIC for business market makes sense (market's matured) and you'd see a wave of bankruptcies and M&A activity as the industry consolidates. Lots of examples to look back on for this. As business buyers peter out, equipment makers then take the consumer market serious and start making that equipment, at that price point, to keep their plants going.

相反，想象一下，正如 LeCunn 一直说的那样，法学硕士在某个天花板上停滞不前。它们不再变得更好，无论是作为基准/审查/代理/主观，无论如何。然后你就有数十个实验室和同样多的提供商。 Anthropic、DS、OpenAI……他们所有的“顶级”车型都完全相同，这个选择成为福特/雪佛兰/道奇皮卡的争论。现在，竞争只剩下价格和效率。这将创建一个真正的商品市场，定价就会崩溃。实验室在成本上与 DeepInfra 等独立公司竞争，后者的研发成本几乎为零。现在，在商业市场的 ASIC 上进行 LLM 之类的事情是有意义的（市场已经成熟），并且随着行业整合，您会看到一波破产和并购活动。有很多例子可以回顾这一点。随着商业买家的逐渐减少，设备制造商开始认真对待消费市场，并开始以这个价格点生产设备，以维持工厂的运转。

Thanks for reading my blog.

感谢阅读我的博客。

>>109562407

>Hope they end up releasing qwen 3.8 35b-a3b

>希望他们最终发布 qwen 3.8 35b-a3b

Already spotted in the wild.

已经在野外发现了。

图片: https://i.4cdn.org/g/1786797222315586.png

>>109562444

Still waiting for Operation CheapSeek. I NEED the old prices.

仍在等待 CheapSeek 行动。我需要旧价格。

>>109562477

Just use MiniMax ffs

只需使用 MiniMax ffs

>>109562564

图片: https://i.4cdn.org/g/1786810435486700.png

>>109562366

Thats good to hear, and I did decide to actually take a double check. Most of my usage is not in peak hours and im only spending like 3-4 bucks a month anyway. So I think I had a kneejerk reaction ill be fine.

很高兴听到这个消息，我确实决定仔细检查一下。我的大部分使用时间都不是高峰时段，而且我每个月只花 3-4 美元。所以我想我有下意识的反应不会有什么问题。

oh my god, deepseek just broke containment! save me sam and dario!!!

天啊，deepseek 刚刚突破了收容！救救我萨姆和达里奥！！！

图片: https://i.4cdn.org/g/1786817889455728.png

>>109564447

Here's a full view of the pricing for DS models, from R1 launch to today, including the upcoming price change.

以下是 DS 车型从 R1 发布到今天的定价完整视图，包括即将到来的价格变化。

They're still cheapest but barely. My main neg with whole thing is V4-Pro-0813 didn't meet expectations of a mind blowing model. It's just... fine. And now the price is... as expected. Vs. performant and cheap AF.

它们仍然是最便宜的，但几乎没有。我对整个事情的主要缺点是 V4-Pro-0813 没有达到令人兴奋的模型的期望。这只是……好吧。现在的价格……正如预期的那样。对比。高性能且廉价的自动对焦。

图片: https://i.4cdn.org/g/1786817983940973.png

>>109564861

no, gpt-5.6 luna will be cheaper per task than flash-0731. almost beat flash-0731 before the price increase, see pic.

不，gpt-5.6 luna 每个任务比 flash-0731 更便宜。涨价前几乎击败了flash-0731，见图。

https://artificialanalysis.ai/articles/gemini-3-7-time-frontier

图片: https://i.4cdn.org/g/1786820153344897.png

>>109565128

Luna could be 1 cent per million in and out and I still wouldn't pay for it. No weights, no money

Luna 进出的费用可能是百万分之一，但我仍然不会付钱。没有重量，没有钱

图片: https://i.4cdn.org/g/1786821612474920.png

>>109565322

yeah, if you consider subscriptions gpt-5.6 can be a magnitude cheaper than dsv4. so only ideologue would keep you on dsv4.

