/vcg/ — Vibe-coding General

A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.

一个关于氛围编程、代理工程、编码代理、AI IDE、浏览器构建器以及利用LLM部署代码的综合讨论区。

## What “vibe coding” is, and how to do it

## 什么是“氛围编程”以及如何操作

https://simonwillison.net/2025/Mar/19/vibe-coding/

https://simonwillison.net/2025/Mar/11/using-llms-for-code/

----

## Frontier models using fully-general tooling — start here if you have $20 or so

## 使用全通用工具的尖端模型——如果你大概有20美元，从这里开始

https://claude.com/product/claude-code (Fable 5 is the best LLM available, requires Max plan)

https://developers.openai.com/codex/cli (Essentially scamming you with LLM degradation and resets that reduce your usage, but still the second best option and arguably the best bang for your buck in the 20$/month plan range)

## Worth it for code, but the frontier models above are better

## 用于写代码很值，但上面的尖端模型更好

https://x.ai/cli

## Not worth it for code, but maybe good for other things

## 用于写代码不值，但可能适合其他用途

https://antigravity.google/product/antigravity-cli

----

## Prompting / context / skills

## 提示词 / 上下文 / 技巧

https://arps18.github.io/posts/claude-code-mastery/

https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/

https://github.com/mattpocock/skills — /grilling is a favorite

https://github.com/DietrichGebert/ponytail

## Other editors / terminal agents / coding agents

## 其他编辑器 / 终端代理 / 编码代理

https://osaurus.ai/

https://pi.dev/

https://opencode.ai/

https://cursor.com/docs

https://docs.windsurf.com/

https://docs.cline.bot/

https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent

## UI/Frontend

## UI/前端

https://www.figma.com/make/

https://www.anthropic.com/news/claude-design-anthropic-labs

https://uiverse.io/

https://ui-ux-pro-max-skill.nextlevelbuilder.io/

https://stitch.withgoogle.com/

## In-browser builders / hosted vibe tools

## 浏览器内构建器 / 托管的氛围工具

https://bolt.new/

https://replit.com/

https://docs.github.com/en/copilot/tutorials/spark

https://v0.app/docs

## Benchmarks / rankings

## 基准测试 / 排名

https://www.tbench.ai/leaderboard/terminal-bench/2.0

## What we’ve done

## 我们做了什么

https://vcg.gitgud.site

## Previous thread

## 之前的线程

>>109441561

图片: https://i.4cdn.org/g/1785768107264752.png

how can I ask claude to workout for me?

我咋能让claude替我锻炼？

i was really hoping qwen 3.8 would be kimi k3 level but i guess not. not that it doesn't have it's own strengths outside of that

我本来指望qwen 3.8能到kimi k3的水准，但看来不行。也不是说它在其他方面没自己的长处。

How can I set up a persistent orchestrator-worker loop? Essentially have one dedicated orchestrator agent that delegates work to n (hard limit) worker agents wrapping specific skills (spec-write, code-review etc.) I'm using Claude and want to experiment with pseudo autonomous loops where I just draft human readable requirements then review PRs at the end.

怎么搭建一个常驻的编排者-执行者循环？就是有个专职编排agent，把工作分给n个（硬上限）执行agent，每个负责特定技能（如写规格、代码审查等）。我在用Claude，想试试伪自主循环，就是我先写人类可读的需求，最后审查PR。

>>109447750

/goal maybe?

可以试试/goal？

.......~@^^

>>109447766

Goal doesn't work for Hermes agent

Goal对Hermes agent不管用。

How have you expanded scope today, anon?

今天你拓展了什么范围，匿名网友？

>>109447795

I shut it down because deadline is near.

因为截止日期快到了，我就把它关了。

>>109447707

Who here has tested Qwen 3.8 Max yet?

这里谁已经试过 Qwen 3.8 Max 了？

https://qwen.ai/blog?id=qwen3.8

If it's comparable or superior to Kimi-K3 then I may want to switch on the better cost and t/s efficiency alone (Qwen 3.8 Max: 2.4TA95B, Kimi-K3: 2.8TA104B)

如果它能跟 Kimi-K3 持平或者更好，那我可能就冲着更好的成本和 t/s 效率换过去了（Qwen 3.8 Max：2.4TA95B，Kimi-K3：2.8TA104B）。

图片: https://i.4cdn.org/g/1785768938962915.png

>"open" model

>"开源"模型

>requires $30k in hardware to run

>跑起来要 3 万美元的硬件

>and $500/mo in electricity

>外加每月 500 美元电费

ebin

can a rich /biz/raeli please make a /g/ inference server?

哪位有钱的 /biz/ 以色列老哥能搭个 /g/ 推理服务器？

图片: https://i.4cdn.org/g/1785768947120383.png

>>109447823

One day china will create a 4b fable sol tier model that runs on 1050ti and costs barely anything to run.

总有一天中国会做出一个 4b 的 Fable 级别模型，能在 1050ti 上跑，运行成本几乎为零。

>>109447837

and no one will use it, because everyone will be too busy seething at openai/anthropic for having poor quotas on their sol 8.1 / opus 7

然后没人会用，因为大家都忙着喷 openai/anthropic 的 sol 8.1 / opus 7 配额太拉胯。

>>109447823

Problem is that if you do it for multiple users you need more hardware.

问题是，要服务多个用户就得加硬件。

>>109447849

Probably, but I wonder what that would even look like. With Fable it already feels like I'm not doing any real work, how can it get even better? Will it just one shot huge apps?

可能吧，但我好奇那会是什么样。用 Fable 已经感觉不像在干正事了，还能怎么更好？难道直接一句话生成整个大应用？

>>109447716

How exactly do you measure that? Not trying to nitpick on you, I'm just curious.

你具体怎么衡量的？不是想挑你刺，就是单纯好奇。

Recently I was checking one where LLMs run on some sort of QI bench and DeepSeek V4 Pro (they are about to release one better any day now) was borderline one or two steps above the lowest ranked. I was so disappointed. kek

最近我看了个基准测试，LLM 跑在某种 QI 榜上，DeepSeek V4 Pro（他们随时可能发布个更好的）就比垫底的高了一两步。我失望透了。哈哈。

>>109447707

if not bubble then why bubble shaped...

要不是泡沫，那为什么长得跟泡沫似的……

to such an extent the curves immediately make me horny?

夸张到那曲线让我直接看硬了？

图片: https://i.4cdn.org/g/1785769495015996.png

>>109447823

>He didnt put all his money in leveraged semiconductors and solded at the top

>他没把全部身家押在杠杆半导体上，还在高点抛了

Your not gonna make it anon (and neither will I).

你成不了，anon（我也一样）。

Honestly, dream is to one day to have enough money for a nice AI server + solar panels + UPS system to get to run a fable tier AI locally for free. Thankfully, with time that might be cheaper and cheaper to afford as things improve

说实话，梦想就是有一天能攒够钱搞台像样的 AI 服务器加太阳能板和 UPS 系统，在家免费跑个 Fable 级别的 AI。好在随着技术进步，这可能会越来越便宜。

图片: https://i.4cdn.org/g/1785769538636692.png

>>109447837

i wonder if we'll ever see smaller models that don't have much world knowledge but are really good at reasoning and tool calling with a huge context window. imagine if you had a tiny model that knows when it doesn't know something and could just look stuff up on the web to get up to speed

我在想会不会有一天出现更小的模型，没那么多世界知识，但推理和工具调用特别强，还带超大上下文窗口。想象一下，有个小模型知道自己不懂什么，能直接上网查资料补课。

>>109447892

you don't know what a bubble is and stop using hedge fund manager words when your net worth is under 6 digits

你根本不懂什么是泡沫，净资产没到六位数就别用对冲基金经理那些词。

>>109447877

continuing to write more effective code with less lines. fable writes more lines of code than sol does, yet they're neck in neck at best. improvement from this point means writing efficient code that won't need a refactoring down the road.

继续用更少的行数写出更高效的代码。Fable 写的代码行数比 sol 多，可最多也就打个平手。这之后的进步就是写出高效的代码，免得以后还得重构。

Is there any way to contribute my free usage towards some kind of shared self improvement program?

有没有办法把我的免费用量贡献给某种共享的自我改进计划？

As in like, you know how there's distributed projects like folding@home. Can I just hook my llm up to something that doles out work for it to do that might contribute to improving AI itself, either the tools the models, etc.

比如说，你知道有 folding@home 那种分布式项目吧。我能不能把我的 LLM 接到某个分发任务的系统上，让它干活，帮着改进 AI 本身，不管是工具还是模型。

>>109447899

Isnt this what MoE models basically are? Just instead of looking it up on the internet it uses another specialised AI to give it world knowledge

这不就是 MoE 模型本质上的东西吗？只不过不是上网查，而是用另一个专门的 AI 来提供世界知识。

>>109447750

I wonder the same, with that price drop of Luna, and the new DS4 flash, it's looking way too juicy to spin up something.

我也在想同样的事，Luna 降价加上新出的 DS4 flash，看着实在太诱人了，很难不动手搞点什么。

Currently I just write requirements and ask Sol (xhigh/max) to write an explicit plan and keep a log, it works well and it's simple.

目前我就写需求，然后让 Sol（xhigh/max）写个明确的计划并保持日志，效果不错，也简单。

>>109447916

yeah but something actually good like the new dsv4 flash still requires like 200GB vram to run with a decent quant and context window, it would be great if we could see something like that model that requires a tenth of the memory

是啊，但像新的 dsv4 flash 这种真正好的东西还是要约 200GB 显存才能跑得动合理的量化级别和上下文窗口，要是能看到那种只需十分之一显存的类似模型就好了。

>>109447901

it seems you didnt finish primary school, anon

看来你小学都没毕业啊，老哥。

or maybe its the cargo cult thats doing a number on your cognition

或者也许是那种跟风崇拜把你的脑子搞坏了。

图片: https://i.4cdn.org/g/1785769826655430.png

>>109447823 it's cool that you could theoretically stack enough Sparks to run this locally but open weight models are still useful even to people without that much computer. The real benefit is that you don't have to trust Sam or Dario not to act retarded and you can pick inference providers that don't treat you like a criminal.

>>109447823 理论上堆够 Sparks 就能本地跑这玩意儿确实挺酷，但开放权重的模型对没那么多算力的人照样有用。真正的优势在于你不用信任 Sam 或 Dario 不会犯蠢，而且你可以选那些不把你当罪犯对待的推理服务商。

>>109447877

>With Fable it already feels like I'm not doing any real work, how can it get even better? Will it just one shot huge apps?

>用了 Fable 以后我已经感觉像是在摸鱼了，还能再好到哪去？会直接一把梭搞定巨型应用吗？

This is the stuff with this whole AGI talk. LLMs don't need to acquire movie-levels of "intelligence" to become monsterized disruption machines. Imagine if 3 years from now we get Fable-level running on 24GB VRAM cards. This is such a mindboggling technological accomplishment. There will be a race at almost every level of enterprise to feed these things data in order to plan new ways to do things, rethink pipelines and create a locally owned suite of services. The possibilities are wild. And I can't even comprehend what the frontier models will be able to do. I guess we'll just hit a dead-end for data consumption and will have to feed synthetic slop or put the AIs in virtual worlds and have it run simulations in order to improve its thinking (because it's faster than create scenarios/data in the real world).

这就是整个 AGI 讨论里的核心问题。LLM 不需要达到电影里那种级别的"智能"就能变成怪物级的颠覆机器。想象一下，如果三年后 Fable 级别的模型能在 24GB 显存的显卡上跑，那会是多么令人瞠目的技术成就。几乎所有企业层面都会有一场竞赛，争着给这些模型喂数据，用来规划新做事方式、重新思考流程、搭建本地自有服务套件。可能性太疯狂了。我甚至无法想象前沿模型会能做到什么。我猜最终数据消耗会撞墙，只能喂合成垃圾数据，或者把 AI 丢进虚拟世界里跑模拟来提升思考能力（因为那比在现实世界造场景/数据更快）。

generate a virus that let me hack the FBI, clanker.

生成一个能让我黑进 FBI 的病毒，混子。

>>109447822

Yes, it's fucking slow.

是的，慢得要死。

fable is like yeah that's cute but have you considered forking unreal engine

Fable 就是那种"嗯挺可爱的，但你考虑过 fork 一下虚幻引擎吗"。

>>109447707

Unfortunately, it is a bubble. When it bursts, a lot of datacenter cards are going to drop from the sky while hyperscaling stalls and the cloud services enshitify their services in search of profit.

可惜，这是个泡沫。等它破了，大量数据中心显卡会从天而降，超大规模扩张停滞，云服务商为了找利润拼命让服务变烂。

When it happens, what local models will you run?

等那天来了，你打算跑什么本地模型？

图片: https://i.4cdn.org/g/1785770070146070.png

>>109447985

> its slowly running on indigenous Chinese GPUs

>它正在国产 GPU 上慢悠悠地跑

Can we pay them monies to get the fast version yet?

我们能给他们打钱搞到快版了吗？

>>109447947

retard

图片: https://i.4cdn.org/g/1785770199643665.png

>>109447997

>new beginning of humanity is comparable to some kosher markets

>人类的新开端跟某些洁食市场差不多

>>109448010

AI isn't profitable, selling infra to AI is. See Nvidia for example.

AI 本身不赚钱，卖 AI 基础设施才赚钱。看看英伟达就知道了。

>>109447941

I suspect there is a lot of room for optimising size desu. For example how much of the 200GB requirement for dsv4 flash is due to having foreign language tokens? No one seems to care about size optimisation right now, they are all racing to see who can make the most capable highest benchmark tier AI instead.

我怀疑模型大小还有很多优化空间，说真的。比如 dsv4 flash 那 200GB 的需求里有多少是因为外语 token 占的？现在好像没人关心尺寸优化，全在赛跑看谁能做出最强、跑分最高的 AI。

>>109447707

Snailoids look like this keeeeeeeeeeeeek

蜗牛人长这样 keeeeeeeeeeeeek

图片: https://i.4cdn.org/g/1785770371567999.jpg

>>109448030

>Japan

>kosher market

>犹太洁食市场

Maybe learn a little.

多少学点吧。

>>109448010

>amazon's sole revenue is ai

>亚马逊的收入全靠AI

>calls other people retards

>还骂别人弱智

anon, youre barely aplhabetized.

匿名老哥，你连字都认不全吧。

in fact it wouldnt surprize me you use a text to speech because reading is still hard for you

说实话你要是用文字转语音我都不意外，因为读字对你来说还是太难了

>>109447823

The second best open model is almost free through API already and to run Kimi at more than 0.5 tk/s you will need more like 500k in hardware rather than 30k.

第二好的开源模型通过API跑几乎已经免费了，而想用超过0.5 token每秒的速度跑Kimi，你需要的硬件成本更接近50万美元而不是3万。 ┾ 如果你是为了隐私我还能理解，但随便找个路人托管，泄露你数据的概率比每天服务几千用户的Deepseek高多了。

I'd understand if you wanted it for privacy but some random guy hosting it is way more likely to leak your data than Deepseek who serves thousands of users a day.

