scriptsboy🇨🇳🇭🇰🇲🇴🇺🇳 @scriptsboy

@Nicole_yang88 美版感觉很难买到

VirtualBox 7.2.18 Released with Linux 7.3 Fixes, Support for RHEL 10.3 Kernel lxer.com/module/newswir…

@realNyarime 看来顺路得去吃几个广西人，顺路在把卡办了

@gwvpsceping 速度如何？

License update: Starryblu’s Canadian licence has been updated with additional permissions. fintrac-canafe.canada.ca/msb-esm/reg-eng

勿忘，九一八！ 1931年9月18日 九一八事变爆发 这个日子 是中国人民抗日战争的起点 更是所有中国人心中的痛 勿忘国耻、铭记历史 砥砺前进、吾辈自强

@realNyarime 正在找这个卡

😩 Still waiting for your Starryblu Card? We know the wait can feel LOOOOONG…⏳ So here’s a little something for those who’ve been waiting. 💙 🎁 We’re giving 10 people the chance to unlock Starryblu Card access sooner — no more waiting. No invite code? That’s okay. Used someone else’s invite code? Also okay. Want in? 1️⃣ Follow @Starryblu_ 2️⃣ Quote this post 3️⃣ Tell us in the comments: What would you use your Starryblu Card for? 📅 The 10 selected spots will be announced on Friday, September 18. 🍀 Only 10 spots. Maybe this little Starryblu perk has your name on it. 👀 #Starryblu #Fintech #GlobalPayments #StarrybluCard

@realNyarime 明天买一张去

@pengchujin 感谢大佬

Brilliant paper by NVIDIA. they found a way to make KV cache transferable between models. the target model skips prefill entirely, and the conversion runs 2.7 to 25x faster than processing the context again. let's understand why this is so important today. LLM APIs are stateless, so every turn sends the entire conversation back to the model. the model reads it again before generating a new token, and all of it is billed as input. prompt caching lets providers hold the KV cache for a stable prefix and bill cache hits at roughly 10% of the base input rate because that compute was already done. this 90% reduction is one of the biggest levers in LLM serving. but the cache only works on the model that produced it. since keys and values depend on that model's weights, another model cannot directly reuse them. this becomes a problem with LLM routing. if traffic moves to another model for cost or capability reasons, the accumulated KV cache becomes invalid. the entire context must be processed again and billed at full rate. NVIDIA's recent paper treats this as a representation problem. prefill's main output is the KV cache, so transferring it means converting one model's cache into the representation another model expects. they first checked whether this conversion had enough structure to exploit. for Qwen3 14B → 32B, a linear regression from a single source layer reconstructed 56% of the variance in the target model's keys. because models can have different layer counts, there is no natural one-to-one layer mapping. so, for every target layer, they rank the source layers by predictive power and use the top eight together. that pushes reconstruction to 79%. the mapper has three parts: > each target layer and head gets its own linear map, solved in closed form instead of with gradient descent. > cross-layer selection chooses the source layers that best predict each target layer. their ablation shows this contributes the most. > RoPE adds a position-dependent rotation to keys. they remove it, fit the mapping in position-free space, then apply the target model's rotation again at inference. across six model pairs from Qwen3, Llama 3.1, and Ministral 3, four retain 73 to 98% of the receiving model's standalone accuracy. conversion is also 3 to 25x faster than processing the context again. prior cross-model KV reuse methods either train a neural adapter for every pair or require architecturally identical models. this approach is closed-form and training-free, but there are still important limitations. every tested pair belongs to the same model family, such as Qwen → Qwen or Llama → Llama. cross-family transfer is still future work. the tested pairs also share KV head count and per-head dimension. mismatched configurations remain untested. and the work currently covers dense full-attention models only, not sliding-window or attention-recurrent hybrids. link to the paper: arxiv.org/abs/2608.03893 i wrote a first-principles breakdown of how KV caching works the article is quoted below.

