TGViewer
Channel Public Channel
V2EX

V2EX

@pushv2ex

推送 V2EX 论坛贴【全部板块】

本频道: @pushv2ex

友频道: @push52pj
Subscribers
548
Photos
24.2K
Videos
3
Links
227K

Showing posts older than #236699 · Back to latest

Older Posts 20 shown
Post #236698 8
[问与答] 突然有个想法,自动化助手股票下单可行性分析;

最近一直再用 AI 鼓捣自动化助手,自动化操作 windows 应用的场景,目前在游戏场景,已经得到应用,目前我试玩的传奇私服的游戏,已经可以完全自动化了; 突然想到,为啥不做一个自动助手,去股票下单;比如解析行情,跑策略,然后操作 windows 版本的应用下单,这样也能自动化; 不知道小弟这个想法,可有可执行性;
Post #236697 7
V2EX [推广] What actually changed in Kling 4.0? I compared the official 3.0 and 4.0 docs I've been going through Kling 4.0 material lately, and a lot of write-ups credit 4.0 with things 3.0 already had. The usual three are 4K, native audio and multi-shot. The official…
Here's what the official material says. Flash has been in limited early access since Sept 28, and the official entry point's notice points to Ultra annual subscribers. For the full model, the release notes say it launches in October, while the Sept 30 introduction says it enters early access first and expands in October. I couldn't find evidence that every regular account has full 4.0 access, but it's also wrong to say nobody can use it. It depends on your account.

Two items are explicitly still coming soon: 10-bit HDR at 1080p/4K, and multiple extensions up to 2 minutes.

5. Who should care

By task, 4.0 is most relevant if you:

● need longer continuous shots, or short stories and previs with a clear arc
● need complex character motion or performance clips
● need several references at once to keep a character or product consistent
● want keyframes to control timing
● make ads, music videos or performance shots
● have multilingual dialogue where voice and lip-sync need to line up

If you mostly make simple 5–10s clips, basic image-to-video or quick idea tests, 3.0 is still fine. There's no need to upgrade just for the version number, and 4.0 coming out doesn't make 3.0 useless.

References:

The information above about Kling 4.0 comes from kling4.tv.

● Kling 4.0 official release notes
● Kling VIDEO 3.0 / 3.0 Omni official guides
● Kling 4.0 feature page and multi-keyframe guide

A question for anyone who already has access to 4.0 full or Flash: on 20–30s continuous action, do the characters and motion stay consistent all the way through? And once you go past five or six keyframes, does it get more controllable or do they start fighting each other? Official specs are one thing; what actually comes out is another.
Post #236696 7
[推广] What actually changed in Kling 4.0? I compared the official 3.0 and 4.0 docs

I've been going through Kling 4.0 material lately, and a lot of write-ups credit 4.0 with things 3.0 already had. The usual three are 4K, native audio and multi-shot. The official docs show all three were there in the 3.0 era: native audio and Multi-Shot are in the 3.0 launch guide, and a 4K mode was added to the 3.0 series this April.

So I lined up the official 3.0 and 4.0 docs side by side. As far as I can tell, the real changes in the full 4.0 model come down to a few things: the single-generation cap goes from 15s to 30s, up to 10 keyframes, up to 15 combined reference assets, plus what Kling describes as better handling of complex motion and big camera moves, and broader audio/language support.

The short version first ("3.0" means the 3.0 series; the reference and editing parts mostly refer to 3.0 Omni):

1. What 30 seconds is actually good for

Going from 15s to 30s sounds like a number doubling, but for anyone making clips the real difference is whether one complete action fits in a single generation. A lot of shots have a natural arc: the build-up, the action itself, someone's reaction, and the moment it settles. Squeezing all four into 15s is tight, so people split it into several generations and stitch them, and continuity at the seams becomes its own problem. A 30s window leaves more room to plan, so continuous product demos, a small single-scene story, a performance clip or one long camera move have a better chance of fitting in one pass.

A few things are easy to overstate, though. 30s is a cap, so it doesn't mean every clip should be 30s. It doesn't mean you can skip editing, and it doesn't automatically mean fewer retries or lower cost. The official material has no numbers on that and I haven't run comparison tests, so the honest claim is just "more room to plan a shot."