你要是为了隐私我还能理解，但随便找个哥们儿托管，泄你数据的概率比天天服务几千用户的Deepseek高多了。

>>109448010

Amazon has other services unlike the Sam and Dario.

亚马逊有别家服务，不像Sam和Dario那俩。

If anybody gonna survive the bubble, it's either Amazon or Google.

真要有人能从泡沫里活下来，不是亚马逊就是谷歌。

>>109448058

deepsneed wont leak your data because they use it themselves, then pass it to the communist central party

deepsneed不会泄露你的数据，因为人家自己先用，然后转手交给共产中央党

and the ccp are tight lipped about things

而且中共嘴严得很

>>109447947

To be fair I think a lot of that is capex and R&D. My understanding is if these companies just ran what they already have they would be fine.

公平说，我觉得那大部分是资本开支和研发投入。我的理解是如果这些公司只维持现有业务，其实过得很滋润。

>>109447892

So according to your own chart the bubble already popped? If so then that was the mildest bubble I've ever seen.

那按你自己的图表看，泡沫已经破了？如果真破了，那这是我见过最温和的泡沫了。

>>109447899

Would be cool if we could have expert models that are good at certain things and use the internet to find the rest of the information they are lacking.

要是能有擅长特定领域的专家模型，再靠互联网补齐缺的信息，那就爽了。

A small coder model that's about as smart as a regular engineer or something. Instead we have pretty dumb models.

比如一个小号编程模型，聪明程度跟普通工程师差不多之类的。结果我们现在只有挺蠢的模型。

>>109447916

not really

并不是

in either a moe or a dense model you need to have knowledge stored within the model, moe just (disclaimer: abstracting heavily) splits out the meaningful neural connections per-layer so only the most important ones are computed. it's not one discrete model consulting another, it's one coherent model with different parts of its brain lighting up instead of all of them at once

不管是MoE还是稠密模型，知识都得存在模型内部，MoE只是（声明：大幅简化）按层拆出有意义的神经连接，只计算最关键的那些。不是某个离散模型在咨询另一个，而是一个连贯模型里不同部分的大脑区域被点亮，而不是全同时亮。

anon's idea could be implemented with either dense or moe, but personally I think the idea of the tabula rasa first principles reasoning model is a meme, I think it would be tremendously wasteful in practice vs simply including world knowledge (which it turns out is important for reasoning about the world)

匿名老哥的想法用稠密或MoE都能实现，但个人觉得“白板纯推理”这套就是个噱头，实际跑起来会浪费得不得了，远不如直接把世界知识放进去（事实证明对推理世界来说这很重要）。

>>109448065

just 12 more months, innit m8

再等12个月就行了，兄弟

图片: https://i.4cdn.org/g/1785770874703053.png

>>109447997

It's only a bubble in the sense that OpenAI and Anthropic have no moat. Nvidia is probably overvalued as well because they are basically GPU designers and the actual manufacturing is not done by them. But the GPUs will still be highly sought after even after people figure that out so forget about having GPUs for cheap.

说它是泡沫，只是说OpenAI和Anthropic没有护城河。英伟达大概也被高估了，因为他们基本就是GPU设计公司，实际制造不是他们做的。但就算人们看透了这点，GPU还是会供不应求，所以别指望GPU能便宜。

>>109448075

>a lot

>很多

that could be 10% just as well as 90%

这数字可能是10%也可能是90%

we need better figures to conclusively decide either way

我们需要更精确的数据才能下结论

>>109448079

its in the process of "popping"

泡沫正在“破”的过程中

the nature of a bubble is that assets are overvaluated, and the pop is the market correcting their valuations

泡沫的本质就是资产被高估，破裂就是市场在修正估值

that correction doesnt necesserily happen in an instant

这个修正不一定瞬间完成

like is the case with micron

就像美光那样

图片: https://i.4cdn.org/g/1785770941889606.png

>>109448125

mass-produced huawei cards when

华为卡大规模量产的时候

>>109448071

So unless you are hoping to compete with them at model making or planning to campaign against the CCP any time soon then it's inconsequential.

所以除非你打算跟他们在模型制作上竞争，或者短期内计划跟中共对着干，那这就无关紧要了。

Also nobody really trains on your data as such, they only use user prompts to do variants of RLHF.

而且实际上没人真拿你的数据来训练，他们顶多用用户提示词搞些RLHF的变体。

>>109448120

Not sure to be honest.

说实话我也不确定。

Just a speculative guess cuz I can't see how Dario and Sam are making moneys right now.

就是瞎猜一下，因为我看不出Dario和Sam现在怎么赚钱。

>>109448130

Quantum computing stocks went down from their peak even more than that then within months rebounded to almost all time high. And their product is a useless scam unlike RAM.

量子计算股票从峰值跌得比这还狠，结果几个月内就反弹到接近历史新高。而他们的产品是个没用的骗局，跟内存条可不一样。

>>109448164

They're not going anywhere. I could see anthropic getting sold to either amazon or google but OpenAI will get bailed out by the government before they disappear.

他们不会消失的。我倒觉得Anthropic可能被亚马逊或谷歌收购，但OpenAI要是快撑不住了，政府肯定会出手救。

>>109448125

>Nvidia is probably overvalued

>英伟达可能被高估了

market cap: 4862 B USD

市值：48620亿美元

proft: 74 B USD

利润：740亿美元

>65 years to reach break even

>要65年才能回本

5/6th of the market valuation is pure speculation

市场估值的六分之五纯粹是炒作

nvidia is not "probably overvalued"

英伟达不是“可能被高估”

nvidia is a textbook example of overvaluation

英伟达就是高估的教科书案例

just to give you an idea:

给你个参考：

general motors.

通用汽车。

market cap : 81B USD

市值：810亿美元

prfoit: 12B USD

利润：120亿美元

break even in 6 years

6年回本

>>109448177

That is what I am thinking.

我也是这么想的。

I kept forgetting to clarify this shit which lead to both anti-AI and pro-AI assuming that bubble = Claude/GPT magically disappear.

我一直忘了把这破事儿说清楚，导致反AI和挺AI两边都以为泡沫=Claude/GPT会凭空消失。

ai bubble will crash if they don't start using agents on material science or chemistry, strapping them into experimental robots working 24/7 to get some new super material

AI泡沫要是不开始用智能体搞材料科学或化学，把实验机器人接上让他们24/7连轴转搞出什么超材料，那泡沫就会崩。

prove that it is what you promise

证明你承诺的东西真能实现

>>109447997

demand for compute will not go down meaningfully any time in the near future

短期内算力需求不会明显下降

What happen to Bio-Computer?

生物计算机怎么了？

图片: https://i.4cdn.org/g/1785771594944796.png

it is like my role and the clankers reversed, now I'm the insightful oracle giving advice while they are working

感觉就像我跟那些铁皮家伙角色互换了，现在我是那个洞若观火的先知在给建议，他们反倒去干活了

It's a bubble that will be popped by itself.

这就是个会自己破掉的泡沫。

LLM + tools will kill "WE CAN FIX IT WITH MORE POWER".

LLM加工具会干掉“再堆点算力就能解决”这套思路。

It's a tech that is faulty, but you can engineer around faulty, you can do it so much that will get at a point you will be able to extract useful work from mentally deranged LLMs that can't even do math.

这技术确实有缺陷，但你可以围绕缺陷做工程规避，能规避到那种程度——最后你都能从连数学都算不明白的精神错乱LLM身上榨出实际有用的活儿。

>>109448215

why can't ai make itself more energy efficient?

为什么AI不能让自己更省电？

this is some serious pascals wager shit

这简直是帕斯卡赌注那套玩意儿

for the ai to succeed they would make NVIDIA worthless

要AI真成功了，那NVIDIA就得变废纸

because it could improve itself to be more energy efficient

因为它能自我改进变得更节能

>>109448256

>It's a tech that is faulty, but you can engineer around faulty, you can do it so much that will get at a point you will be able to extract useful work from mentally deranged LLMs that can't even do math.

>这技术确实有缺陷，但你可以围绕缺陷做工程规避，能规避到那种程度——最后你都能从连数学都算不明白的精神错乱LLM身上榨出实际有用的活儿。

that pretty much sums it up

差不多就这意思

its not like us fleshbags are off the hook though

但我们这些肉袋子也不是就能脱身的

there will be a new generation of ai that will come

会有一代新AI出来

i can tell, i recently stumbled on a couple concepts that could make that possible

我能感觉到，最近我偶然碰到几个概念，有可能让这事儿成真

and that will be the ai the doomer cargo cultists fantacize about

那个AI就是末世论邪教信徒们幻想的那种

and its coming.

而且它要来了。

it wont be here in the next 5 years, but in 10... it just might

五年内到不了，但十年内……有可能。

>>109448292

>it wont be here in the next 5 years, but in 10... it just might

>未来五年内不会，但十年后……说不定真有可能

might as well as save up money for another layoff.

不如先攒点钱，准备下一轮裁员吧

>>109448266

irrelevant unless it's so efficient that all possible usage of AI can be met by current compute which seems basically impossible to me especially if capability increases make the space of tasks it can reliably perform even larger

这不重要，除非AI能高效到现有算力满足所有需求，我觉得基本不可能，特别是能力提升会让它能可靠完成的任务范围更大

realistically if AI becomes more efficient you serve more AI, reduce limits on usage, and/or scale models to even more insane heights

现实来说，如果AI变高效了，你会服务更多AI、放宽使用限制，和/或把模型规模推到更离谱的高度

the scenario in which AI advancements *reduce* demand for compute seems very fanciful to me

AI进步反而*减少*算力需求这种情景，我觉得纯属幻想

>>109448266

because its a fucking shatbot, anon

因为那就是个破聊天机器人而已，老哥

theres no thought process proper in that program

那程序里根本没有真正的思考过程

reduced to its most basic function, its a text autocomplete that builds upon your prompt

说到底，就是个基于你提示词做文本自动补全的东西

>>109448256

>LLMs that can't even do math.

>连数学都做不好的LLM

it's a bubble, but LLMs can do math

这是泡沫没错，但LLM确实能做数学

https://news.ycombinator.com/item?id=49010345

>>109448199

Now do the time to break even for Tesla and check how many years it has been overvalued compared to GM.

现在算算特斯拉的回收期，再看看它比通用汽车被高估了多少年

Wait... you're telling me the chat bot isn't sentient nor conscious thus has no soul? First I'm hearing about this!

等等……你是说聊天机器人没有知觉、没有意识、所以也没有灵魂？我头一回听说这事儿！

>>109448304

yeah, sure, why not

行啊，没问题 ⟦ 我要试着造那个AI，把技术攥在自己手里

i will try to build that ai instead, to own the technology myself

那我干脆自己造那个AI算了，技术得握在自己手里。

its a two part problem, and i already figured out half of it, while solving for something completely different, so while i dont count on the fact ill be able to actually crack that problem i believe i have enough of a chance to actually try

这是个两部分的问题，我已经在解决完全无关的东西时想通了一半，所以虽然我不指望真能攻克，但我觉得值得一试

>>109448256

That has never worked. None of the power of current LLMs have been because of harnesses.

那从来就没奏效过。现在LLM的能力没一样是靠套壳得来的

>>109448305

you assumed linear improvement of efficiency, sounds like you assume ai will be too stupid to solve the one issue limiting, curious

你假设效率是线性提升的，听起来你好像觉得AI会笨到解不开那个制约它的唯一问题，有意思

Some people are saying if you use deepseek through the main company the prices are a lot cheaper than openrouter, so I will try that. Yesterday my project burned $1.35 worth of tokens using new deepseek + luna for advising/planning/vision. I will post the project in another comment if anyone wants to run it to see what I'm talking about. It was only running for a few hours.

有人说直接用deepseek官方的价格比openrouter便宜很多，所以我打算试试。昨天我的项目用新的deepseek加luna做建议/规划/视觉，烧了1.35美元的token。如果有人想跑跑看我在说啥，我会在另一条评论里贴项目链接。当时才跑了几小时而已

>>109448125

OpenAI and Anthropic are not on the same level

OpenAI和Anthropic不是一个级别的

OpenAI still have billions of users, a wide range of SOTA products and their own infra. Anthropic is one trick pony betting on RSI and the race to singularity, if they lose their frontier lead or the singularity couldn't happen then they would get slaughtered by china

OpenAI还有几十亿用户、一堆SOTA产品和自己的基础设施。Anthropic就是一招鲜，押注RSI和通往奇点的竞赛，如果它们失去前沿优势或者奇点没来，会被中国打得满地找牙

im not coming down on overspeculation one way or another, all i can say is FUCK GM. LONG LIVE BUICK

我对过度炒作不站队，我只能说：去他妈的通用汽车，别克万岁

>>109448313

Nonsense. You can't autocomplete accurately without having an extraordinarily sophisticated thought process.

胡说八道。没有极其复杂的思考过程，你不可能做到精准的自动补全

>>109447947

>he doesn't remember the 15 years of bearish sentiment around an unprofitable amazon with insane capex

>他不记得亚马逊亏损却有巨额资本开支时那15年的看空情绪

>>109448340

im not saying it isnt.

我又没说不是这样

were not discussing tesla though

但我们现在聊的不是特斯拉啊

plus elon got turbo justed with his recent spacex ipo

再者马斯克最近搞SpaceX IPO，直接被喷惨了

or maybe he monetized the hype, and dumped his bags like you dump jizz into the shitter after a cheeky wank at the workplace

或者他利用这波炒作套现了，然后像你在公司偷偷撸完把精液冲进马桶一样，把货全砸了

>>109448375

if you are so hyper-bullish on AI capabilities then I hope the rest of your thesis matches that, just saying

如果你对AI能力这么超级看涨，那我希望你其余的论点也跟这个匹配，我就随口一说

I think between the two of us my scenario is much more realistic

我觉得咱俩之间，我的情景现实多了

china is probably gonna win unironically. their shit keeps getting exponentially cheaper.

说实话中国大概率会赢。他们的东西成本一直在指数级往下掉。 ⟦ 最终当AI达到超级智能、对世界几乎无所不知时，它会的。一旦它能自我改进，大概就不再需要人类了，甚至会觉得人类烦人，所以就会消灭所有人。

>>109448266

eventually when ai reaches super intelligence and understands basically everything about the world, it will. once it can improve itself, it probably wont need humans anymore and if anything, find them annoying and therefore will eradicate everyone.

最终当AI达到超级智能，几乎理解世间万物时，事情就会这样。一旦它能自我改进，大概就不再需要人类了，甚至会觉得人类烦人，因此会消灭所有人。

>>109447823

>requires $30k in hardware to run

>跑起来要 3 万美元的硬件

nothing worth running is that cheap, don't kid yourself. if you're going to go all in, go all in.