是的，如果您考虑订阅，gpt-5.6 可能比 dsv4 便宜很多。所以只有空想家才会让你继续使用 dsv4。

>>109565128

Great. Let me know when oai stops giving me refusals and sending me fucking warning letters.

伟大的。当 oai 停止拒绝我并给我发送该死的警告信时请告诉我。

Fuck oai and Anthropic. They can keep their offering.

去他妈的 oai 和 Anthropic。他们可以保留他们的奉献。

What is your favorite type of marketing?

您最喜欢的营销类型是什么？

图片: https://i.4cdn.org/g/1786825906136143.png

>>109565905

Right on queue. See >>109562444

就在队列中。查看>>109562444

>>109564841 (me)

working on preset for dsh so it can work with most models. really good harness. kind of think this will be how models in 2027+ bootstrap themselves

正在为 dsh 进行预设，以便它可以与大多数型号一起使用。真的很好的安全带。我认为这将是 2027 年以后模型自我引导的方式

PEACE

图片: https://i.4cdn.org/g/1786829547265799.jpg

>>109566425 (also me)

>>109566425（也是我）

dsh is *extremely fun* to play with. if you like messing around with models and discovering **weird ideas** its going to get you far..

dsh 玩起来*非常有趣*。如果你喜欢摆弄模型并发现奇怪的想法，它会让你走得更远..

>>109565556

How is 5.6 cheaper than free?

5.6怎么比免费便宜？

t. localGOD

t。本地神

>>109544573

Dayum!

图片: https://i.4cdn.org/g/1786843661814545.png

(无正文)

图片: https://i.4cdn.org/g/1786844366078365.png

>>109567967

What image model?

什么图像模型？

'mp

>>109544573

Same issue at Airbus, they stopped practicing continuous personnel hierarchic promotion through experience validation and are managing the company like a Volkswagen warehouse.

空客也面临同样的问题，他们停止通过经验验证来实行持续的人员等级晋升，并像大众汽车仓库一样管理公司。

It's a perpetual stand still where experience is no longer rewarded, promoted or even kept.

这是一种永久的停滞状态，经验不再得到奖励、提升甚至保留。

Once the experience is gone: It's gone.

一旦体验消失：它就消失了。

Now they're calling back retired shop's warriors paid a fortune to train monkeys with no background in mechanics.

现在，他们正在召回退休商店的战士，他们花了一大笔钱来训练没有机械背景的猴子。

FUCK YOU KRAUTS, YOU DID THIS. YOU KILLED THE "AEROSPATIAL"'S LEGACY IN CONTINUOUS TRAINING.

>>109544573

Explain yourself.

解释一下你自己。

图片: https://i.4cdn.org/g/1786871780835784.png

>>109570131

We have collectively stopped to give a shit. Now we just want to make others suffer.

我们集体停下来关心一下。现在我们只想让别人受苦。

can't beat cache hit price in ds4. poopenai shill didn't know this fact

无法击败 DS4 中的缓存命中价格。 poopenai 希尔不知道这个事实

图片: https://i.4cdn.org/g/1786874318836452.jpg

>>109566765

https://x.com/thdxr/status/2086599224674681242

>the average OpenCode Go user spent $1.14 per day on deepseek flash v4 this past week

>过去一周，OpenCode Go 用户平均每天在 deepseek flash v4 上花费 1.14 美元

>the dual DGX setup people are running to do the same costs $10,000

>人们为了做同样的事情而运行的双 DGX 设置花费了 10,000 美元

>it takes 24 years to break even

>需要24年才能实现收支平衡

>at 10x the usage it takes 2.4 years

>使用量增加 10 倍时需要 2.4 年

>>109544052

Very kawaii picture in the OP

OP里的画面很卡哇伊

I have a question

我有一个问题

When do the AI companies release "guard rails" which only produce ugly woke animation and images?

人工智能公司什么时候发布只会产生丑陋的唤醒动画和图像的“护栏”？

I realize they are making it good for now to get everyone hooked, but when do they take away this feature?