Also, the "multiple extensions up to 2 minutes" Kling mentions is a separate feature, currently marked coming soon. It is not a native 2-minute single generation.

2. 10 keyframes vs 15 references: what each one controls

This part confused me at first too. Put simply, keyframes control what the frame should look like at a given point on the timeline. References control what the character, product, motion, style or voice should be like.

Two common misreadings. First, 10 keyframes is not 10 shots. Keyframes are anchor points in time, shot count is a different thing, and 3.0 already did multi-shot. Second, 15 references is the combined total, so you can't add up the per-category caps and get twenty-something.

Kling's keyframe guide also warns that references and keyframes shouldn't contradict each other, and that the numbers are limits. Using all of them doesn't make the result better.

3. Don't mix up Flash and the full model

Plenty of write-ups blend the full model's specs with Flash's, so they read like one model. Per the official feature page:

● Full: 3–30s, 720p/1080p/4K, first/last frame and up to 10 keyframes, the more complete set of controls.
● Flash: 3–20s, 720p, 8-bit SDR. It supports text-to-video, first-frame image-to-video and Omni Reference, but the feature page lists first/last frame and multi-keyframe as not supported. It's positioned for fast iteration.

So when you see "Kling 4.0 does 30s, 4K and 10 keyframes," those are full-model specs and don't carry over to Flash. Going the other way, Flash being 720p doesn't mean the full model is 720p only.

4. Can you actually use it right now?

"Officially announced," "Early Access," "Coming Soon," "open to all regular accounts" and "available on third-party platforms" are five different things.
Post #236695 4
[推广] 做了个宽高比计算器 根据比例快速算出另一边应该是多少像素

剪视频、出图时常算另一边应该是多少像素,自己算有时候其实挺烧脑的,于是做了这个站。

核心是:已知宽高求比例,或已知比例和一边求另一边;另有转换(裁切 / 补边)、分辨率、DPI 、码率、屏幕尺寸、PPI 、图片尺寸检测等。社交平台页挂官方帮助出处,公开数据集可下 JSON / CSV 。

采用的静态站方案,计算在浏览器本地完成。

欢迎试用和拍:Aspect Ratio Calculator
Post #236692 4
[Local LLM] strata 极致优化指南(参数篇)64GB RAM + RTX pro 5000 48GB, IQ3_XXS 量化, prefill 4500t/s, decode 180t/s 同时缓存 5 个 262k 会话不重新 prefill

前言:strata 最近已经火出圈了,火到 RAM 又翻了个倍。目前京东自营,海力士颗粒的 64GB 的 DDR5 6000 内存已经几近破万。

十一长假回来,笔者也是暂时小试牛刀了一把,发现有一些关键参数,会直接影响整个 LLM 引擎的本质体验,但是默认值并不是特别合理。所以发个 tips 小文供大家交流。欢迎批评指正。

本文全程手打,无任何 AI 含量,请放心阅读。

----------------------

1 、Harness 会频繁唤起子代理( sub agent ),导致切换会话 prefill 浪费太多时间,怎么办?

答:开启--conversation-cache-mib

这个参数会让 strata 使用 RAM 来缓存历史对话,从而不让 LLM 反复重新 prefill 老的 session 。增加会话切换时的速度,对于 agent 、harness 场景有奇效。

开启 conversation-cache-mib 后,monitor 面板会有如下展示:

分别表述当期缓存了多少个会话,以及目前占用了多少 RAM 。

这个机制有两个痛点:1 、不支持多显卡。2 、最大只支持 20GB 的对话缓存(如果是 8bitkvcache ,大概就是 4 各半的 262k context )。

----------------------

2 、我的显卡显存总是占不满,浪费一些,怎么破?

答:配置--expert-cache ,不要用 auto 档位。手动指定

让更多的专家进驻显存。注意,手动指定的是下限,所以要反复测试量力而行。

----------------------

3 、我想使用上文的 conversation-cache ,但是我 RAM 都快满了,怎么办?

