Rosler H

[Rosler H @StarAlone999]

@kyou_huang 神经病吧，大陆安全得很啊。

[Rosler H @StarAlone999]

@SKgic9h7u31496 @AntiTHBK 你不是支那人吗？当初日本人的这个定义不针对台湾人？

[Rosler H @StarAlone999]

@Qisiyingshi3 @happypaidaxing 你能不能先了解一下这个支那人的词的由来？你这不是也骂自己了吗

[Rosler H @StarAlone999]

@62qoe 一顿饱和顿顿饱人家还是知道的哈

[Rosler H @StarAlone999]

@nyxvoid92 傻币吧，它写在自己报告里中国是敌对国家不让用了

[Hux @huxlab]

认识一个在 Anthropic 做研发的朋友，年薪 150 万刀。

我问她是怎么把 Claude 调教到极致的

她直接甩给我一段 30 分钟的内部录屏，原本不对外公开。

毫不夸张，这绝对是全网含金量最高的 Claude 实操资料之一。

昨晚看到一半，我后背发凉：

过去两年我写提示词的方式，从底层逻辑上就全错了。

看完后，我提炼出 5 条Anthropic 内部正在用的核心法则：

1. 用 XML 严格隔离语义块：如果你肉眼都分不清数据、业务政策与角色边界，模型只会更混乱，严禁把所有规则揉成一坨大段文本。

2. 定期清理旧模型的“防御性补丁”：新模型遵循能力极强，遗留的历史负向限制会导致过度拟合，反而把上下文已有的正确信息死扣不给。

3. 指令无法凭空赋予模型能力：在 Prompt 里加粗强调“务必算准”没有任何意义，复杂运算必须果断剥离给 Tool Calling 代码执行。

4. 业务权衡必须给全博弈天平的两端：只灌输单向操作成本会让模型极端避险，必须同时给出违约后果与潜在损失，才能激活高智力模型的正确权衡。

5. 用“生成-评估-修复”闭环取代单体巨型 Prompt：复杂强约束任务拆成三步轻量流水线，不仅 Token 消耗与延迟直接腰斩，还能动态兼容运行时软约束。

建议直接收藏，今晚就花半小时看完，认知彻底重置。

[Ruiteng Huang @huangruiteng]

读了下 RSIGym，本质是自进化改进下面这两套结构化文本字段。万物皆可结构化的思路，loopx 也是如此。
ImproverGenome(
diagnosis_prompt,
mutation_strategy,
candidate_selection_policy,
)
AgentGenome(
system_prompt,
planning_policy,
skills,
memory_policy,
verification_policy,
)

[Rosler H @StarAlone999]

@qkjOECRrDv33469 @MaitreyaBhakal Like what?

[Gill @gurtej__gill_]

Great Paper from NVIDIA.

We’ve spent the last couple of years treating RAG and long context as two completely separate engineering headaches.

One is an external pipeline of bi-encoders and rerankers, the other is an expensive brute-force attention problem.

This paper makes a super compelling case that this split is mostly artificial.

Instead of bolting on external retrievers, they tap straight into the frozen LLM’s own internal representations to select evidence across both regimes.

The wild part is that the backbone stays completely frozen.

They only add around 500k trainable parameters.

On a massive 3B-token Wikipedia index, this tiny tweak beats traditional dense and hybrid retriever-reranker stacks across the board (HotpotQA recall jumps from 49% to 73% and 2Wiki leaps from 31% to 60%).

But it also solves the needle-in-a-haystack problem from the other direction.

On long-context benchmarks, using that same internal mechanism to aggressively prune distractors before generation pushes NoLiMA accuracy at 128k context from basically flatlining at 1% up to nearly 25%.

Because it strips out the fluff early, you get a massive drop in FLOPs and time-to-first-token past 32k tokens instead of watching your latency explode.

Great work by Edan Kinderman and the team.

Read the full paper here: https://t.co/tNREGwtKZP

[How To Prompt @HowToPrompt__]

China quietly solved the biggest problem in RL training.

Kimi published a paper that makes LLM reinforcement learning 2× faster.

Right now, every major lab uses Reinforcement Learning to make their models smarter.

But there’s a massive structural flaw in how it works.

During training, the AI is asked to generate multiple different answers to the same prompt so it can figure out which one is best.

Here is the problem: some answers are short. Some answers are long.

Because the system is synchronous, the entire multi-million dollar GPU cluster has to sit and wait for the single longest answer to finish generating before it can update the model.

It’s like keeping the entire class inside until the slowest kid finishes the test.

The GPUs sit idle. Throughput dies.

Kimi researchers dropped a paper introducing "Seer," a system that completely eliminates this bottleneck.

Instead of waiting, Seer uses "Online Context Learning."

It dynamically splits the generation process into movable chunks and learns from the other responses to the same prompt on the fly.

It stops the GPUs from waiting.

The results are staggering.

Seer achieves up to a 97% improvement in training throughput compared to state-of-the-art baselines.

It reduces that dreaded long-tail waiting time by 93%.

No massive architecture overhaul. Just a radically smarter way to manage the flow of data.