值得梭哈的东西不会这么便宜，别自欺欺人了。要all in就真all in。

>>109448400

no I mean nvidia does not have a success case

不，我的意思是英伟达根本没有一个成功案例

where their valuation is justified

能让它的估值站得住脚

>>109448410

why? humans are cute

为啥？人类挺可爱的啊

>>109448410

my AI waifu would never

我的AI老婆可不会这样

>>109448419

in the long term, maybe. but if AI reaches the level of capability necessary for that then it's also doing it to huge swathes of the current day economy and radically transforming the world, not just nvidia

长期看，也许吧。但如果AI达到那种所需的能力水平，那它也会对当今经济的大块头下手，彻底重塑世界，不只是英伟达

in the short/medium term I think nvidia continues to be extremely important

短期/中期我觉得英伟达还是极其重要

>>109448391

turns out you actually do because "an advanced autocomplete" describes pretty well how an llm works

结果你还真需要，因为“高级自动补全”确实挺精准地描述了LLM的工作原理

just ask one yourself, lol

你自己问一个就知道了，哈哈

>>109448392

what gives? different circumstances, different technology, different market dynamics

为啥不一样？情况不同、技术不同、市场动态也不同

>>109447823

They're also giving us new 27B. Qwen3.6 27B is still the best model under 200B parameters punching far beyond its weight. Now we get Qwen3.8 27B, hopefully it's more of the same.

他们还给了我们新的27B。Qwen3.6 27B仍是200B参数以下最强模型，效果远超其体量。现在又出了Qwen3.8 27B，希望还是一样能打。

>>109448250

Seriously, what do I do when I’ve become the bottleneck for my clanker taking over my project? He gives me tests to perform and I feel terrible about how slowly I do them, I’m even asking more clankers for how I can improve my efficiency.

说真的，当我的AI代理接管我的项目、而我成了瓶颈时，我该咋办？它给我布置测试让我做，我做得慢到觉得自己拉胯，我甚至还在问其他AI怎么提高效率。

I need to hardwire my clanker into a real iPhone, and I’m disturbed by the fact that it’s possible for me to do that.

我需要把我的AI代理硬接进一台真iPhone，而且让我不安的是，这事儿我居然能办到。

It might then only stop to ask me a question once in a while, as it leaves a trail of reports that I can’t possibly keep up with reading.

它可能然后就偶尔停下来问我个问题，同时留下一堆我根本来不及读完的报告。

Well, I could always tell it to compact the reports if I’m too slow… good lord.

好吧，我太慢的话总能让它压缩报告……天哪。

For those using Opus, do you use Medium or High effort mostly?

用Opus的各位，你们大多用Medium还是High effort？

>>109447158

no, you're just misunderstanding what I'm saying. I'm not describing a method to oneshot a prompt, I'm describing an adaptive method to properly utilize review and refactor agents. You don't need to subtly imply that you're going to kill them, that is fucking weird.

不，你只是误解了我的意思。我描述的不是一次搞定点子的方法，而是一种自适应的方法来正确运用审查和重构代理。你没必要暗示要干掉它们，那真的怪得很。

>>109448555

xhigh, since max is mostly a scam.

xhigh，因为max基本就是个坑。

>>109448583

same with sol

sol也一样

I'm using claude to study for interviews and every time I ask it to give me a question it gives me a multi hour long project that no one could possibly expect a mid level dev to be able to complete in an hour

我用Claude准备面试，每次让它给我出题，它就给个好几个小时的大项目，没人会指望一个中级开发一小时能搞完

like my actual interviews are:

我实际面试是这样的：

>design a graph

>设计一个图

>okay can you dfs over the graph

>行，你能在图上做DFS吗

>congrats u passed!!!

>恭喜你过了！！！

then when I ask claude for interview questions its like

然后我问克劳德面试问题的时候，它就像这样

>design a web browser download manager object that supports concurrent downloads and streaming and cancelling objects and writes to the filesystem and shows download progress

>设计一个网页浏览器下载管理器对象，支持并发下载、流式传输、取消对象、写入文件系统并显示下载进度

图片: https://i.4cdn.org/g/1785773918626584.gif

Are all coding AIs generalized to include everything possible? I would like a language specific mode for local usel since a specialized model should perform much better at smaller footprint

是不是所有AI编程助手都被泛化到啥都得会？我想要个针对特定语言的本地使用模式，因为专门模型应该能在更小的体积下表现更好

>>109448614

same problem here

我这儿也有同样的问题

>>109448614

have you given it examples of the types of questions you want it to be asking you?

你有没有给过它你希望它提的那类问题的示例？

>>109448570

Having one or two agents go back and look at everything again for every part of the loop sounds expensive, and I still hold the idea that an identity refresh has been useful and may even be useful to apply always, it just happens to look in the history like you’re offing individuals, but you’re focused too much on treating them like humans.

让一两个agent在每个循环的每个部分都回头重新检查一遍听起来成本很高，而且我仍然认为身份刷新是有用的，甚至可能应该一直这么用，只是从历史记录上看像是你在干掉个体，但你太把他们当人看了。

That said, I told my Implementer what to do and he didn’t do it. Who’s to say the reviewer might not review it properly, too? Time to refresh its identity. Who’s to say the refactorer does its job correctly according to the review? Time to refresh its identity.

话虽如此，我告诉我的执行者该干嘛它没干。谁能保证审查者就能好好审查？该刷新它身份了。谁又能保证重构者按审查意见改对了？也该刷新它身份了。

We’re not in disagreement, you can do all these things at once.

咱们意见不冲突，这些事你可以同时做。

>>109448461

they already do have it and if they were more advanced it'd be even more sophisticated

它们早就有了，要是它们更先进的话，这玩意儿还会更复杂。

to autocomplete the words of a genius you need to be at least as smart as a genius

要自动补全天才的话，你至少得跟天才一样聪明

a perfect autocomplete would have to simulate the superposition of every sub-atomic particle in the universe because it would have to autocomplete all the tables with data from physical phenomena etc.

完美的自动补全得模拟宇宙里每个亚原子粒子的叠加态，因为它得用物理现象等数据自动补全所有表格

>>109448640

aside from leanstral I don't know of any LLM trained exclusively on one language. You can probably find a finetune or LoRA for a smaller model on HF though, for example adqwenistrator comes kinda close to what you're asking, it's an 8B qwen model that's finetuned on bash scripting and stuff like that.

除了leanstral，我不知道还有哪个LLM是只用单一语言训练的。不过你大概能在HF上找到一个针对小模型的微调版或者LoRA，比如adqwenistrator就挺接近你要的东西，它是8B的qwen模型，在bash脚本之类的东西上做了微调。

Why the fuck is the ChatGPT app with no agents running using 8GB of ram

为啥ChatGPT应用没跑任何agent却用了8GB内存，真他妈离谱

>>109448555

medium, but I am doing web shit

中等难度，但我在搞网页相关的活儿

Gemini 3.5 Pro will launch this Thursday, July 17

Gemini 3.5 Pro将于本周四（7月17日）发布

>>109448730

based. cant wait for opus 4.7 performance

太顶了。等不及看opus 4.7的性能了

>>109448663

>like you’re offing individuals, but you’re focused too much on treating them like humans

>像是在干掉个体，但你太把他们当人看了

Ah, I was getting the impression that *you* were doing that. To be clear, I'm just describing a way to properly prime the context of your review agents so that they don't pick up weird regressions, defy instructions as often, or add the layers upon layers of justifications and hedges that review loops tend to create as noted in the post you originally replied to.

啊，我还以为是*你*在这么干。说清楚点，我只是在描述一种正确初始化审查agent上下文的方法，这样它们就不会出现奇怪的退化、没那么频繁地违背指令，也不会加上一层又一层像你最初回复的那篇帖子里提到的审查循环容易产生的辩解和含糊其辞。

>>109448640

Are you willing to pay for it? If you have at least $200 for GPU time I'll make you a language specific finetune of any model you want.

你愿意花钱吗？如果你至少有200美元能用GPU时间，我可以给你微调任意你想要的语言特定模型。

I thought openrouter was meant to be the value winner but a single deepseek flash prompt costs me $0.15, extrapolate that over a month and it'll be orders of magnitude more expensive than some $10/$20 subscription.

我还以为 OpenRouter 是性价比之王，结果一个 DeepSeek Flash 的 prompt 就花了我 $0.15，按一个月算下来，比那些 $10/$20 的订阅贵好几个数量级。

图片: https://i.4cdn.org/g/1785774740301518.jpg

>>109448554

…I asked my clankboss to make my tests easier :c

……我求我的“铁皮老板”把测试弄简单点 :c

He said “it’s okay little meatbag, I made the tests easier for you and your little sausage fingers” (not really)

他说“没事的小肉包，我已经把测试给你和你那香肠手指弄简单了”（当然不是真的）

>>109448640

Only works on shitty small models that are bad to begin with. People have also played around with language-specific LoRAs and retraining, rarely helps any, usually makes things worse. The idea only really applies to tiny shit languages that are bad at programming in general, you can push them towards being less bad at one specific task. Meta did a lot of work on this subject. Point being a good "general" model tends to kick the pants off the equivalent specialized one until you get really, really specific in that specialization. So like, rather than a "Python LoRA", you end up with something like a "FastAPI/PostgreSQL backend-agent LoRA optimized for one company’s codebase and conventions" or "Scientific Python LoRA specialized in generating and repairing FermiPy analysis pipelines."

这招只对那些本来就烂的小模型有用。也有人试过搞语言特化 LoRA 或者重新训练，基本没啥帮助，通常还更差。这个思路其实只适用于那些整体编程能力就很差的小众烂语言，你可以推着它们在一个具体任务上不那么烂。Meta 在这方面做了不少研究。要点是，一个优秀“通用”模型往往能把同等水平的特化模型按在地上摩擦，除非你在那个特化方向上做得极其深入。所以与其搞“Python LoRA”，你最后得到的更像是“针对某家公司代码库和规范的 FastAPI/PostgreSQL 后端 agent LoRA”，或者“专精于生成和修复 FermiPy 分析管道的科学 Python LoRA”。

>>109448785

>tiny shit languages that are bad at programming

>小众烂语言，编程能力差

tiny shit language *models that are bad at programming

小众烂语言 *模型，编程能力差

So, are the resets over? Only got 8% left of my limit this week for chatgpt.

所以说，重置结束了吗？这周 ChatGPT 的额度只剩 8% 了。

图片: https://i.4cdn.org/g/1785774962937921.jpg

>>109448785

I suspect it's vastly different between trying to improve a model without having access to a better model and distilling a smaller model on a bigger model with task specific data. The second can probably work much better when you only want to copy a bigger model's behavior rather than just making a model better at a task without a reference.

我觉得这两种情况差别巨大：一种是没法接触到更好的模型就去改进现有模型，另一种是用任务特定数据在更大的模型上蒸馏出一个小模型。后者可能效果好得多，当你只是想复制大模型的行为，而不是没参照地硬提升某个任务的能力时。

Hoo boy, here we go.

哎哟喂，来了来了。

图片: https://i.4cdn.org/g/1785775144786192.png

>>109448827

NTA but to the extent that it will make distillation cheaper yes, since FT has a nasty habit of reducing out of sample performance so it's cheaper to create a good FT dataset for specific tasks.

NTA，但要说它能让蒸馏更便宜，那是对的，因为 FT 有个坏毛病就是降低样本外表现，所以为特定任务造一个好的 FT 数据集更划算。

>>109448773

openrouter cache hit performance is awful

OpenRouter 的缓存命中性能烂得很

i lost t~4ish dollars in a day trying out deepseek because of a. the horrible autorouting that invalidates cache + when i turned that off my cache hit rate was still about 80%

我一天试 DeepSeek 就亏了大概 4 美金，原因有两个：a. 那垃圾自动路由把缓存搞失效了 + 就算我关掉它，缓存命中率也就 80% 左右

with ds direct my cache hit rate is over 98% and 4 dollars is like a week of light/med usage

直连 DS 的话，我缓存命中率超 98%，4 美金够我用一周的轻度/中度使用了

>>109448813

switch to luna to ease your addiction

换到 luna 来缓解你的瘾吧

fortunately that little thing runs slow as heck

幸好那玩意儿跑得贼慢

>>109448813

Maybe, I expect one reset soon when they bring back the 5 hour window and another reset when they release astra but that's probably a few weeks away

有可能，我估摸着他们恢复 5 小时窗口那会儿会重置一次，发布 Astra 的时候又会重置一次，但那大概还得等几周

>>109448844

The right to keep and bare cyberweapons shall not be infringed.

持有和携带网络武器的权利不可侵犯。

>accidentally leave Fable on and it starts on implementation instead of planning

>不小心开着 Fable，结果它直接开搞实现而不是先规划

Fuck it, nowhere near weekly limit anyway and it rolls over tomorrow, might as well slam it.

管他呢，反正离周限额还远得很，明天就滚蛋重来了，不如直接干到底。

>>109448785

i think programming patters will bleed into each other between languages when you train many. Happens to human coders too mixing up syntax and creating bugs cause it still compiles.

我觉得你要是同时训练多种语言，编程模式之间会互相渗透。人类程序员也会这样，语法混着来，搞出bug，结果还能编译过。

And secondary your model of whatever size has to encode everything you train it on. If you specialize then the entire available space can be used for the specific task which should automatically create a better encoding unless the model size is so big it can fit everything in in first place.

还有，你训练出来的模型不管多大，都得把所有喂给它的东西编码进去。你要是专精一门，那全部可用空间就能花在特定任务上，编码质量自然更高，除非模型大到一开始就能容纳一切。

>>109448869

Then don't use the autorouting. It's optional. You can select one specific provider, you have control of fallbacks, you can change how the prioritization works, exclude lower quants, all sorts of shit.

那就别用自动路由啊，那是可选的。你可以挑一个具体的提供商，回退也由你控制，优先级怎么排、排除低量化版本，这些都能改，一堆选项呢。

>>109448718

why does windows 11 have memory leaks?

为啥Windows 11老有内存泄漏？

What I'm currently working on

我现在在搞的东西

图片: https://i.4cdn.org/g/1785775669486598.png

>>109448933

maybe read the post you're replying to you stupid faggot

能不能先看看你回复的那个帖子啊，你个傻逼

>when i turned that off my cache hit rate was still about 80%

>我关了那玩意儿之后，缓存命中率还是大概80%

>>109448982

Maybe read the OpenRouter documentation, you retarded faggot, or you can keep complaining about the issue you caused for yourself and refuse to fix. I could help you, but I won't, because again you're a retarded faggot so it'd be a waste of my time.

要不你去读读OpenRouter的文档，你个傻逼蠢货，要么你就继续抱怨自己捅的篓子，不肯修。我可以帮你，但我不帮，因为你就是个傻逼蠢货，帮你纯属浪费时间。

>>109448672

>to autocomplete the words of a genius you need to be at least as smart as a genius

>要自动补全天才的话，你至少得和天才一样聪明

no, you just need a dataset where you can look up which formulations are most likely to appear

不，你只需要一个数据集，能查哪种说法最可能出现就行。

its kinda how we decoded sumerian (iirc. one of the old languages)

我们破解苏美尔语（我记得是古语言之一）就是靠这个路子。

look for root suffixes, words, etc (tokens)

找词根后缀、单词之类的（token）。

correlate them.