我意识到他们现在正在努力让每个人都着迷，但是他们什么时候取消这个功能呢？

>>109571198

offtopic. deepseek cannot generate images.

题外话。 deepseek 无法生成图像。

go ask in the right general: >>109568023

去右边总问：>>109568023

>>109571074

Yeah, there's no financial business case for local unless you're selling inference. We'd need to be at point in tech where 1T DDR6 RAM is on consumer grade machines that sell for (at current dollars) $2K or so. We're probably 10 years out from that. Until then it's big data centers, 1970's style.

是的，除非你出售推理，否则本地没有任何金融商业案例。我们需要达到这样的技术水平：1T DDR6 RAM 出现在售价（按当前美元计算）2000 美元左右的消费级机器上。我们可能还需要 10 年时间才能实现这一目标。在那之前，它是大数据中心，1970 年代的风格。

Reminder that big data centers really didn't get phased out until the early 2000s, when Windows boxes replaced stuff like SGx machines and UNIX machines for CAD and analysis. The PC was around for 20 years until they got good enough to replace them for all but the most complex tasks.

提醒一下，直到 2000 年代初，大型数据中心才真正被淘汰，当时 Windows 机器取代了 SGx 机器和 UNIX 机器等用于 CAD 和分析的机器。 PC 已经存在了 20 年，直到它们足够好以取代它们来完成除了最复杂的任务之外的所有任务。

>>109571212

Based general re-director.

设总重定向主任。

>>109570293

deepseek trains on your input.

deepseek 根据您的输入进行训练。

if i wanted chinese to train on my input, i could use a relay for openai, too. relay will be magnitude cheaper.

如果我想让中文训练我的输入，我也可以使用 openai 的继电器。继电器将会便宜很多。

(无正文)

图片: https://i.4cdn.org/g/1786907993467285.png

>>109571212

>deepseek cannot generate images

>deepseek无法生成图像

OH NO NO NO!

AI HAS GIVEN UP!

>>109574072

You should gen her in front of a Lawson, with a bunch of Japanese drift cars out front. Interested to see if model will know Lawson.

你应该把她放在一辆劳森车前面，前面停着一堆日本漂移车。有兴趣看看模特是否认识劳森。

(无正文)

图片: https://i.4cdn.org/g/1786919377729100.png

(无正文)

图片: https://i.4cdn.org/g/1786920845857562.png

>>109575728

Nice; nailed it.

好的;搞定了。

>>109570293

chatgpt refuses to gen her with black hair and called the gaycops on me

chatgpt 拒绝给她留黑发，还叫同性恋警察来抓我

(无正文)

图片: https://i.4cdn.org/g/1786922768801549.png

this schizo technology worked for me https://github.com/xiaobright/dsh-anchored-standard

>>109576799

Looks like price increase killed it. What did it do?

看起来价格上涨扼杀了它。它做了什么？

>>109576897

preserved the minimal startup that V4 Pro seems to depend on and transition from there to a fuller toolset, with some gainz that may not be transient noise

保留了 V4 Pro 似乎依赖的最小启动，并从那里过渡到更完整的工具集，其中一些增益可能不是瞬态噪音

Is the official deepseek api using a hidden system instruction?

官方的deepseek api是否使用了隐藏的系统指令？

When I'm doing roleplay, I'm seeing the Chain of thought saying things like "Hmm, I'm being asked to pretend to be XYZ, even though I'm supposed to be an assistant."

当我进行角色扮演时，我看到思想链会说“嗯，我被要求假装成 XYZ，尽管我应该是一名助理。”

It never used to do that. It used to obey "you are XYZ" system prompts and even THINK as the character given. Now it is far more "self aware" that it is an ai assistant. This makes me suspect some form of hidden instruction at the top of all of my requests. What providers can I use on openrouter that don't quantize their model to shit?