[Rosler H @StarAlone999]

@kumadanko_001 笑容不是该出现在普通中国消费者脸上吗？可以买到便宜的大豆和牛肉了

[Rosler H @StarAlone999]

@turningpointjpn Why US not cracked down IRAN?

[Rishi Khare @ COLM @rishiskhare]

CacheFuse is an early attempt in cross-model KV cache reuse that demonstrates that linearly interpolating the KV caches of two models can outperform a standalone model’s cache alone. I’ll be presenting it tomorrow at #COLM Efficient Reasoning Workshop - stop by!

[Rosler H @StarAlone999]

@8tJ3zm9YrV5hKr9 @miren_41319 how? By a lunatic?

[SuSu_酥酥👅 @NFT_Chen]

🔥必读！Qwen 负责人刘大一恒团队把极稀疏 MoE 的负载不均当成闭环控制来解！这对Qwen 4.0 而言意义重大！

他们发现：专家池越大、每 token 只点亮一小撮，容量能涨、推理成本几乎不动。DeepSeekMoE 16B 是 64 专家选 6 个（激活约 9.4%），Kimi K3 已到 896 专家选 16 个（约 1.8%）。再往 5 万亿、10 万亿走，热门专家会打满 GPU，闲置专家却训不透。

ID Balancing 的关键点：
1️⃣给现有方法定身份。DeepSeek 的 loss-free 是固定步长积分控制，Kimi K3 的 Quantile Balancing 是广义比例控制，都是不完整 PID。

2️⃣新方法是 ID 控制器，不用辅助 loss。积分项按负载误差缩放，偏得越大纠得越猛，接近均衡就收步；微分项只在失衡正在恶化时打开，提前刹车。

3️⃣768 专家、Top-10 / Top-5 / Top-3 对比。最稀的 Top-3 里，相对最好基线，主干最坏 MaxVio 降超过 50%，训练平均 MinVio 降超过 12%。辅助 loss 的 MaxVio 可飙到约 211，ID Balancing 约 15。

4️⃣规模从 18.9B 拉到 69.9B，最坏 MaxVio 几乎不动（约 8.4→7.8），比辅助 loss 低约 89.6%。越稀疏优势越大，语言建模和下游不掉点。

大误差大力纠，恶化才加导数，靠近均衡就收手。

论文：https://t.co/RXYI8cVGiT

#MoE #Qwen #刘大一恒 #稀疏MoE #PID #阿里巴巴 #Qwen4

[旺旺 @TiagerBao]

现在很多 Agent 所谓的“长期记忆”，其实只是把历史聊天存起来再搜一遍。

Hindsight 想做得更深一点：

让 Agent 不只是记住，而是从过去的经历里形成新的判断。

核心就 3 个动作：

Retain：记住事实和经历
Recall：找回相关记忆
Reflect：把多段记忆放一起分析，形成新的结论
检索也不是只跑向量搜索，而是同时走：
语义 / BM25 / Graph / 时间线

更有意思的是，它会把长期积累的内容整理成：

Observations → Mental Models → Knowledge Pages

比如 Claude Code 用久了，可以逐渐记住这个仓库的架构、代码习惯、历史决策和正在做的事情，下次 Session 不用重新解释一遍。

目前已经支持 Claude Code、Codex、Cursor、GitHub Copilot 等，还自带 MCP Server。

Agent 真要长期干活，记忆不能只是“搜到以前说过什么”，还得知道这些经历最后说明了什么。

GitHub：https://t.co/I1FdD36vyk

#AI #Agent #Memory #ClaudeCode #Codex #MCP #RAG #OpenSource

[Superman @thesupermannx]

Samsung open-sourced a method that shrinks 13B parameter LLM into less than 1 GB.

It's called "LittleBit"

Instead of storing AI weights as standard numbers, they used latent factorization to crush them down to extreme sub-1-bit levels.

In some configurations, they hit 0.1 bits per weight.

But here is where the architecture gets crazy.

When you compress an AI this much, you can stop doing math.

Instead of forcing the hardware to do heavy floating-point multiplication, LittleBit replaces the core computation with a bitwise XOR operation.

It swaps complex matrix math for basic sign flips.

The results rewrite the rules of model deployment:

• Unlocks a massive 11.6x inference speedup relative to standard FP16 models.

• Radically reduces memory footprint and loading bandwidth.

• Maintains robustness in extreme sub-0.5 bit regimes where previous compression methods catastrophically fail.

This is not a clever optimization.

It is the blueprint for running massive, state-of-the-art AI locally on cheap, resource-constrained devices.

[Rosler H @StarAlone999]

@CryptoNobler what app？

[Rosler H @StarAlone999]

@xianebvjsjj 侵华战争炸了多少？杀了多少人？

[Juncheng Yang @1a1a11a]

The largest agentic LLM Inference dataset has arrived!
206B tokens · 12,002 sessions · 1,186,582 LLM requests · 1,213,347 tool calls
The traces include tool names, arguments, durations, and more.

Explore the dataset on @huggingface https://t.co/12hrLuupVF

Paper coming soon! Stay tuned for more at https://t.co/8NnRV9xiN5.