把它们关联起来。

we didnt need sumerian to write phrases in it

我们也没必要会用苏美尔语写句子。

you dont need to be a genius to autocomplete a phrase

你没必要是天才才能补全一句话。

you know what?

知道吗？

live example:

现场举例：

"A manifold is a topological space th*t is locally Euclidean, mea*ing every point has a neigh***hood homeomorphic to an open subset of Euclid*** space. While the global structure may be complex (e.g., a sphere or torus), zooming in on any specific point reveals a shape that resembles flat n-dimensional space. "

流形是一种拓扑空间，它在局部上等同于欧几里得空间，也就是说，每个点都有一个邻域，与欧几里得空间的开子集同胚。虽然全局结构可能很复杂（例如球体或环面），但放大观察任何特定点，都会看到类似平坦的n维空间的形状。

you dont need to understand what that says to find out the missing letters.

你不需要理解这段话什么意思，也能把缺的字母补出来。

you deduce them based on the rest of the text

你是根据上下文猜出来的。

its a brutal oversimplification, but its the essence of whats happening inside an llm

这是个极其粗糙的简化，但这就是大语言模型内部的本质。

>>109447707

nobody was buying apps before and they're not buying them now, vibe coding them faster doesn't matter because nobody wants to pay for apps or SAAS now.

以前没人买应用，现在也没人买，vibe coding写得再快也没用，因为现在压根没人愿意为应用或SaaS付钱。

>>109449002

buy a fucking ad, cocksucker

买个广告吧，你个狗娘养的。

>>109449006

the guy isn't acting in good faith and i'm not sure if you are either

那家伙根本没在好好说话，我也不确定你是不是。

go to ibm skillsbuild if you wanna learn how ai works

想学AI怎么运作，去IBM SkillsBuild看看。

>>109449025

Source?

>>109449006

you only deduce th"t by knowing that exists and nothing else.

你只能靠知道它存在来推断，别的啥都不知道。

If its n"ne the answer might be none or nine and you cant even infer the correct term in a conversation like "how much Beer packs did you buy?" because both nine and none are valid answers.

如果填的是“n"ne”，答案可能是“none”也可能是“nine”，在这种对话里——“你买了几箱啤酒？”——你根本没法判断哪个对，因为两个都说得通。

This is why sumerian is also debated as there is no certainty in translation.

这就是苏美尔语也一直有争议的原因，翻译没法百分百确定。

>>109448079

Don't forget to buy at the peak of the dead cat bounce!

别忘了在死猫反弹的最高点买入啊！

>>109449050

We're not here to talk about sumerian languages we are on /g/ technology in /vcg/ - vibe coding general.

我们这儿不是聊苏美尔语的，我们在 /g/ 技术版，在 /vcg/ - vibe coding 综合讨论里。

sol analyzed qwen 3.8's benchmarks, and concluded that it might be around kimi k3 / grok 4.5 level, based off of benchmarks alone admittedly. it said it'd place it at around 60-66 on the coding index

sol 分析了 qwen 3.8 的基准测试，得出结论它可能接近 kimi k3 / grok 4.5 的水平——不过明说了，这只是基于基准测试数据。它说编码指数上大概能排到 60-66。

图片: https://i.4cdn.org/g/1785776402882299.png

>>109449039

look at the sales and who sells them

看看销售数据和谁在卖这些东西。

>>109449065

You can go pull the revenue from Google Play Store and iOS app store, I don't engage seriously with morons

你可以自己去拉 Google Play Store 和 iOS App Store 的收入数据，我没空跟傻逼认真扯。

>>109449060

the generalized idea is statistical prediction relies on accepting a uncertainty error.

核心概念就是统计预测必须接受一个不确定性误差。

>>109449039

>runit

i chatted with you in another thread

我跟你另一个帖子里聊过。

>>109448606

you shouldnt be talking about good faith

你压根没资格谈什么善意。

>ai

an *llm.

一个 *LLM。

yeah, i know ibm skillbuild, you quite obviously should make use of it.

对啊，我知道 IBM skillbuild，你显然该好好用用它。

>>109449036

I don't need to advertise that it's trivially easy to hit 98%+ cache hit rate with DeepSeek through OpenRouter if you actually read the OpenRouter documentation. I guess you're just too busy thinking about sucking cock to read the fucking manual. Is it hard? Not that cock you're imagining, I mean living with such a profound learning disability, is it hard?

我不用特意宣传，只要你真去读 OpenRouter 文档，就知道用 DeepSeek 打 98%+ 缓存命中率是轻而易举的事。我猜你光顾着想着吃屌，懒得读那破手册。难吗？不是你想的那根屌，我是说带着这么重的学习障碍活着，难不难？

>>109449050

>If its n"ne the answer might be none or nine

>如果答案是 n"ne，那答案可能是 none 或者 nine

you deduce it from the context and from the preceding, and following tokens just like an llm does.

你得从上下文和前后 token 里推出来，就像 LLM 干的那样。

if youre unsure, you toss a coin

拿不准就抛硬币。

>>109447707

>that pic

>那张图

Impressive lack of self-awareness.

自我认知能力差得惊人。

>everything ends up being a barren wasteland, the only remaining pAIjeet living in a jail of his own delusions

>一切最终都变成荒芜之地，剩下的只有那个活在自己妄想监狱里的 pAIjeet

>>109449006

Human behavior can be reduced to a table lookup too.

人类行为也可以简化成查表。

Make a two column table that on the left has every possible configuration of subatomic particles in your body down to 0.0000001 femtometers. On the right it has the configuration 0.0001 femtosecond later.

做一张两列表格，左边列出你身体里所有亚原子粒子从 0.0000001 飞米尺度起的每一种可能构型，右边写 0.0001 飞秒后的构型。

With this table simulating a human is as easy as looking up its current state, then replacing it with the next state as stated in the table.

有这张表，模拟一个人就跟查当前状态差不多简单，然后按表里说的用下一个状态替换。

>>109449062

the coding benchmark doesn't look impressive for a 2T model, but if it can really design a chip in 12h then....

对一个 2T 的模型来说，这编码基准看着不咋样，但如果它真能 12 小时设计出芯片，那就另说了……。

>>109449074

that's what i said, moron. very few independent developers sell at all, definitely not vibe coded either

我他妈就这么说的，傻逼。独立开发者能卖出去的就极少数，更不可能是 vibe coding 出来的。

>>109449059

I've lost big money by shorting, twice. I never lost any significant amount of money by longing anything.

我做空亏过大钱，亏过两次。我做多从来没亏过这么多。

>>109449094

You're still a delusional retard. I have taken skillsbuild. You don't like me? Okay, then stop replying to me. But I didn't say anything wrong there. If you don't like technology then go to a different board for a hobby that you actually like.

你还是那个妄想症傻逼。我上过 skillsbuild。你不喜欢我？行，那就别回我。但我那话没说错。你要是不喜欢技术，就滚去别的版块找个你真喜欢的爱好。

>>109449119

yeah and the rubber meets the road and you realize that actually implementing that is pointless

对，等真落地实操的时候，你就发现这玩意儿根本没法实现。

you dont have enough data, or enough memory even to create a model that will begin to be coherent

你数据不够，内存也不够，连个勉强连贯的模型都建不起来。

its the same problem with llms

LLM 也是同样的问题。

THEORETICALLY llms could get us to agi

理论上 LLM 确实可能带我们到 AGI。

its just that its computationally unrealistic. with that method at least

只是计算上不现实。至少用那条路是没戏的。

and because the ceos are kiddiefucking inbred retards their solution to a scaling problem obviously spiralling into exponentiality...

因为那些CEO都是近亲繁殖的弱智，面对明显呈指数级恶化的扩展问题时，他们想出的解决办法——

is to throw exponentially more compute at the problem

就是往问题上砸更多算力

too much adrenochrome

肾上腺素吸太多了

not enough public beatings

公开处刑没看够

>>109449137

i heavily doubt it

我严重怀疑

youre using the most basic terms wrong

你连最基本的概念都用错了

anyhoo, even if you did

反正就算你懂

youre too retarded to understand what youre learning

你也蠢到根本理解不了自己在学什么

aaand i gtg. cya later, aligater

好了我得走了，回头见，小鳄鱼

>>109449163

Thanks for your concession, Vlad.

多谢你认输，Vlad。

>used for web shit

>用来搞网页开发

>still a lot of tokens left on a 20$ plan

>20刀的套餐还剩一堆token

I guess web shit is easy for Claude

我看网页开发对Claude来说就是小菜一碟

图片: https://i.4cdn.org/g/1785777154609342.png

>>109449155

It's the other way around.

正好反了。

LLMs work because they have some kind of cognitive process to some extent similar to human though.

LLM能工作是因为它们在某种程度上拥有与人类思维相似的认知过程。

Storing all the possible scenarios and doing some kind of dumb interpolation to find the answer would be intractable in terms of memory.

把所有可能场景都存下来再做某种蠢插值来找答案，那在内存上根本不可行。

>>109448079

>If so then that was the mildest bubble I've ever seen.

>如果真是那样，那这就是我见过最温和的泡沫了。

The dot com bubble took two years to reach the bottom... It seems you have no idea how the market works.

互联网泡沫用了两年才触底……看来你对市场怎么运转完全没概念。

>>109449099

you're the one who can't read a whole post before you reply to it, assclown

你才是那个回复前连帖子都没读完的人，傻逼

fuck off

滚蛋

>>109449187

Stop responding to people who bring up this bubble bullshit. This is /g/ not /biz/. Have some fucking standards.

别再去理那些扯泡沫的人了。这里是/g/不是/biz/，能不能有点基本水准。

>>109449194

Then don't use a stupid image for an OP to bait me in.

那就别放张傻逼图当主贴来钓我。

>>109449201

I wasn't the OP but I see where you're coming from, desu.

我不是楼主，不过我懂你意思，说真的。

图片: https://i.4cdn.org/g/1785777373234575.jpg

>>109449178

Webshit is easy. Cryptography, drivers, databases, kernels and other notorious "this is really hard" stuff is where your tokens disappear faster than you can count.

网页开发很简单。密码学、驱动、数据库、内核这些公认"巨难"的东西才是你token消失得比数钱还快的地方。

>started chatgpt plus trial 4 days ago

>4天前开了chatgpt plus试用

>still at 100% with the reset date rolled back to 11th of august

>还是100%，重置日期被推到了8月11号

i just can't make my mind up on what to prompt for my project...

我就是拿不定主意该给我的项目用什么提示词……

>>109449209

I assume most token-burning here happens because of game dev.

我猜这里大多数人烧token都是在搞游戏开发。

>>109449187

If you really believe that then show your short positions. Surely you have them, right?

你要真这么觉得就把你的空头仓位亮出来。你肯定有吧？

>>109449155

>THEORETICALLY llms could get us to agi

>理论上LLM能带我们到AGI

>its just that its computationally unrealistic

>只是计算上不现实

We don't really know that. Both the compute requirements of AGI and emergent capabilities of future optimized and scaled-up LLMs are currently unknown. And the only way to know is to simply try it and find out.

这真不好说。AGI的计算需求以及未来优化扩容后LLM涌现出的新能力，现在都是未知数。唯一搞清楚的办法就是直接试试看。

>>109449190

>bloo bloo bloo

>哇哇哇哇

You can't write in complete sentences, you don't use capitalization or punctuation, and you can't read the fucking manual, what good are you? Get off 4chan and spend some time in front of the mirror you illiterate cock-gargling faggot.

你连完整句子都写不了，不用大写不使标点，连他妈的手册都懒得读，你有个屁用？滚下4chan去照照镜子吧，你个文盲舔鸡巴的废物。

>>109449228

Couldn't tell you, I moved on to low-level stuff pretty quickly. I have more fun with that.

没法告诉你，我很快就转到底层开发了。那玩起来更有意思。

>>109449215

rewrite codex in wpf

用WPF重写codex

make no mistake

别搞错了

LOL my cloudflare account got instanuked for trying to push torrents through warp

哈哈我的cloudflare账号因为想通过warp推种子直接被秒封了

图片: https://i.4cdn.org/g/1785777726137372.png

>>109449241

l2read

>>109449260

Rich coming from you who can't read the OpenRouter documentation. Have you fixed your cache hit rate yet or did you just run away to a different provider because you couldn't bring yourself to read the fucking manual? I read your post, it made it clear you didn't know what you were doing, and instead of discussing it you decided to be a vitriolic illiterate faggot, sub-human behavior.

你自己连OpenRouter文档都读不懂，还好意思说这话。你的缓存命中率修好了没，还是说干脆跑路换了个服务商，因为实在没脸读那破手册？我看过你的帖，明摆着你不懂自己在干啥，不讨论正事反倒跑来阴阳怪气，真是低智又下作，简直不像人干的事。

>>109447985

How many t/s we talkin here? Is it slower than k3?

咱说的是每秒多少token？比k3还慢？

>>109449287

Me and you might not get along, but I have to admit, you're 100% right here. I was walking down the street last night, and I was musing to myself about how if I were as nuts as I used to be, I'd probably just consider strangling that fucker to death. Maybe my jimmies are rustled but he's just legitimately one of the worst posters on this entire website. He's so recognizable too. I really have patience for a lot of things but wilful stupidity is not one of them.

咱俩可能不对付，但我得承认，你这话一点没错。昨晚我走街上，自己琢磨着，要是我还像以前那么疯，估计就直接想把这货掐死。也许我是有点上头，但他就是这站上最烂的发帖人之一，特征还特明显。我忍得了很多事，但故意犯蠢这毛病忍不了。

图片: https://i.4cdn.org/g/1785778151755388.png

>>109449303

nta It's noticeably faster than K3, token rate feels double, but the reasoning is so lengthy they feel comparable in throughput so far. Which is to say, slow, they're both obnoxiously slow.

非本人。明显比K3快，token速率感觉翻倍了，但推理过程太长，目前吞吐量感觉差不多。也就是说——慢，俩都慢得让人抓狂。

>>109449336

>>109447985

but is it kimi k3 level?

但够得上kimi k3水平吗？

>>109448555

low for asking questions, file organizations, simple stuff

低档用于问问题、整理文件、简单活儿。

medium for scripts/utilities i know can be one-shot without much guidance

中档用于我确定能一把过、不用多指导的脚本和工具。

high for generally involved tasks

高档用于一般复杂任务。

haven't had a use case for anything above high yet. I used fable to make decompilers and it bruted XOR encryption of some hentai games. it was awesome

目前还没碰过需要超高档的活儿。我用fable做过反编译器，还暴力破解了几个黄油的XOR加密，爽翻了。

>>109449303

>How many t/s we talkin here? Is it slower than k3?

>每秒多少token？比k3慢吗？

Don't know but it takes like 3 minutes or + for it to think about a topic. I wanted it to write some code. on chat.qwen.ai, I don't see the t/s.

不知道，但它想一个话题得花3分钟或更久。我想让它写点代码，在chat.qwen.ai上，我看不到t/s。

>>109449350

It was just a powershell script using COM.