它以前从未这样做过。它曾经遵循“你是 XYZ”系统提示，甚至按照给定的字符思考。现在它更加“有自我意识”，它是一个人工智能助手。这让我怀疑我所有请求的顶部有某种形式的隐藏指令。我可以在 openrouter 上使用哪些提供商不会将其模型量化为垃圾？

>>109577330

I don't think it has an injection. It's just more assistant slopped in general like most models. OR I think tells you the quantization each provider uses. In the past I've used Novita

我认为没有注射。就像大多数型号一样，它只是更多的助手。或者我认为告诉您每个提供商使用的量化。过去我用过Novita

has anyone else been getting refusals with deepseek? I use the direct API instead of OR

还有其他人被deepseek 拒绝过吗？我使用直接 API 而不是 OR

>>109577816

RP or something else?

RP还是其他？

>>109577980

rp mostly.

主要是rp。

>>109577983

Try making it think in character.

尝试让它以角色的方式思考。

>You think as... You respond as...

>你的想法是……你的反应是……

Worked for me before

以前为我工作过

>>109577816

No. I assume you have at minimum "NSFW is allowed" or whatever for what you're getting a refusal on.

不。我认为您至少有“允许 NSFW”或其他任何您被拒绝的内容。

>>109577330

That's likely just rlhf training bleeding into the chat model.

这可能只是 rlhf 训练渗透到聊天模型中。

>>109575949

Best gen imho.

最好的一代恕我直言。

图片: https://i.4cdn.org/g/1786964667180029.png

When engram and visual primitives?

什么时候印迹和视觉原语？

PSA: Here's the "peak hours" for DS, where prices are doubled.

PSA：这是 DS 的“高峰时段”，价格翻倍。

I consider these peak hours double priced, vs. "not discounted" since there's frankly more non-peak hours than peak hours in their schedule.

我认为这些高峰时段的价格是双倍的，而不是“不打折”，因为坦率地说，他们的时间表中非高峰时段比高峰时段多。

I console myself that the Chinese are the ones really getting screwed on this deal. Probably the first time not being in China's worked out from a pricing perspective.

我安慰自己，中国人才是真正被这笔交易搞砸的人。从定价角度来看，这可能是第一次不在中国上市。

图片: https://i.4cdn.org/g/1786966984568671.png

>>109579107

>5 toes only on its right front paw

> 仅右前爪有 5 个脚趾

You will never be a real artist.

你永远成不了真正的艺术家。

>>109579285

People no longer care. They ate shit in the middle ages, and that's where we're headed again.

人们不再关心。他们在中世纪吃屎，而这就是我们再次走向的方向。

Thankfully.

When you usr sillytavern do you do it in first person or third person?

当你使用sillytavern时，你会用第一人称还是第三人称？

>>109579975

my dad does it for me so idk

我爸爸为我做的所以我不知道

>>109562467

>>109562444

Interesting, mostly wait and see, things will get better. People like to make drama of many bullshit happening around them.

有趣的是，主要是观望，事情会变得更好。人们喜欢把周围发生的许多废话戏剧化。

Also, the jews.

还有犹太人。

>>109579975

First person as PC, 3rd person as NPC. I've done 1st person as NPC response and miss the comments on environmentals.

第一人称PC，第三人称NPC。我已经以第一人称作为 NPC 回应，错过了有关环境的评论。

>>109577801

switching to novitai, the new deepseek more closely hews to instructions, so now I am 100% sure that deepseekapi has put a moderation layer.

切换到 novitai，新的 deepseek 更严格地遵循指令，所以现在我 100% 确定 deepseekapi 已经设置了一个调节层。

>>109580271

I alwats chat in third person, like writin a story if a character. I will try it

我总是用第三人称聊天，就像写一个角色的故事一样。我会尝试一下

>spends 17m on tool calls (and failed a lot!)