答:启用低内存模式:--resident-experts

小科普,正常默认情况下,strata 加载的专家是内存+SSD 加载全量,vRAM 加载高命中概率专家。相当于 RAM 和 vRAM 之间,是有一定的重复加载的。

官方这样指定默认值是有道理的,一旦 VRAM 的专家没能命中,直接可以读取 RAM 的进入 vRAM 。 代价就是这会导致 RAM 的极大浪费。

通过配置低内存模式,能够让 vRAM+RAM 分别加载不同专家,分工合作。这样能节省巨量 RAM 。代价是一旦 VRAM 中的专家没能命中,除了从 RAM 读取以外,还要从 vRAM 剔除一个回写到 RAM 。但相较于直接节省 40GB 的 RAM 空间而言,这点回写简直可以忽略不计。而且你显卡的显存越大,遇到回写的概率越低。 笔者的是 48GB 的 rtx pro 5000 ,专家命中率一直保持在 99%。

----------------------

4 、我想多个聊天会话一起聊(并发),怎么办? 答,配置 parallel: X ( X 是并发会话数,最大 8 )

会占用显存,我个人实测,IQ3_XXS 量化下,会话长度 262k ,每个并发占用 5.2GB 显存。

以上,欢迎列为交流,指正。最后列出我个人的全套参数,供各位参考。
"args": [
"--pack", "C:\\work\\Strata-data\\packs\\iq3_xxs",
"--native", "C:\\work\\Strata-data\\models\\IQ3_XXS\\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf",
"--ple-gguf", "C:\\work\\Strata-data\\models\\IQ3_XXS\\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf",
"--expert-profile", "C:\\work\\Strata\\data\\expert-profile.bin",
"--resident-experts",
"--expert-cache", "19000",
"--prefill", "auto",
"--spec", "4",
"--mtp", "C:\\work\\Strata-data\\mtp\\rt",
"--max-context", "262144",
"--kv", "k8v4",
"--vision",
"--vram-reserve-mib", "3400",
"--pcie-frac", "0.00",
"--spec-min-p", "0.70",
"--conversation-cache-mib", "20480",
"--conversation-cache-slots", "6",
"--conversation-cache-min-free-mib", "13824",
"--prompt-cache-every", "8192"
],
Post #236691 5
[程序员] 前 AI 时代挖的坑, AI 时代填上了:一个算法可视化网站

这个项目其实是前 AI 时代就开始做的。起因是我自己学算法,总是学完就忘,就想着做一个算法可视化项目,既能顺便学习,又能刷一点开源项目。那一阵子新拟物风格比较火,正好拿来练练手。

结果做完第一个算法(冒泡)就搁置了,工作太忙,一放就是好几年。后来有了 AI 才重新捡起来,今年补到了 92 个条目,也算是圆梦了吧。

现在大概是这样:

● 9 大类:数据结构、排序、图、DP 、回溯、字符串、数论、计算几何、查找
● 每一步都能前进、后退、拖进度条,倍速可调
● TS / Python / Go / Rust 四种代码跟着动画逐行高亮,还有变量面板
● 排序可以输入自己的数组,二分和快排会在关键步停下来出题
● 中英双语,手机也能看

地址: https://algo.illegalscreed.cn/

源码: https://github.com/IllegalCreed/algorithms-visualization

说实话,目前的效果还远远没达到我想要的样子,后面有机会再慢慢优化。

最讽刺的是,都用 AI 做了,反而背离了我的初衷:做完这个网站,我还是没学会算法。不过转念一想,大 AI 时代,学算法还有个屁用呀……这么一想,这个网站好像也没啥用了 😂

还是那句话,大 AI 时代,学算法还有个屁用呀。欢迎来反驳。
Post #236689 4
[汽车] 紧急情况,不应该一脚踩死么,还要点点刹?

以前遇到老司机总说我踩刹车,老是一脚踩死,这个习惯不好。应该踩一脚松一下再踩死,这样更安全

今天网上又看到老司机争论说:高速踩刹车要点点再踩死,说这样既不会踩断刹车,也不打滑。

能理解多次点刹肯定更不易损坏刹车,也更安全,问题是遇到紧急情况本能就一脚踩到底啊,再松一下不是又往前多冲出距离了么。

再说,高速情况,踩>松>踩 那得多块的脚速啊,这™是老司机炫技,还是挖坑呢,有老铁说说真的能做到吗?
Post #236688 5
[分享创造] 因朋友是地理老师,开发了这款「中国昼夜地图」欢迎大佬试用并且提意见

可以随时查看中国以及各省,任何时间的昼夜分布图,以及经典的夏至、冬至对比分布图

比如夏至 黑龙江 3 点多天就亮了,而夏至的新疆 晚上将近 10 点天才黑了

都可以清晰看出,欢迎大佬试用并且提出意见

给朋友上课用的不知道 OK 不 OK

地址: https://www.gdufe888.top/sun/

源码: https://github.com/wenbochang888/china-daylight-atlas
Post #236687 5
[分享创造] 这绝对是 Window 最炸裂的终端效果了!