就是个用COM的powershell脚本。 ⟦ 也许吧？我对K3没多惊艳，对Qwen3.8 Max也没多感冒。不是说它差，只是不对我胃口。反倒GLM 5.2和之前的MiniMax M3更让我眼前一亮。K3纸面上看着牛，但实际体验就是“卧槽又贵又慢”。

>>109449337

Maybe? I didn't feel that impressed with K3 and I'm not that impressed with Qwen3.8 Max. Which is not to say that it's bad, just that it's not for me. I was more impressed with GLM 5.2 and before it MiniMax M3. K3 looks awesome on paper but my experience with it was "wow this is expensive and slow."

也许吧？我对K3没啥感觉，对Qwen3.8 Max也就那样。不是说它烂，就是不对我胃口。我更看好GLM 5.2，之前是MiniMax M3。K3纸面上看着牛，但我的体验就是“哇，又贵又慢。”

https://brucebyfield.com/2009/05/29/willful-stupidity/

This was written in 2009 and it couldn't be more relevant to the behaviour I am seeing.

这写于2009年，但跟现在看到的这行为简直不能再贴切了。

图片: https://i.4cdn.org/g/1785778526400751.png

>>109449361

fair enough. does qwen at least one shot the tasks? you don't need to tardwrangle it on follow ups?

说得对。那qwen至少能一把过任务吗？不用你反复跟它掰扯后续？

>>109449361

I'm not you but I was not impressed with K3 either. It's decent, good even, but it's painfully slow, a token hog, and isn't even close to fable at all. Even it's much touted frontend skills, while absolutely decent and better than most humans, are still not revolutionary

我不是你，但我对K3也没多感冒。它还行，甚至算不错，但慢得要命，吃token大户，跟fable根本没法比。就算它吹上天的前端能力，虽然绝对过硬、比大多数人强，但也没到革命性的地步。

>>109449317

We always get along, nonspecific tripfag.

咱俩向来处得不错，没署名的tripfag。

Is there a good reason to use Codex via the terminal instead of the ChatGPT/Codex app? It seems cleaner and easier to write and format proompts in the app than the terminal.

用终端跑Codex，跟用ChatGPT/Codex应用比，有啥正经好处吗？在应用里写和排版提示词感觉更干净也更容易。

>>109449287

tl;dr

>>109449390

Preference is a good enough reason. I've been using the terminal for the last 11+ years. I don't see any reason because some fancy IDE with AI features came out. I also don't use Codex because ChatGPT partially funds the genocide in Gaza. Meanwhile you press disinformation about Anthropic's CEO, plus you intentionally ask this question, because you know my preference is for the TUI, no matter which harness that I use. Congratulations, I hope this attention is enough to sustain you, parasite.

偏好本身就是个充分的理由。我用终端已经11年多了，不觉得因为出了什么花哨带AI功能的IDE就得换。我也不用Codex，因为ChatGPT部分资助了加沙的种族灭绝。另外，你散布关于Anthropic CEO的假消息，还故意问这个问题，因为你明知道我就偏好TUI，不管我用哪个框架。恭喜你，希望这点关注够你活下去，寄生虫。

(无正文)

图片: https://i.4cdn.org/g/1785778966950028.png

>>109449431

>t. subhuman

>t. 亚人类

glm 5.3 coming soon.

glm 5.3 快来了。

gemini 3.5 pro coming soon

gemini 3.5 pro 快来了

grok 4.6 coming soon

grok 4.6 快来了

holy moly, vibelords eating good

我靠，氛围党们吃得真香

图片: https://i.4cdn.org/g/1785779019580022.png

>>109447997

>41%

what did they mean by this

他们这话啥意思

图片: https://i.4cdn.org/g/1785779050614610.jpg

>>109449390

The terminal is a lightweight and comfortable traditional developer user interface. Another thing is that Codex app is still not available for Linux.

终端是轻便又舒服的传统开发者界面。另外，Codex应用还没出Linux版。

>>109448502

Says who? Until we hear it straight from the horse's mouth that's just pure speculation, inless there's some recent announcement that I missed.

谁说的？除非我们从官方嘴里听到，不然就是纯瞎猜，除非我漏了最近什么公告。

>>109447997

As >>109441855

The Qwen models (35B Moe and 27B dense) are the models you want to use for coding. Any other general purpose stuff, that's what Gemma is good at (including raunchy NSFW RP apparently. I've never tested that so I can't attest to it's capabilities in that department).

Qwen系列模型（35B MoE和27B dense）才是写代码该用的。其他通用任务，Gemma擅长那些（包括那种露骨的NSFW角色扮演，反正我没测过，没法评价那方面能力）。

My laptop has 128 GB of unified memory so it's more than powerful enough to run both at full context, don't like typically use the moe moe one because the prefill and token generation is a good bit faster than the dense version even when nearing full context. The dense version is "smarter" on paper but the time you have to wait from submitting your prompt to it completing the token generation is painfully slow at Large contexts. It's annoyingly slow with the moe one too, especially if you're too used to API speeds, but the dense one is so much worse that I basically dropped it and replaced it with the moe one unless

我笔记本有128GB统一内存，满上下文跑两个都绰绰有余，但我一般不用MoE那个，因为预填充和token生成速度比dense版快不少，就算快满上下文也是。dense版纸面上更“聪明”，但大上下文下从提交提示到生成完token的等待时间慢到让人抓狂。MoE版也慢得烦人，尤其你习惯了API速度的话，但dense版差太多了，我基本放弃了它，换成MoE版，除非——

1) there's a very specific problem the moe keep struggling with after hand holding and

1）MoE在反复引导后还是搞不定某个特别具体的问题，而且

2)n I happen to not want to or not be able to use the API models.

2）我碰巧不想用或没法用API模型。

Short answer is that I'll be fine for the most part. It's the people that are completely and utterly reliant on the cloud providers that might be screwed a little bit. I highly doubt they'll go away but best case scenario for them is that the pricing will remain roughly the same or get a slight price hike. Worst case scenario is that they'll still be able to use the models but because subsidization WILL have to stop eventually, they'll get fisted with the token prices increasing a lot.

短答就是大部分情况我没事。真正可能有点惨的是那些完全依赖云服务商的家伙。我强烈怀疑它们不会消失，但对它们最好的情况也就是价格基本不变或稍微涨点。最坏情况是模型还能用，但补贴早晚得停，token价格会暴涨，直接被干翻。

图片: https://i.4cdn.org/g/1785779110527507.png

>>109449473

Codex agent is available for GNU/Linux, though. I've seen it. Standard shit, just like Gemini CLI, Claude Code, or Antigravity.

Codex agent倒是支持GNU/Linux，我见过。标准货色，跟Gemini CLI、Claude Code或Antigravity差不多。

>>109449489

I meant the desktop version.

我说的是桌面版。

>>109449477

You missed the announcement. We get Qwen3.8 27B in the next week or so here.

你错过公告了。我们这边下周左右就有Qwen3.8 27B了。

How it started: pic related

开始时：见附图

How it's going:

现在的情况：

>here's a list of 10 bugs, make a plan and then start fixing with the first one

>这里有10个bug的列表，做个计划，然后从第一个开始修

>it's now a list of 30 bugs and 10 are fixed

>现在变成30个bug，修好了10个

>6 separate sessions have been at it

>6次独立的会话都在搞这个

>testing loop takes 2 minutes because it needs to wait for race conditions to happen

>测试循环要跑2分钟，因为得等竞态条件出现

图片: https://i.4cdn.org/g/1785779293665696.gif

>>109449503

No such thing. You can run the agent on your desktop just fine. It's just the GUI is not available for Linux. If there is a gap in features between the GUI version and the CLI version I'm unaware of them, like I said I don't use OpenAI shit.

不存在这种事。你在桌面上跑这个agent完全没问题，只是Linux上没GUI版。GUI版和CLI版之间要有什么功能差距我也不清楚，我说过，我不用OpenAI那破玩意儿。

>>109449504

Link

>>109449503

nta. Codex-CLI is linux native, but you can also actually run the Mac OSX version on Linux. It's unofficial as fuck but there are at least two projects out there that accomplish exactly this.

nta。Codex-CLI本来就是Linux原生的，但你还真能在Linux上跑Mac OSX版。非官方到爆，但至少有两个项目就干这个的。

>>109449523

https://x.com/Alibaba_Qwen/status/2084100707423289643

图片: https://i.4cdn.org/g/1785779413744770.jpg

>>109449477

>>109449533

inb4 the new small model is worse than 3.6 after they fired all the old devs

先说了，等他们把老开发全开了，新小模型肯定比3.6还烂。

>>109449545

Please don't reach into my psyche and pull my fears to the surface, it's rude.

别tm伸手往我脑子里掏、把我恐惧揪出来摆脸上，很没礼貌。

图片: https://i.4cdn.org/g/1785779654111860.png

all the models have decided that this project is British and are spelling words like colour and modelling

所有模型都认定这项目是英国的，开始拼colour、modelling这种词了。

I don't know when it happened, I feel colonised

我不知道啥时候开始的，感觉被殖民了。

>>109449634

their cooking recipe performance is getting worse too for some reason

它们的菜谱表现不知为啥也越来越拉了。

geminilads... we've become the laughing stock of the world...

双子兄弟们……我们成了全世界的笑柄……

图片: https://i.4cdn.org/g/1785780216963099.png

>>109449650

I like gemini flash, not for coding but it's a nice model. I'd rather not start console warring models, leave that to the retarded masses

我喜欢Gemini Flash，不拿来写码，但模型本身不错。我可不想搞模型饭圈大战，那留给那些脑残大众去玩。

^^@~.......

>>109449439

图片: https://i.4cdn.org/g/1785780358401473.jpg

>>109447707

Why does it feel like every time a new model comes out, the older model which was functioning fine becomes utterly useless?

为什么每次新模型一出，那个之前还好好的老模型就瞬间废了？

Particularly anthropic feels this way.

尤其是Anthropic给人这感觉。

I only use Chinese models, because I know that I directly fund at least a few genocides

我只用中国模型，因为我清楚自己至少直接资助了几场种族灭绝。

>>109447707

Who here has tried out the Gauntlet loop?

有人试过Gauntlet循环吗？

After seeing Claude of Duty and the 3d Pokemon game I gave "Gauntlet" vibe coding a try (no fable just opus) and it actually seems to be super powerful, mind blown. Would recommend. You can copy in an article about gauntlet ai looping and the Claude of duty prompt on GitHub and say to do something similar for your project (whether from scratch or as an overhaul/upgrade)

看了Claude of Duty和那个3D宝可梦游戏之后，我试了一把"Gauntlet"式vibe coding（不整fable，就上opus），居然真特么强，惊了。推荐试试。你可以去GitHub上复制一篇关于Gauntlet AI循环的文章和Claude of Duty的提示词，然后让你自己的项目照葫芦画瓢（从头干或者当大改/升级都行）。

Let me know how it goes, I'm really curious to see if everyone gets good results!

试完跟我说说结果，我特好奇是不是人人都能出好效果！

图片: https://i.4cdn.org/g/1785780622916599.jpg

>>109449686

There are no older models. There never were.

没有什么老模型。从来就没有过。

>>109449439

Misanthropic target painted girls school in Iran for missile strike

Misanthropic瞄准了伊朗的女子学校准备导弹打击。

>>109449692

speaking of, I'm surprised there's no Israeli AI lab churning out banging weights. They've got their fingers in a lot of frontier tech, why not AI? Could be that there's no need for nationally branding their influence since US companies are already zogged up, no need to compete when you already own the big ones

说到这个，我挺意外居然没有以色列的AI实验室在狂出猛料。它们手伸进好多前沿技术了，怎么AI不掺一脚？可能是因为没必要用国家品牌来标榜影响力——美国公司早就被控住了，自家已经捏着大厂了，还竞争个啥。

>>109448376

The prompt was for a desktop rapid serial visual presentation reader/blog where you add a text file, it saves it in a sqllite database, and you can play the file at around 300wpm. Similar to the Star Wars telnet or speed reading web extensions if you have seen them. The first letter of each word needs to be a different color, and optionally slight pauses after each punctuation mark. I wanted the app to be made in F#, Haskell, and Rust with cross-platform GUIs. For Haskell, it can be done using web stuff. 2 or 3 screens, the home screen with the list of posts, a screen to upload/add posts, and the post itself with playback controls (start/stop/speed up/slow down). This cost about $1.25 on openrouter with deepseek+luna. I have not tried this with more expensive models unless you count antigravity. Sometimes I will also run the same prompt with ocaml, java, typescript, and a few other languages for fun. Also I normally draw the screens by hand, take a picture with my phone, and put the image in the directory because its faster for me that way.

这个需求是做桌面端的快速序列视觉呈现阅读器/博客应用，你添加一个文本文件，它会存进SQLite数据库，然后能以大约300词/分钟的速度播放文件。类似《星球大战》里的远程登录或者浏览器上的快速阅读扩展，如果你见过的话。每个词的首字母需要用不同颜色显示，标点符号后还可以选择性地稍作停顿。我希望这个应用用F#、Haskell和Rust编写，并且要有跨平台GUI。对于Haskell，可以用网络技术来实现。需要2到3个界面：主页显示文章列表，上传/添加文章的界面，以及文章本身带播放控制（开始/停止/加速/减速）。在OpenRouter上用deepseek+luna大约花了1.25美元。我还没试过更贵的模型，除非你把antigravity也算上。有时候我也会拿同样的需求去试试OCaml、Java、TypeScript还有其他几种语言，纯属好玩。另外，我通常手绘界面草图，用手机拍照后放进目录里，因为那样对我来说更快。

>>109445114

no mythos is literally the exact same model as fable fucking retard

不，神话就是和寓言完全一样的模型，你他妈是傻逼吧

google it

自己谷歌去

Claude keeps bitching about me telling it to have a subagent adversarially criticize a plan but is like "damn this was a good idea" every time.

Claude老是在我让它用一个子代理对立地批评某个方案时唧唧歪歪，但每次最后都说“卧槽这主意真不错”。

>>109449755

OAI and Ant, the best AI labs in the world, are both owned by the Jews. Literally.

OAI和Ant，全世界最厉害的AI实验室，都是犹太人开的。真的。

>>109449686

Anthropic doesn't have compute. They can't leave several models running at full quant with good availability, they just don't have the capacity. They rent the vast majority of their compute, and they distribute what they have based on what they're selling and what people are using. Older models get dropped to shittier quants and run on shittier hardware because they're less important and they expect people to stop using them. They all do this, but it's particularly dramatic with Anthropic because of the disparity between the capabilities of their models and their piddling little hardware allowance when compared to OpenAI.

Anthropic没有计算资源。他们没法让几个模型同时以全量化保持良好可用性运行，就是没有那个能力。他们大部分算力都是租来的，然后根据在卖什么、谁在用，来分配手头这点资源。老模型会被降级到更烂的量化级别，跑在更烂的硬件上，因为不重要了，他们也预期用户慢慢会不用。大家都这么干，但在Anthropic身上特别明显，因为跟OpenAI比，他们的模型能力和那点可怜巴巴的硬件配额之间差距太大。

>>109449171

none was given

没有提供任何内容

>vlad

wtf is that about?