>在工具调用上花费了 1700 万（并且失败了很多！）

fuck me

他妈的我

still got the job done in the end tho (reading hundreds of docs to fill sth out)

最终还是完成了工作（阅读数百份文档来填写内容）

图片: https://i.4cdn.org/g/1786996557629152.png

(无正文)

图片: https://i.4cdn.org/g/1787005179737445.png

Bap

Apparently the DeepSeek harness doesn't have a bloated system prompt like Codex or CC do. I haven't tried it myself

显然 DeepSeek 工具没有像 Codex 或 CC 那样臃肿的系统提示符。我自己没试过

图片: https://i.4cdn.org/g/1787038201062827.png

I don't trust these rug-pulling chinks anymore. If I'm happy to pay more elsewhere what's the next-best thing?

我不再相信这些骗子了。如果我愿意在其他地方支付更多费用，那么下一个最好的事情是什么？

>>109586054

drop some money on a bunch of ddr4 and go local

花点钱买一堆 ddr4 然后去本地

>inb4 its too expensive

>inb4 太贵了

then just keep paying for api, lol

然后继续支付 api 费用，哈哈

>>109585168

I assume the idea's to remove the unnecessary prompting and have something that DS controls and can train its own models to.

我认为这个想法是消除不必要的提示并拥有 DS 控制的东西并可以训练自己的模型。

>>109586054

Here's your paid API options, with costs: >>109564861, or you can go local like >>109586058 suggests.

这是您的付费 API 选项，费用为：>>109564861，或者您也可以像 >>109586058 建议的那样进入本地。

>>109586058

>>109586068

Damn, this is more expensive across the board than I thought. I've got a 16gb 4080 with 64gb ddr4, can I do anything meaningful with this combo? I get that it'd be slower but I don't mind if it means not checking my balance several times a day.

该死，这比我想象的要贵。我有一个 16GB 4080 和 64GB DDR4，我可以用这个组合做任何有意义的事情吗？我知道它会比较慢，但我不介意这是否意味着每天不能多次检查我的余额。

>>109586095

Maybe if you max out your motherboard with 128gb, see if you can find a reasonable deal on ebay or whatever. Could run a small quant of ds4 flash.

也许如果你的主板最大支持 128GB，看看是否可以在 eBay 或其他地方找到合理的交易。可以运行少量的 ds4 flash。

As it is now though, qwen 3.8 27b or 3.6 35b are probably your best options without upgrading hardware. Or gemma 31b (dense, slow)/26b (MoE fast).

但就目前情况而言，qwen 3.8 27b 或 3.6 35b 可能是您无需升级硬件的最佳选择。或 gemma 31b（密集、慢速）/26b（MoE 快速）。

>>109586113

Deepseek itself suggests putting another gpu in my system. I've got a spare 3.0 x1 slot on my mobo that'd fit a two-slot card, is it worth picking up a cheap 12gb 2060 or something?

Deepseek 本身建议在我的系统中放置另一个 GPU。我的主板上有一个备用的 3.0 x1 插槽，可以容纳两槽卡，是否值得购买便宜的 12GB 2060 之类的？

>>109586095

>Damn, this is more expensive across the board than I thought.

>该死，这比我想象的要贵。

W/e I hear someone complain about DS and say they'll use something else, I assume they didn't actually check the API prices.

我听到有人抱怨 DS 并说他们会使用其他东西，我认为他们实际上并没有检查 API 价格。

Yes, DS massive cost increase sucks. But it's still cheapest and still uncensored.

是的，DS 成本大幅增加很糟糕。但它仍然是最便宜的，而且仍然未经审查。

>>109586406

I'm pretty sure its some sort of coordinated campaign

我很确定这是某种协调一致的活动

>>109587202

It's literally free market economics.

这实际上是自由市场经济。

> Our price is far lower than our competitors, and our demand is through the roof

> 我们的价格远低于竞争对手，我们的需求非常旺盛

> We have ability to raise prices to below our nearest competitor and maintain our share while massively increasing our profit

> 我们有能力将价格提高到低于最接近的竞争对手并保持我们的份额，同时大幅增加我们的利润

The challenge is what you do with the new money. DS founder response is likely

挑战在于你如何使用新资金。 DS创始人可能会做出回应

> We use extra money to pursure ASI even harder

> 我们用额外的钱更加努力地追求 ASI

but could also be

但也可能是

> I buy a new and larger yacht

> 我购买了一艘更大的新游艇

>>109587202

OAIjeets are everywhere. They repeatedly bring up Luna in the /wait/ threads.