项目出发点其实是眼馋 mac 的各种优秀的终端,分屏,终端持久化,消息通知,命令补齐等等,由于 win 上实在没有太全面的终端;于是乎我又来 vibing 啦 本次更新重大,加入了非常多的内容

项目介绍

PC 端

欢迎各位朋友建议+提 issue ui 方面希望给出更多的建议,争取赶上 mac 的体验!

注: 使用前一定要先安装压缩包内的 maple font ,不然会显示异常 喜欢的朋友可以点个 star 哦

项目地址: https://github.com/Kuddev/pebrel
Post #236686 4
[投资] 有没有必要配置一枚 BTC

从资产配置的角度来说,大家觉得有没有必要配置一枚完整的 BTC 并长持。我说的长持是传家的那种,就是不遇到不卖出它就过不去的坎,就终生不卖了,将来挂了就留给子孙后代,当传家宝。

另外想听听大家对 BTC 未来价格走势的看法,觉得极值是多少,能达到或突破 100w 美元吗?
Post #236685 6
[分享创造] 做了个网页版练枪靶场,有无 FPS 大佬来挑战一下

最近沉迷 FPS 游戏,我想每天玩游戏前练练枪,又不想为了这几分钟装个客户端,就搓了个网页版的:

https://aimrange.org

浏览器打开就能玩,不用下载,也不用注册。

● 6 个常见场景:Gridshot 、Sixshot 、Microshot 、Spidershot 、Motionshot 、Strafetrack
● 灵敏度默认按瓦的算法,瓦学弟直接填游戏里的数就行。go 学长切到 CS2 填自己的数,几个游戏之间也能互相换算
● 准星可以直接粘贴瓦的准星代码导入,不用一项项调
● 有周榜和总榜。没有账号体系,起个名字就能上榜
● 顺手增加了反应时间测试等小工具

欢迎大佬们体验和批评,有 bug 、手感不对,或者灵敏度换算对不上的都欢迎评论
Post #236683 2
[职场话题] 最终公司还是要挺不住了😢

99 的今年 27 岁,21 年出来工作现在五六年了,一直干前端

第一家是一个可视化厂商,用自己的框架做三维大屏

现在第二家是做水利系统,主要是 cesium 地图+点线面的绘制,技术要求也不高,今年 ai 基本也可以搞定了。

近两年公司接不到什么项目,也没年终奖了,只有每个月的固定工资。

今天领导挨个办公室通知全员降薪 20%,公司的现金流不健康,大饼是是等现金流好起来会重新涨回来,争取再涨薪。

目前我个税在 12k ,实际到手刚好 10k ,降 20%下去估计到手就 7 、8k 了,和我 21 年刚毕业一个水平,可能还不如刚毕业的时候,毕竟当时五险一金最低缴纳到手会多一点,而且当年还能收到一两个月的年终奖😢

现在纠结两个点
1 、如果有裁员名额是不是直接争取裁员好一些
2 、如果没有裁员/争取不到裁员,在一定会降薪的情况下,合同固定工资为 8k ,其他的收入有 3k 左右的绩效,500 左右的补贴,300 的工龄工资,这些要扣是不是反对无效,只能呆到自己离职。

自己和 ai 聊天分析了半天感觉就两个选项,要么能争取到裁员走人,要么只能接受降薪呆着。
不知道我上面的想法有没有问题。
Older posts →
Threads Profile ViewerView any public Threads profile without an account.Open ThreadLook →Writing with AI? Make it sound human.Metric37 rewrites AI drafts so they read naturally. Free AI detector, 1,500 words free.Try Metric37 →