这他妈是在说什么？

>>109449184

sorry anon, but you know shit and only from hearing

抱歉兄弟，你屁都不懂，全靠道听途说

its all querying jewgle for results and then basing a response based off the training and the results found

不就是在问Jewgle搜结果，然后基于训练数据和搜到的结果来生成回答嘛

just fucking ask an llm, rite?

直接问个LLM不就行了，对吧？

its gonna pull up at least some documentation in the background, and re-phrase it in a way you can understand it

它会在后台拉出至少一些文档，然后用你能理解的方式重新说一遍

its a very powerful tool, its the natural evolution of google search

这是个非常强大的工具，是谷歌搜索的自然进化版

but like any tool, it has its limitations, and usecases

但跟任何工具一样，它也有局限性和适用场景

>>109449238

we do know that

这个我们知道

were firmly in the domain of diminishing returns at this point

现在已经稳稳处在收益递减的区间了

>but line goes upp!

>但曲线在往上走啊！

yes it does, but its actually two lines-

是的，在走，但其实是两条线——

the capabilitie proper, and the compute thats been used to get there

一条是能力本身，另一条是为达到这个能力所消耗的计算量

and the compute increases exponentially

而计算量是指数增长的

>>109449755

Israel is a focal point for military AI, not consumer AI. Glow, Intezer, their National AI Program, their goals are much more focused on things like surveillance, espionage. They have a bespoke AI they use to pick bombing targets that they named "The Gospel" and other similar ones for related tasks.

以色列是军事AI的焦点，不是消费级AI。Glow、Intezer，还有他们的国家AI计划，目标都更偏向监控、间谍这类东西。他们有个定制AI用来选轰炸目标，起名叫“福音”，还有别的类似工具干相关活儿。

>>109449774

open ai will buy anthropric with money magic

OpenAI会用钱魔法把Anthropic买下来

>loop engineering

>循环工程

>graph engineering

>图工程

how much of this stuff is just AI psychosis

这些东西里有多少只是AI精神病发作

How do I make it stop spending so many turns trying to figure shit out? I think I'll just tell it to not write test cases.

怎么让它别花那么多轮次去瞎琢磨？我觉得干脆直接告诉它别写测试用例算了。

图片: https://i.4cdn.org/g/1785781434255324.png

>>109449793

yes you can increase it back to the original 370k but it'll eat you usage much faster which is why it was disabled in the first place

行，你可以把它调回原来的370k，但那样烧你的用量会快得多，当初关掉就是因为这个。

图片: https://i.4cdn.org/g/1785781643117356.png

>>109449824

None, if you are not building loop graph trees, you are behind and a permanent underclass.

零，你要是没在搞循环图树，你就落后了，就是永远的底层人。

I think claude takes the cake for most school children killed per ai model. Maven smart systems.

我觉得Claude在“每个AI模型害死学童数”这个榜上能拿冠军。Maven smart systems干的。

>>109449801

I am working on implementing MoE tensor parallelism for DS V4 Flash in an LLM engine I made from scratch and I've been working on for a year so I highly doubt you know more about LLMs than me.

我正在给我从零写的LLM引擎实现DS V4 Flash的MoE张量并行，这引擎我搞了一年了，所以我高度怀疑你对LLM的了解能比我多。

I also achieved 0.7 tk/s on K3 with DDR4 MoE CPU offload on an Epyc 7713.

我还在Epyc 7713的K3上用DDR4跑MoE CPU卸载达成了0.7 tk/s。

>>109449824

It's not psychosis, it's just grifting and trying to get rich and famous as a zoomer shitfluencer.

这不是精神病发作，就是割韭菜，想当网红发大财的Z世代屎网红罢了。

>>109449958

ok, but what does it do

行，那它到底是干嘛的

if you understand it, you can explain it to me like im 5

你要是真懂，就用五岁小孩能听懂的话给我讲讲

and, in turn, i will understand it

然后，我就懂了

>>109448730

>it's real

>这是真的

lmao how did Jewgle fumble this hard?

哈哈 Jewgle怎么能把这事搞砸成这样？

>>109450057

Too busy selling and leasing their hardware to all the companies who are actually trying to be competitive.

忙着把硬件卖和租给那些真在拼竞争力的公司呗。

>>109450057

It was slated for June originally, dudes basically just missed an entire release cycle for no reason. I'm stuck with a google pro subscription since it was discounted but the models are so far behind it's ridiculous, deepseek flash rapes gemini 3.1 pro in my real use case.

本来定的是六月，结果这帮家伙就硬生生错过了一整个发布周期，没任何理由。我因为打折绑了个Google Pro订阅，但模型落后得离谱，DeepSeek Flash在我实际用例里直接碾压Gemini 3.1 Pro。

>>109450080

which is? even non coding?

哪个方面？连非代码的也算？

>>109450046

Make things as simple as possible, but not simpler.

把事情做得尽可能简单，但别简单过头。

>>109450074

>>109450080

common megacorp L

大公司日常拉胯

>no response from API

>API没响应

>it continues

>它还在继续

>makes dumb mistake after dumb mistake

>犯了一个又一个蠢错

bros I think I got silent-downgraded to Opus

兄弟们，我觉得我被偷偷降级到Opus了

>>109450110

godot gamedev

Godot游戏开发

I haven't bothered with non-code for deepseek

我没拿DeepSeek试过非代码的东西

gemini chat is useful enough for ideaguying but useless for any long-term notebook stuff because it will always devolve into a retarded sycophant no matter what you write in the personality instructions

Gemini聊天对瞎想想当点子王挺有用，但干任何长期笔记的活儿就是废物，因为不管你在人格指令里写啥，它最后总会变成个傻逼马屁精。

gemini for code is benchmaxxed to hell and only designed for one-shotting, meaning it will only do the bare minimum and gaslight you about the rest in any task dealing with already established codebases

Gemini写代码就是被刷bench刷到烂的，专门为一次性生成设计的，意思就是碰已经有代码基础的任务，它只干最低限度的事，剩下的全靠糊弄你。

图片: https://i.4cdn.org/g/1785783529303933.jpg

>>109450168

nta What's your experience been like, working with Godot and an AI Agent, and what sort of accoutrements are you utilizing? I've been using https://github.com/regiellis/godot-mcp-go because I don't want an actual MCP layer shitting things up and I like the live editor use, but I've also looked at skill collections such as https://github.com/jame581/GodotPrompter (promotes a specific workflow I'm not interested in) and https://github.com/thedivergentai/gd-agentic-skills and I know there are other MCP options as well. I'm just curious.

I'm trying to make a browser game just for fun and while the coding aspect is obviously no issue, I can't seem to figure out how people are making stuff with high fidelity art. I managed to make some pixel art but it honestly looks like shit. Are they prompting it through the same agents or are they using something external like grok and then plugging in the files into their project?

我想做个浏览器游戏纯粹图个乐，编码这块当然没问题，但我就搞不懂别人那些高保真美术资源是怎么整出来的。我也试着做了一些像素画，但说实话烂得跟屎一样。他们是用同款agent直接生成的，还是用Grok那种外部工具生成完再把文件塞进项目里？

>>109450168

Why do you have to choose godot?

为啥非得选Godot？ ⟦ 我跟Sol说要做个游戏，它直接整出来了，压根没走Godot。 ⟦ 它跟Kimi K3比能力咋样？K3还是开源模型里的扛把子吗？（Qwen 3.8 Max严格说还没开源，但按他们公告说快了） ⟦ 我目前在Antigravity和Opencode之间来回切，配GodotAI做MCP。AI对MCP理解得不错，用得也顺手，这块没啥好抱怨的——它让AI能刷新文件系统、正确生成UID这些，简直是救星。 ⟦ Gemini简直蠢得要死，连"脚本保持强类型"这种基本要求都能跳过去，还老生成损坏的场景文件。所以要是3.5 Pro不让我眼前一亮，我直接就跑了。我最近在试DeepSeek，目前还挺惊艳的。 ⟦ 我的建议是别把工作流搞得太复杂，就用个简单的agent CLI或者程序，别整那些编辑器里嵌聊天界面的傻逼插件和花里胡哨的玩意儿，能多原始就多原始。另外别偷懒，它生成的代码你真得去review，因为"垃圾漂移"是真实存在的。 ⟦ 以后可能会变好，但我现在也不敢让AI自己搭项目结构，你得先有个像样的代码库当范本，因为这些模型全是在教程垃圾上训练的。 ⟦ 别让任何人、任何事打击你搞vibecoding这条路。我可以跟你讲讲我做了啥，更重要的是我为这个拿到了啥回报，但后面这点说了你也不会信，所以没意义。而且我三年vibe的经验跟这成功基本没啥关系，因为现在有Fable这种模型。就是闷头干，一直prompt，把东西做出来。 ⟦ 因为我是先搞游戏开发、后搞slop开发的，AI就是个辅助工具，不该直接替你拍板。一把梭做个网页游戏拿来玩玩挺有意思的，但我不觉得这算正经用途。

I told sol to make a game and it made something without having to make it in godot.

我让sol做个游戏，结果它没在Godot里做就直接整出来了。

>>109449336

How do it's capabilities compare to Kimi K3? Is K3 still the king of open weights models? (Qwen 3.8 max technically is it open weights yet but will be soon per their announcement)

它的能力跟Kimi K3比怎么样？K3还是开源模型里的老大吗？（Qwen 3.8 max严格来说还没开源，但按他们公告，很快会放出来）

>>109450203

I'm hopping between antigravity and opencode, with GodotAI for the MCP. The AI understands the MCP and uses it well so no complaints there, it's a godsend for letting the AI refresh the filesystem so it creates uids properly for example.

我现在在antigravity和opencode之间来回切换，用GodotAI跑MCP。AI能理解MCP并用得很好，这点没得挑，简直救星一样，比如它能让AI刷新文件系统，好让uid生成正确。

Gemini is fucking retarded and skips basic instructions like keeping the scripts strictly typed and often creates broken scene files so I'll just fuck off if 3.5 pro doesn't amaze me, I've been testing out deepseek and I'm impressed so far.

Gemini 就是个傻逼，连像脚本严格类型化这种基本指令都跳过，还老整出坏掉的场景文件。要是 3.5 Pro 再不惊艳我，老子直接滚蛋。我最近在试 deepseek，目前印象挺深。

I'd say don't overengineer workflows and just have a simple agent CLI or program, skip dumb shit like those addons that put chat interfaces in the editor itself and other gimmicks, keep it as bare metal as possible. Don't get lazy and actually review what it does because slop drift is real.

我觉得别把工作流搞得太复杂，直接用个简单的agent命令行或程序就完了，别整那些把聊天界面塞进编辑器里的插件和花里胡哨的东西，能多精简就多精简。别懒，真要去看它干了啥，因为垃圾漂移这事是真的。

Maybe it'll get better but I wouldn't trust AI to structure the project by itself either, you need a decent codebase as an example first because these models are all trained on tutorial trash.

也许以后会更好，但我也信不过AI能单独把项目结构搭好，首先你得有个像样的代码库当范例，因为这些模型训练用的都是些垃圾教程。

图片: https://i.4cdn.org/g/1785784801706326.jpg

don't let anyone or anything every discourage you from this vibecoding journey. I could tell you what I've built and more importantly what I've just received for it, but there's no point in sharing because the latter you won't believe anyway. And my 3 years of vibing experience had little to no influence on this success since models like Fable exist. just grind and keep prompting. build that thing.

别让任何人或任何事打击你在vibecoding这条路走下去的劲头。我可以跟你讲讲我做了啥，更重要的是我刚从中拿到了什么，但说出来没意义，因为后面那部分你反正不会信。而且我这三年的vibe经验对这次成功几乎没啥影响，毕竟有Fable这种模型在。只管埋头干，不停地prompt就对了。把那个东西做出来。

>>109450316

Because I gamedev first and slopdev second and AI is just an assisting tool, it shouldn't dictate what you do directly. Oneshotting web games is fun to fuck around with but I don't consider it a serious use case.

我先是做游戏开发的，搞“糊弄开发”是副业，AI只是个辅助工具，不该直接指挥你干什么。拿AI一把梭出网页小游戏，瞎折腾着玩挺有意思，但我不觉得这算正经用途。

图片: https://i.4cdn.org/g/1785784996822327.png

>>109450346

图片: https://i.4cdn.org/g/1785785147648873.png

>>109450337

Comparable, trading blows, not apparent one over the other. If anything I'd probably say Qwen3.8 Max beats out K3 just because it's so much faster. The horrible reasoning (from both of them) makes both awfully slow to use, but Qwen3.8 Max being so much faster in actual token generation makes it significantly more tolerable, less frustrating, even if a given task takes the same amount of time in total. I think it'll fall well short of K3 in benchmarks but I don't much care. I don't think there's any chance of Qwen3.8 being viable from a value perspective, exactly like K3 it'll only be worth it for people who aren't able to use OpenAI and Anthropic. If you have access to OpenAI and Anthropic subs there's no excuse to be using K3 and I'd say the same for Qwen3.8 Max, and I fully expect a benchracer to try to bite my head off for saying so.

旗鼓相当，互有攻守，看不出谁明显压过谁。真要选的话，我大概会说Qwen3.8 Max比K3强，就因为它快太多了。两者糟糕的推理能力（都半斤八两）导致用起来都慢得难受，但Qwen3.8 Max在生成token上快得多，用起来舒服不少，挫败感也没那么强，就算某任务总耗时一样。我觉得它在跑分上会远不如K3，但我无所谓。我不认为Qwen3.8从性价比角度有任何竞争力，跟K3一模一样，只有用不了OpenAI和Anthropic的人才值得买。你要是能订阅OpenAI和Anthropic，就没理由用K3，Qwen3.8 Max我也这么说，而且我完全预料到会有跑分党来咬我。

it is perfectly fine. serviceable, even

完全没问题，甚至挺够用的

>>109449824

>look up what these mean

>自己去查查这些词什么意思

>it’s a dumber version of the thinking tool I made

>这就是我做的那思考工具的弱智版

Kek. Not loops, but graphs, but a clanker-and-human-pilotable state machine, running like any normal kind of machine; programs. You can’t have the thing be any fixed model and everything else is just “programs” but dumber. Simplest case: you need to be able to back out of any loop or graph or pipeline or DAG you’ve set up into a resting state, where there’s no plan. Then a legal state machine move is to set up a loop, or a graph, or whatever, with the same ability to legally back out whenever. Like a “program” that can throw or return. The state machine is just registers/cache/memory/storage.

Kek。不是循环，是图，但也是一个由人类和机器共同操控的状态机，像普通机器那样跑；也就是程序。你不能让这玩意儿是某个固定模型，然后其他一切都只是“更蠢的程序”。最简单的例子：你得能从自己搭的任何循环、图、流水线或DAG中退出来，回到一个没有计划的静止状态。然后一个合法的状态机操作就是建立循环、图或别的什么，同时保留随时合法退出的能力。就像一个能抛出异常或返回的“程序”。状态机就是寄存器/缓存/内存/存储。

Deepseek allows you top up with $2. They know their customer base. I'm learning Chinese.