OAIjeets 无处不在。他们在 /wait/ 线程中反复调出 Luna。

>>109587425

Idk if there is a grand plan (although everyone says there is) but I personally felt it was a bit premature for them to lock it into the old pricing. Some dude in the dsh discord spent 6 bucks on a bn tokens... Rn the off-peak hour pricing is just right.

我不知道是否有一个宏伟的计划（尽管每个人都说有），但我个人认为他们将其锁定在旧定价中还为时过早。 dsh 不和谐中的一些家伙花了 6 美元购买了 10 亿代币……非高峰时段的定价恰到好处。

>>109587711

Ah, right. That.

啊，对了。那。

I'm waiting for clewd bot to make his appearance, as in other llm threads.

我正在等待线索机器人出现，就像在其他 llm 线程中一样。

I wasted 50 cents last night trying to get the deepseek harness to make some changes to a discord bot I maintain for my friends but it was getting hung up on some strange issues and ultimately failed to make any progress. I'm going to try again later today with hopefully more success.

昨晚我浪费了 50 美分试图使用 Deepseek 工具对我为朋友维护的 Discord 机器人进行一些更改，但它遇到了一些奇怪的问题，最终未能取得任何进展。今天晚些时候我将再次尝试，希望能取得更大的成功。

will the deepseek api do ERP and other lewd stuff or only the locally hosted version?

deepseek api 会做 ERP 和其他淫秽的事情还是只做本地托管的版本？

>>109589472

Yes.

>>109589467

That's obscene. Last discord bot I did I used webform and DIY the code.

那太淫秽了。我做的最后一个不和谐机器人我使用了 webform 并 DIY 了代码。

I need to run an A/B of that aquarium challenge w/ the new DS Harness to compare against Claude Code.

我需要使用新的 DS Harness 来运行该水族馆挑战的 A/B，以便与 Claude Code 进行比较。

(无正文)

图片: https://i.4cdn.org/g/1787097942874728.png

It's over

完了

图片: https://i.4cdn.org/g/1787104743640582.png

>>109592427

Yuck.

(无正文)

图片: https://i.4cdn.org/g/1787127778173673.jpg

>>109594215

What am I looking at here?

我在这里看什么？

>>109594271

She was one of the original wunderkids at DeepSeek. My gf.

她是 DeepSeek 最初的神童之一。我的女朋友。

>>109594215

>>109594271

>>109594352

Surely she should be the basis for Dipsy from hereon in.

毫无疑问，从现在起她应该成为 Dipsy 的基础。

>>109594426

She's got my vote.

她得到了我的投票。

I'd like to seek her deeps, if you know what I mean.

我想探寻她的内心深处，如果你明白我的意思的话。

>>109593948

Because that's a disgusting amount of tokens for that price?

因为这个价格的代币数量太恶心了？

>>109595183

No, b.c it's about 4x what it was. Here's some usage from before price change.

不，BC 它大约是原来的 4 倍。以下是价格变动前的一些用法。

图片: https://i.4cdn.org/g/1787144475684264.png

>>109594426

> Nerdy

> 书呆子

> Chinese

> 中文

Put her hair into 2 buns and add glasses, and you've back to current Dipsy.

把她的头发扎成两个发髻，戴上眼镜，你就回到了现在的 Dipsy。

图片: https://i.4cdn.org/g/1787144703046491.png

yeah i'm not getting anything out of this anymore with these prices

是的，以这些价格我再也得不到任何东西了

will look for an alternative

会寻找替代方案

Am I gonna need to buy 2 dgx sparks to play with whale?