Akshay 🚀 @akshay_pachaar

8 months ago

x.com/i/article/2015…

21 182 1K 645K 2K

今天看到的最漂亮的网站： kubernetes3d.com/rack 把 k8s 集群做成了机架的样式，而且高度可操作，真的是 eye candy，还分了 front back 和 side 三个面，每个面讲的是同一套集群的不同切面，是一件精致的艺术品。 注意上面的各种按钮和场景都是可交互的，用来学习概念也很棒，我已经把玩许久了，谁做的这么棒啊。

@hejushang8 土耳其的有必要搞一个

This guy built a trading bot on Polymarket and made +$295,910 His bot trades 5 and 15 minute crypto Up/Down markets using a late-resolution strategy, where most of the position size is deployed when the outcome is getting close to final: 1. It starts with a small directional position At the beginning of each market, it records the reference price and tracks Chainlink TWAP along with external spot/perp feeds During the early stage, it builds a small position. This acts like a probe while the model’s confidence is still relatively low 2. It increases the position as the outcome gets closer to certainty The algorithm constantly recalculates probability. As resolution gets closer, it can estimate more accurately whether the result can still change 3. If the signal changes, it buys the opposite side In 88.2% of markets, it trades only one outcome. In 11.8% of cases, it buys both Up and Down. The opposite side works as a hedge His Polymarket nickname: BoneOhio Follow this account and trade on Polymarket through this app: t.me/PolyGunSniperB… The main edge of this guy’s HFT algorithm appears during the final minutes of the market. It estimates how close the outcome is to being locked in and concentrates most of its position size where the model sees mispricing

VALIX @RetroValix

a month ago

x.com/i/article/2091…

12 36 180 1.9M 511

some seriously good AI models are sitting at $0 right now 😳 KiosAPI is showing free access to: - MiniMax M3 - Kimi K3 - GLM 5.3 Flash - DeepSeek V4 Flash getting started: 1. go to kiosapi.com 2. create your account 3. generate your API key 4. base URL: kiosapi.com/v1 5. select a $0 model > OpenAI-compatible API worth checking 👀

Auto-generate Web Recon Reports - EyeWitness Tool

Cola Skill 收藏了超多实用的 Agent Skill 全部是社区和创作者们精心挑选 每天都有新的 skill 推送 找不到想要的 skill 的时候 就可以来搜搜看 colaskill.com

@TechElyra Send

卧槽！微软官方亲自下场做AI 量化研究员了——RD-Agent，14.4K stars。 以前搞量化，挖因子、调模型、写代码、跑回测，基本都得人一轮轮手动做。 RD-Agent 直接把这套研究流程交给 Agent： 提假设 → 写代码 → 跑实验 → 看结果 → 再迭代 让因子和模型一轮轮自动迭代。 核心亮点： 1️⃣ 因子 × 模型协同进化：不是只挖因子或只调模型，而是两边交替优化 2️⃣ 自动挖 Alpha：能从研报等文本中提取因子，再生成可运行代码 3️⃣ 背靠 Qlib：因子、模型、回测、信号整套量化流程直接串起来 4️⃣ 实验结果不错：论文量化场景中，用更少因子跑出了约 2× 基准 ARR 5️⃣ 直接能跑：pip install rdagent，官方提供量化场景入口 我比较在意的不是那个 2× ARR。 而是它把 AI 从“帮你分析股票”，变成了一个会提假设、写代码、跑实验、看结果，再继续改的量化研究员。 这才是 RD-Agent 真正有意思的地方。 github.com/microsoft/RD-A…

爱吃折耳根的Ace @_zheergen

4 weeks ago

挖到一个 A 股开源选股工具：a-share-quant-selector 做技术选股最烦的，往往不是没有想法，而是每天对着全市场 K 线，自己判断碗口、放量和 KDJ。 它的核心很直接：用 AKShare 拉 A 股数据，按碗口反弹规则扫描，结果推到钉钉，也能开 Web 看图。 ✅ 碗口反弹：上升趋势 + 放量阳线 + KDJ 低位 ✅

27 25 87 38K 114

APRENDE A HACKEAR!!! Hay un catálogo de máquinas vulnerables, montadas a propósito, donde puedes practicar intrusión sin cometer ningún delito. Laboratorios gratis, entorno real. Y cuando te atascas, existen las soluciones escritas paso a paso.