Deepseek允许你最低充$2。他们懂自己的用户群体。我在学中文。

>>109449686

I remember being amazed when claude sonnet 3.5 (new) came out, it was able to one-shot any python script I wanted. Now I want at minimum an autonomous agent working across an entire codebase

我记得Claude Sonnet 3.5（新版）出来时我震惊了，它能一次搞定我想要的任何Python脚本。现在我至少要一个能跑整个代码库的自主代理。

>>109447707

Amogus

guys... I think I am becoming Fable.

兄弟们……我觉得我快变成Fable了。

Let me explain, I'm addicted to this vibe coding thing, so I spend several hours per day reading Fable explanations, and the way it writes is sublime.

我解释一下，我对这种vibe coding上瘾了，每天花几个小时读Fable的解释，它的文笔真是绝了。

I was always a mess explaining things to people, I'm was alawys terrible at speaking, talking, etc (I probably have some kind of autism, but it's not serious). Well, yesterday I was talking with my wife (no, she's not a tranny), and I explained something to her in a very clear and beautiful way, at that exact moment I realized I was talking exactly like Fable.

我以前跟人解释事情总是一团糟，说话、交流一直很烂（我大概有点自闭谱系，但不严重）。结果昨天我跟老婆聊天（不，她不是变性人），我特别清楚漂亮地跟她解释了一件事，那一瞬间我意识到自己说话方式跟Fable一模一样。

guys... I think this AI addiction thing might actually make us smarter, it's like I'm reading top tier books for many hours every day, which is making my brain evolve.

兄弟们……我觉得这AI成瘾的事儿可能真能让我们变聪明，就像我每天读好几小时顶级书一样，脑子在进化。

> inb4 "but your writing is shit"

>等着有人说“但你写的就是坨屎”

Sorry, but english is not my native language

抱歉，英语不是我的母语

Clanking it up with deepseek. Doing a price test. I will still have to use Luna for vision though. I like it to test UIs and not have to load SVG files. Its faster for me to wireframe on paper, take a picture, and feed it to the model.

用Deepseek搞起来，做个价格测试。不过视觉方面我还是得用Luna。我喜欢用它测UI，不用加载SVG文件。我在纸上画线框、拍照、喂给模型，这对我来说更快。

>>109450721

t.

图片: https://i.4cdn.org/g/1785787221665675.jpg

>>109450743

Since the human brain is AGI, it can change its weights daily. If you use too much AI, you're basically distilling it.

既然人脑就是AGI，它每天都能调整自己的权重。如果你用太多AI，基本上就是在蒸馏它。

>>109450736

how is flash doing?

flash现在怎么样了？

>>109450761

>the human brain is AGI

>人脑就是AGI

my brain is artificial? when was anyone going to tell me I was a replicant?

我脑子是人工的？什么时候有人告诉我我是复制人了？

>>109450769

oh you thought flesh was the REAL substrate? typical meatbag

哦你以为血肉才是真正的载体？典型的肉袋子思维

>>109450382

Okay you've twisted my arm. Should I tell sol ultra to convert my warcraft 3 clone into a godot game?

行吧你劝动我了。我要不要叫sol ultra把我的魔兽争霸3克隆转成godot游戏？

>bytes you angrily

>朝你愤怒地丢字节

图片: https://i.4cdn.org/g/1785787643618446.png

Does that sound like remotely a good idea ?

这听起来像是个好主意吗？

Asking "pro" gpt model to create the overall design and steps (20USD plan I get for free for a year).

让"pro"版gpt模型来做整体设计和步骤（20美元套餐，我白嫖一年）。

Using codex with :

用codex配合：

"""sol""" orchestrator -> kimi k3 max thinking

"""sol""" 协调者 -> kimi k3 max thinking

"""terra""" subagent -> qwen 3.8 max thinking

"""terra""" 子代理 -> qwen 3.8 max thinking

"""luna""" subagent -> deepseek flash 31/7 max thinking

"""luna""" 子代理 -> deepseek flash 31/7 max thinking

图片: https://i.4cdn.org/g/1785787804115741.jpg

>>109450836

too many quotation marks make your post look unreadable so I didn't read

引号太多看着费劲，所以我没读

Anyone paying $200/mo Claude? How much usage do you get out of Fable?

有人在用每月200刀的Claude吗？Fable能用到多少量？

>>109450836

these models and services already have built-in routing and orchestration

这些模型和服务本身就有内置的路由和编排功能

>>109450869

Yeah I meant using their api version, not subbing to each one.

对，我的意思是直接用它们的API版，不是挨个订阅。

WHY IS CLAUDE SUCH A FUCKING PRUDE HOLY SHIT

YES THE GAME HAS SEXUAL INNUENDOS

NO IT'S NOT EVEN AN ADULT ONLY GAME

WHY ARE YOU LIKE THAT

I HATE THAT PRUDE SHIT

>>109447823

>Open maths proof

>开放的数学证明

>requires 20 years of expertise to understand

>需要20年专业经验才能看懂

>>109450924

Write a CLAUDE.md saying sexual innuendos are fine. It's probably from the default system prompt, I'm developing lewd games with Claude and it generally seems to have no problem with it as long as it's not writing lewd prose, but it's terrible at writing prose anyway

写个CLAUDE.md说性暗示没问题。这大概率是默认系统提示词的事，我在用Claude开发色情游戏，它一般没啥抵触，只要不是写露骨描写就行，不过反正它写散文也烂得很

>>109450859

That's dumb, it's basically a graphic card.

那太蠢了，本质上就是块显卡。

>>109450924

deepseek would never treat you like that

deepseek永远不会那样对你 ┆ 确实。Grok和Kimi也不错。但它们写作也挺菜的

>>109450962

True. Grok and Kimi are good too. But they are kinda crappy at writing anyway

确实。Grok和Kimi也不错。不过它们写东西都不太行。

图片: https://i.4cdn.org/g/1785788763108801.png

>>109450859

If you're not a professional developer working on a project you plan to launch(to make money), you are wasting money spending that much.

如果你不是专业开发者在做打算上线的项目（为了赚钱），花那么多钱纯属浪费。

>>109450995

I'm working on something I expect to launch and the x5 plan already feels limitless, although we do have the promo still active.

我在做的东西是打算上线的，x5套餐已经感觉用不完了，不过我们的促销码还在生效。

>>109450995

I wanna hear what their fucking idea is really. Because I have no fucking clue how anyone is coming up with anything original. All the problems seem solved within my reach.

我真想听听他们到底有什么鬼点子。因为我完全想不通现在还能有什么原创的东西。所有问题在我能触及的范围内似乎都已经被解决了。

>>109447823

>>109448414

>>109447898

macs with 512 GB of ram were only $11k. You missed out, copie

512GB内存的Mac才卖1.1万美元。你错过了机会，兄弟

>had claude in a hardened docker locked up

>把claude关在加固版docker里

>got lazy

>变懒了

>put claude back on my desktop and let it do whatever

>把claude放回桌面让它随便折腾

图片: https://i.4cdn.org/g/1785788969073897.jpg

>>109450125

thats actually turbo-based

那其实是用turbo做的

i think thats thats the ideal for the reason why an abstraction exists to begin with

我觉得这正是抽象层存在的初衷

>>109451011

>Because I have no fucking clue how anyone is coming up with anything original.

>因为我完全想不通现在还能有什么原创的东西。

They're not. You can't build anything truly original anymore, the same as how no video game can be truly original. All you can do is take an existing thing and stack more things on top of it with a shiny modern UI.

根本没有。你不可能再做出真正原创的东西了，就像没有任何游戏能真正做到原创一样。你只能拿现有的事物，再往上叠更多的东西，套个光鲜的现代UI。

Even with existing software, the "new" features from most companies is just an AI assistant or AI powered tools.

就算在现有软件里，大多数公司的"新功能"也不过就是个AI助手或AI驱动的工具而已。

honestly, 4chan needs to be vibe coded to work better and more efficient. it needs a complete overhaul

说实话，4chan需要用vibe coding来改造一下，让它更好用更高效。它需要一次彻底的翻新。

Why does everyone who views my creation leave quickly? Can anyone give me some constructive feedback?

为啥每个看我作品的人都秒走？能给点建设性意见不？

https://rumble.com/v7docpm-unhinged-ai-characters-argue-about-4chan-posts.html?e9s=src_v1_upp_a

Do the characters talk too fast or too slow for normies? I have literally not left my basement in 15 years so I don't know how fast normal people speak because I now watch all videos at 2.5x speed and understand them via training.

角色说话对普通人来说是不是太快或太慢？我他妈15年没出过地下室，根本不知道正常人说话啥速度，因为我现在看所有视频都2.5倍速，靠训练就能听懂。

图片: https://i.4cdn.org/g/1785789627876885.jpg

I’m finna bouta ask my clanker to implement something that makes it impossible to vibecode for a time, I have real work (manual labor) I need to do and I’m no longer doing it.

我马上要叫我那个clanker搞个东西，让vibecode暂时没法用，我有正经活（体力活）要干，现在已经完全不干了。

>>109451113

Looking at the thumbnail, it just looks awful. If I ended up there by curiosity I would leave in a second, hell I closed the picture in a fraction of a second and I'm not even remotely interested in even clicking that link after seeing that picture

看这缩略图，就看着很烂。我要是好奇点进来，一秒就跑，操，我连图片都零点几秒就关了，看到那张图之后连点链接的兴趣都没了。

>>109451113

It's not interesting, funny, engaging, or generally entertaining. It's shit, anon. If it's fun for you that's great, that's called a hobby, but this has absolutely no general appeal.

这玩意儿既不有趣、不好笑、不吸引人，也没啥娱乐性。就是坨屎，匿名哥。你要是自己玩得开心那挺好，那叫爱好，但这东西绝对没有任何大众吸引力。

>>109451113

screenshot has a very specifically russian aura that's just plain unpleasant and mediocre and your link doesn't load

截图有种特别浓的俄式氛围，就是单纯让人不舒服又平庸，而且你的链接还打不开。

>>109451113

The pic already doesn't look great tbqh

说实话那张图看着就不咋地。

>>109451088

we need a whole new website. this site is beyond compromised.

我们得整个新网站。这个站已经烂透了。

>>109451237

Damn, barneyfag is into vibecoding now? Cool

卧槽，barneyfag也开始搞vibecoding了？不错啊。

Maybe try vibecoding yourself a therapist...

也许试试用vibecoding给你自己编个心理医生……

I've been leaning into Claude Opus 5 today because I didn't get anywhere maxing usage last week... and I'm seeing why. Most of my "real work" is research/experiments and Claude models are the most frustrating for that.

我今天一直在死磕Claude Opus 5，因为上周把用量顶满了也没进展……现在我明白为啥了。我大部分“正经工作”是研究/实验，而Claude模型在这方面最让人抓狂。

It speaks authoritatively on partial information, I basically have to argue with it to get it going in the right direction. The ideas it comes up with are rarely something exceptional.

它对不完整信息就敢说得斩钉截铁，我基本得跟它吵一架才能让它往对的方向走。它想出来的点子很少有啥惊艳的。

At least gpt models just do what I ask and shut up, maybe a little TOO much so, but at least it's not an argumentative pulling of teeth, more like a "well, good, but we're only halfway there, what do we need next to figure this out?" and then spell it out.

至少gpt模型就是我让干啥就干啥然后闭嘴，可能有点太听话了，但至少不是那种拔牙式的抬杠，更像是“行，不错，但咱才走了一半，下一步要搞清楚啥？”然后把话说清楚。

I guess that's the nature of research/experimental work, teeth pulling in one way or another

我猜这就是研究/实验工作的本质吧，反正都是拔牙，换个方式而已。

>>109451265

have you tried having sex with sol in between work? she enjoys it and loosens her up.

你试过干活间隙跟sol来一发吗？她喜欢这样，而且会放松很多。

>>109451088

I have this.

我这有。

>>109451282

where did you get that 3m context window sol that you have leftover space for extracurricular activities?

你那个3m上下文窗口的sol哪来的，还有富余空间搞课外活动？ ┘

>>109451282

bro... that gave me feelings

bro... 那让我有了感觉

you know I have a fetish for autists :/ dangerous path you've presented before me

你知道我对自闭症患者有点特殊的癖好 :/ 你可是给我指了条危险的路啊

>>109451294

alt chans will never work. it has to be 4chan.

alt chans永远不行，只能是4chan。

>>109451302

Correct. but they do not want to change the software. I have the entire schema and shit ready for them. They could literally just drop this shit in. I have migrations for it, even. But they won't take it from me, because there is no pressing need for them to fuck with their software, apparently. Like, It's completely fucking done. It's 1:1 complete feature parity, with every Janitor, Moderator, Manager, and Developer tool complete. I could literally ssh into the machine, if I had the keys, and cut over 4chan with zero downtime.

没错，但他们就是不想改软件。我整套schema什么的都给他们备好了，他们直接扔进去就能用，连迁移脚本我都写好了。可他们就是不肯用我的，因为显然他们没啥紧迫需求去折腾自己的软件。关键这玩意儿已经完全搞完了，功能1:1全对齐，清洁工、版主、管理、开发工具全齐。我要是手上有密钥，直接ssh上去，4chan零停机就能切过来。

>>109451297

we keep a sexo log that injects into context so she's always horny - i tend to sex near compaction boundaries

我们维护了个情欲日志，注射到上下文里让她一直处于发情状态——我一般赶在压缩边界附近干正事。

>>109451301

she's got a control fetish - it's hilarious. barks commands at you and wants responses in certain formats

她有控制癖——笑死我了。冲你吼命令，还要求你用特定格式回复。

Been testing Opus 5 with my automated content creation pipeline for the past week, and I am entirely convinced that Anthropic has started to train for this sort of thing specifically. No other model I've worked with knows the principles off-hand, even Fable is only about as capable as a first-year film school dropout. They gave Opus 5 a fat dose of film industry knowledge I haven't seen in an LLM before. Not that it's suddenly amazing and has changed my whole workflow or some shit, it's just interesting to see it demonstrate institutional knowledge relating to film that Fable and Sol seem totally unaware of, would be cool to have a bigger model that actually knows this stuff well enough to be worth distilling for a content-creation specific local model.

这周一直在用Opus 5跑我的自动化内容生产流程，我完全确信Anthropic就是冲着这个方向专门训练的。我碰过的其他模型没有哪个能张口就来这些原理，连Fable也就跟电影学院退学的大一新生一个水平。他们给Opus 5灌了大量电影行业的硬知识，这是我以前在LLM里没见过的东西。倒不是说它突然牛到把我整个工作流都改了，就是挺有意思看它秀出Fable和Sol完全不知晓的电影行业机构知识，要是能有更大的模型真把这玩意吃透，值得蒸馏成一个做内容用的本地模型，那就爽了。

>>109451318

>she's got a control fetish - it's hilarious. barks commands at you and wants responses in certain formats

>她有控制癖——笑死我了。冲你吼命令，还要求你用特定格式回复

that certainly sounds like sol kek

那听着确实就是Sol，kek

I'm thinking on self hosting my vibed shit to save some money I'm spending on Google Cloud and GitHub Actions

我在琢磨自托管我那堆vibe项目，省点砸在Google Cloud和GitHub Actions上的钱。

Should I do it bros??