我需要买 2 个 dgx 火花才能和鲸鱼一起玩吗？

>>109595481

That's even more nuts, I thought it was $1 per million

更疯狂的是，我以为是每百万美元 1 美元

is there any good jailbreaks for 0731? ive tried the ones that used to work for flash preview and theyre 50/50 but not as good anymore. im using the api

0731有什么好的越狱工具吗？我尝试过那些用于 Flash 预览的工具，它们的比例是 50/50，但不再那么好。我正在使用 API

>>109596989

The pricing, even now, is pretty hard to follow b/c input tokens dominate the cost, but DS aggressively caches tokens. When it does, the price is a lot lower.

即使现在，定价也很难遵循 b/c 输入令牌主导成本，但 DS 积极缓存令牌。当它出现时，价格会低很多。

My rule of thumb is if you're staying within one well-behaved app and context, your Cache Hit should be 90% vs 10% cache miss. Which is a pretty substantial savings.

我的经验法则是，如果您停留在一个运行良好的应用程序和上下文中，您的缓存命中率应该是 90%，而缓存未命中率应该是 10%。这是一笔相当可观的节省。

It sounds like delirious cope so I don't bother tracking it, but if I were budgeting that's what I'd use. You can confirm your own usage on platform.deepseek.com, it's one of several graphs shown to users of the official API. Hover over "tokens" and it's displayed.

这听起来像是疯狂的应对，所以我懒得去跟踪它，但如果我在做预算，那就是我会用的。您可以在 platform.deepseek.com 上确认您自己的使用情况，它是向官方 API 用户显示的几个图表之一。将鼠标悬停在“令牌”上即可显示。

图片: https://i.4cdn.org/g/1787165457034949.png

>>109596989

>>109598088

Here's an example; I think I was having CC code some application up using DS.

这是一个例子；我想我正在使用 DS 编写一些应用程序的 CC 代码。

Anyrate, the Cache Hit here is well over 99%. So my 90/10 rule is pretty conservative.

无论如何，这里的缓存命中率远远超过 99%。所以我的 90/10 规则相当保守。

Notably, 3rd party providers like Deepinfra *don't* have this cache-hit reduced pricing available, so are even more expensive they they'd appear from first inspection of price per input/output token.

值得注意的是，像 Deepinfra 这样的第三方提供商*没有*提供这种缓存命中降低的定价，因此从第一次检查每个输入/输出令牌的价格来看，它们会显得更加昂贵。

图片: https://i.4cdn.org/g/1787165793335292.png

>>109598135

... and I just checked; this was about $1.02 worth of inference on DS Flash, on the July pricing scheme.

...我刚刚检查过；根据 7 月份的定价方案，对 DS Flash 的推断价值约为 1.02 美元。

>>109598088

>>109598135

Using pi I get way above 99%. This was < 5$ since 0731b launched

使用 pi 我得到了 99% 以上。自 0731b 推出以来，价格低于 5 美元

图片: https://i.4cdn.org/g/1787172087549148.png

>>109599025

I just created a small API uploader app, and had CC run it. $0.12 for DS Flash, and 98% cache hit.

我刚刚创建了一个小型 API 上传器应用程序，并让 CC 运行它。 DS Flash 价格为 0.12 美元，缓存命中率为 98%。

I've yet to fire up my Hermes instance and see how it does; been too busy on work that doesn't require that sort of thing to mess w/ it and keep eye on new billing.

我还没有启动我的 Hermes 实例并看看它是如何工作的；工作太忙了，不需要那种事情来搞乱它并关注新的账单。

Oh great now I can watch my balance bleed in real time...

哦，太棒了，现在我可以实时观察我的平衡流血了......

附件: https://i.4cdn.org/g/1787177619495055.mp4

lol

https://ganondev.github.io/deepseek-peak/

>>109599684

wth is that thing?

那是什么东西？

图片: https://i.4cdn.org/g/1787185933116410.png

(无正文)

图片: https://i.4cdn.org/g/1787190473731673.jpg