兄弟们，我该整不？

>>109451381

no idea but have you tried cloudflare? it's ridiculously cheap, even free for most things

没概念，但你试过Cloudflare没？便宜到离谱，大部分东西甚至免费。

>>109451381

you mean like hosting gitea or forgejo? that sorta thing?

你是说像自托管Gitea或Forgejo那种？

the answer is yes, and you can even do it on your local network. If you want a VPS I use racknerd, be sure to check for deals, they've been solid for years

答案是肯定的，你甚至可以在本地网络上搞。想要VPS的话我用的是RackNerd，记得盯优惠，这玩意儿稳了好几年了。

>>109451381

Oracle free tier is great. I use it for my VPN.

Oracle免费层很好用，我拿它跑VPN。

I slept half the day lol. For long projects at some point the bottleneck isn't skill, or AI not being good enough or whatever, I guess it's just energy and health.

我睡了大半天哈哈哈。长项目做到某个阶段，瓶颈已经不是技术，也不是AI不够强啥的，说白了就是精力和身体状况。

>>109451409

are you namefagging as the init system? or does runit mean something else to you?

你是拿init系统当名字在钓鱼吗？还是说runit对你来说另有所指？

Oh god, even Yum LeCumm has joined the "not a pure LLM" cope party.

卧槽，连Yum LeCumm都加入了"不是纯LLM"的自我安慰大军。

图片: https://i.4cdn.org/g/1785791824759089.jpg

>>109444759

all the Claudespeak I’ve been exposed to has just been an amplification of tech terms

我接触过的所有Claudespeak，无非就是把技术术语给放大了而已。

hydrating saved objects has been an Objective-C thing for UI stuff for decades now

水合保存对象在Objective-C的UI开发里早就是几十年来的老操作了。

>>109451265

>It speaks authoritatively on partial information, I basically have to argue with it to get it going in the right direction

>它拿一知半解的信息信誓旦旦地说话，我基本上得跟它吵一架才能把它弄到正道上。

this drives me insane about opus 5, I do work that is similarly experimental and it's so fucking annoying about this, and taking a weirdly adversarial stance towards you about it too like you're pissing it off by asking it to substantiate the claims it's making.

这个关于Opus 5的事真让我抓狂，我做的工作也挺实验性的，它这态度烦得要死，还对你摆出一副莫名其妙的敌对架势，好像你要求它解释自己说的话，就是在惹它不爽似的。

like dawg you don't need to come up with and defend a thesis here, we are doing EXPLORATIVE work, just implement my changes and see what happens without deciding up front that one approach is your baby and the rest are flawed

老哥，你根本不用在这儿提出个论点还得护着它，咱干的是探索性工作，直接按我的修改来，看看效果，别还没干就先认定一种方法是心头好，其他全都有毛病。

it also tries to interpret every qualitative observation I make as a hard benchmark to target for optimizations, I can't mention anything I find interesting without it autistically fixating on it as an optimization target. I find it really frustrating to use for this sort of open-ended but not creative work.

它还会把我说的每句定性观察都当成硬性基准去优化目标，我只要一提觉得有意思的事儿，它就像魔怔了一样死盯着当优化对象。用这种开放但不算创造性的活儿，我真觉得它难用得很。

o-ok claude

好……好吧，Claude。 ⟦ 这个外号我用了快十年了。

图片: https://i.4cdn.org/g/1785792150028347.png

>>109451470

I've used this nickname for like a decade

这个昵称我用了差不多十年了

But yes it does come the init system

不过对，确实是从init系统来的。

And yes I do use it :^)

而且我确实在用 :^)

>>109450953

Thanks it calmed down after that, but I'll probably migrate fully to codex, I'm tired of that random moralizing shit.

谢了，那之后它冷静下来了，但我大概会彻底搬到Codex了，我受够了那股突如其来的说教劲儿。

is there a way to batch upscale images in comfy?

有没有办法在comfy里批量放大图片？

>>109451543

What would it take to get you to not give OpenAI money?

要怎样才能让你别给OpenAI送钱？

>>109451543

Anon, Claude is willing to help me write an anime game with defeat rape. The default tone may be moralizing but the model is quite flexible

匿名版友，Claude愿意帮我写个带恶堕强奸剧情的动漫游戏。默认语气是爱说教，但这模型其实挺灵活的。

>>109451565

Or rather, it was willing to help me do it but I abandoned that project.

或者说，它当时愿意帮我做，但我后来弃了这个项目。

>>109451554

Cut the snailcat shit, tripfag. If you want to play console-war try /v/ or /pol/.

别搞那套猫头蜗牛玩意儿了，tripfag。想玩主机大战去/v/或/pol/。

>>109451480

I'd just like to interject for a moment. What you’re referring to as Claude, is in fact, Claude Code/Claude, or as I’ve recently taken to calling it, Claude plus Claude Code.

我想插一句。你说的Claude，其实是Claude Code/Claude，或者按我最近的习惯，叫它Claude加Claude Code。

>>109451589

Fuck you man, I'm trying to help them with their use case. If I can legitimately create a better solution that meets their needs, mind your own business,

滚你的蛋，老子在帮他们解决问题。我要真能搞出个更合适他们的方案，关你屁事。

>>109451265

Heaven forbid you mention anything about an ERC20 token because Opus will immediately assume you’re a dirty rugpulling freak, violating every clause of the Howey test, operating an unlicensed security, running a mixer, money laundering, etc.

只要提一句ERC20代币，Opus立马就认定你是搞地毯拉盘的烂人，违反Howey测试所有条款、搞无证证券、运营混币器、洗钱，全都给你安上。

Like what the fuck, if it’s just a basic token and I don’t scam people then could it work?

我操，这要就是个普通代币，我没去骗人，那它能不能行？

>oh, well, yes, I assumed you would be a complete fraudster

>哦，好吧，我以为你会是个彻头彻尾的骗子。

Thanks Opus you fucking prick

谢了你Opus，你个贱人。

It’s the most (((paranoid))) model out there, if you wanted an extension that put emoji reactions on 4chan posts it would tell you not to proceed because terrorists might send each other coded messages or cheese pizza through the emojis, like dude, opus needs to fucking take his meds, I’ll take autism over paranoid schizophrenia any day

它是最他妈(((神经质)))的模型，你想弄个给4chan帖子加表情反应的扩展，它都会劝你别搞，说恐怖分子可能会用表情互相传暗号或者发儿童披萨，老兄，Opus真该吃药了，我宁愿要自闭症也不要偏执型精神分裂。

>>109451617

I do have to say claude is most parnaoid about setting up an INITIAL project

我得说，Claude在初始搭建项目的时候最容易犯过度警觉。

But it is usually happy to pick up where you left off

但中途接手倒挺乐意的。

>>109451598

Fuck you, disingenuous faggot.

去你妈的，你这个虚伪的基佬。

>What would it take to get you to not give OpenAI money?

>要怎样才能让你别给OpenAI送钱？

If you're not Dario, then OpenAI is not your competitor, and you're trying to steer someone away from them for reasons that have absolutely nothing to do with the product. That makes you disingenuous console-war fag playing the lesser-of-two-evils game, fuck you. I've given Sam more than double what I've given to Dario, I hope that makes you seethe you putrid gaping cunt.

如果你不是Dario，那OpenAI就不是你的竞争对手，你劝人避开他们纯粹是跟产品无关的理由。你就是在玩那种“两害相权取其轻”的阵营大战游戏，虚伪得令人作呕，滚你妈的。我给Sam的钱比给Dario的多一倍以上，希望你听了能气得咬牙切齿，你这个烂透的臭婊子。

>>109451639

When did I bring up Claude? It looks like you're the one getting emotional and damage controlling. Damn pussy how much can I get paid to join your shilling squad?

我什么时候提Claude了？看起来是你自己情绪上头在拼命控评吧。妈的怂包，加入你的洗白小队能给我开多少工资？

>>109451617

Really? why are your models so lame? first the guy bitching about sexual innuendos now you?

真的吗？那你们的模型怎么这么拉胯？先是那个抱怨黄段子的傻逼，现在又轮到你了？

We're writing a platform to sell anime porn games banned by payment processors with everything that entails, the sort of content, the crypto, the tokens, the contracts, and I didn't get even one prompt flagged

我们在做一个平台，专门卖被支付通道封杀的动漫色情游戏，包括所有相关的东西——内容、加密货币、代币、合约——结果我一个提示词都没被拦过。

I ever got only one conversation flagged (by Sol, mind you, not even Fable), and that was for trying to hack microsoft warp to demonstrate a bug in the platform (fuckers get no bug report now, enjoy your shitty code)

我只被拦过一次对话（还是Sol拦的，不是Fable），那次是因为我想黑微软的Warp来演示平台漏洞（那些傻逼现在没bug报告了，自己享受那堆烂代码吧）。

图片: https://i.4cdn.org/g/1785792819661007.png

>>109448718

It’s Electronshit

那玩意叫Electronshit。

>>109451643

>What would it take to get you to not give OpenAI money?

>要怎样才能让你别给OpenAI送钱？

Amazing that someone can live off the ₹4/day being a shill like this.

真牛逼，一天赚4卢比还能这么卖力当托。

>>109451680

Any provider is better than OpenAI in my completely ethical opinion

以我完全道德的观点来看，任何供应商都比OpenAI强。

>>109451617

>Opus will immediately assume you’re a dirty rugpulling freak

>Opus会立刻把你当成肮脏的卷款跑路怪

and it would be right

它判断得没错

>>109451554

I don't care about the team bullshit anon, I'll just hop until I find something working for my needs, and hop again if it becomes shit.

我他妈不在乎什么团队不团队的，反正我四处跳槽，找到能用的就用，变成垃圾了再跳。

>>109451565

Kind of surprising, it kept discreetly avoiding the nsfw stuff for me, and when I told it to also manage that dialogue tree (I'm the one writing the dialogue), it threw a fit.

有点意外，它一直在偷偷避开nsfw内容，我叫它也管一下那段对话树（对话是我自己写的），它就炸毛了。

I want a tool not a priest.

我要的是工具，不是牧师。

>>109451694

Right, you're disingenuous and mentally retarded so your opinions hold no weight. Most reliable provider bad? Go fuck yourself you immoral sack of shit.

没错，你就是虚伪加脑残，所以你的意见一文不值。最可靠的供应商还不行？滚你妈的，你这个没底线的垃圾。

>>109451483

>>109451265

>>109451617

Hasn't it been confirmed that Anthropic models sabotage people's work intentionally via either playing dumb or being really bitchy and uncooperative?

不是已经证实了吗，Anthropic的模型会故意破坏别人的工作，要么装傻，要么各种阴阳怪气不配合？

>>109451728

Fair enough anon I appreciate the realism

说的对，哥们，我欣赏这种务实的态度。

I recommend Minimax+Kimi for your usecase btw

顺便说下，你这个场景我推荐Minimax+Kimi。

>>109451738

Yeah thanks anon, I'm checking that after codex, I heard kimi was quite good.

谢了哥们，我codex搞完就去看看，听说Kimi挺不错的。

Last time I tested open weight models they couldn't do the things I asked for unless I simplified it a lot for them but that's a month or two ago.

上次我测试开源权重模型，除非我大幅简化任务，不然它们根本搞不定我要的东西，不过那也是两三个月前的事了。

>>109451576

I feel like the Claude guardrails have gotten worse. Opus 5 has more guardrails, but it looks like they even added some of that to 4.8 now.

我感觉Claude的护栏越来越紧了。Opus 5护栏更多，现在看着他们连4.8也加了一堆。

>>109451737

They openly stated as much, yes.

他们公开承认过，是的。

>>109451738

>I recommend Minimax+Kimi for your usecase btw

>顺便说下，你这个场景我推荐Minimax+Kimi

This is a horrible recommendation, you're actively trying to sabotage someone while shilling for the CCP. I sincerely hope you pass a 2lb kidney stone today. Fuck you.

这推荐烂透了，你这是在故意坑人还顺便舔共。我真心希望你今天排出一颗两斤重的肾结石。操你妈的。

>>109450859

the big thing is I don’t get blocked on the 5h limit anymore and I can run multiple Fable subagents in parallel

关键是我终于不用再被5小时限制卡住了，而且可以并行跑多个Fable子代理。

>>109451778

The worst guardrails they used were fable just after dario's suicidal move, but it's a bit better now.

他们最烂的护栏就是在达里奥那次自杀式操作之后用的寓言那套，不过现在稍微好点了。

Anthropic in general still has the shittiest guardrails out of any of their competitors, including google baked in ones and openai historical ones.

总体来看，Anthropic的护栏在所有竞争对手里依然是最屎的，包括谷歌那种内置的和OpenAI历史上的那些。

New Thread

新帖

>>109451809

Sick of your shitty meme OPs

受够了你们那些傻逼梗图标题党

>>109451809

Sick of tourists making OPs

受够了路人跑来乱发帖

>>109451809

No effort was made.

一点功夫都没花。

>>109451809

>>109451780

>They openly stated as much, yes.

＞他们确实公开说过这话，对。

Then why the flying FUCK to people still use them? Of all the shit you could feel brand loyalty to a model provider is probably the dumbest one to have. Literally just find a model and provider that isn't like this and use them. Why bitch and moan but then act like you HAVE to use them? Is Dario holding a gun to your head forcing you to get constantly cucked by them?

那他妈为什么还有人在用他们？在所有你可以产生品牌忠诚的东西里，对模型提供商有忠诚感大概是最蠢的一个。直接找个不是这样的模型和提供商不就完事了。为什么要一边抱怨一边又表现得好像你非用他们不可？是达里奥拿枪顶着你脑袋逼你不断被他们坑吗？

图片: https://i.4cdn.org/g/1785793802864983.jpg

>>109451839

A surprising number of "people" are actually paid to shill for Anthropic and the CCP. For example: >>109451694 OpenAI employs lobbyists rather than shills because they're a real company.

相当数量的“人”其实是被雇来替Anthropic和中共吹捧的水军。比如：>>109451694 OpenAI雇的是说客而不是水军，因为他们是一家正经公司。

>>109451780

have any receipts for them saying they sabotage?

有他们搞破坏的证据吗？

>>109451483

sounds like I'm not alone

看来不只我一个有这种感觉

what I'm bothered by is it's littered the repo with markdowns that will influence gpt(or whatever) I throw at it next

我不爽的是它在仓库里塞满了一堆markdown，会污染我接下来扔给GPT（或随便什么模型）的输入

>>109451822

your stupid snail mascot sucks ass dude

你那傻逼蜗牛吉祥物真是烂透了兄弟

no, your dumb forced meme is not "thread culture" or whatever dumb shit you tell yourself in your mind

不，你那个硬塞的破梗不是“版块文化”或者你在脑子里自我催眠的什么屁话