Han Xiao's banner

Han Xiao

@hxiao • 20,794 subscribers

VP, AI @Elastic prev: founder & ceo @JinaAI_

Shorts

I love Clawdbot, but most parts can be just Claude Code --dangerously-skip-permissions + pipe via Telegram. Made a simple version using cloudflare tunnel + tmux + StopHook.

I love Clawdbot, but most parts can be just Claude Code --dangerously-skip-permissions + pipe via Telegram. Made a simple version using cloudflare tunnel + tmux + StopHook.

317,139 Aufrufe

Saw a poster at ICLR showing contrastive learning (InfoNCE included!) ≈ a closed-form spectral decomposition in RKHS. Got curious whether a map could adapt embedding spaces across model families, while preserving NDCG retrieval perf. Did some experiments on my flights back to sf. Video shows diff maps on simple swiss roll data.

Saw a poster at ICLR showing contrastive learning (InfoNCE included!) ≈ a closed-form spectral decomposition in RKHS. Got curious whether a map could adapt embedding spaces across model families, while preserving NDCG retrieval perf. Did some experiments on my flights back to sf. Video shows diff maps on simple swiss roll data.

14,349 Aufrufe

Pure MLX of UMAP, t-SNE, PaCMAP, TriMap, DREAMS, and CNE for Apple Silicon. Beyond the algorithms, I also built a circle-splatting renderer in MLX - scatter-add alpha blending on Metal, piped to H264 hardware encoding. So 70K points, 784 dimensions, from raw data to rendered video < 5s on M3 Ultra. pip install mlx-vis Paper: Code:

Pure MLX of UMAP, t-SNE, PaCMAP, TriMap, DREAMS, and CNE for Apple Silicon. Beyond the algorithms, I also built a circle-splatting renderer in MLX - scatter-add alpha blending on Metal, piped to H264 hardware encoding. So 70K points, 784 dimensions, from raw data to rendered video < 5s on M3 Ultra. pip install mlx-vis Paper: Code:

15,950 Aufrufe

Videos

Anya Rossi

sweetdream.ai

SweetDream.ai•Sponsored•Livecam

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Exclusive private shows

1.2k viewers online

Private Show

Join now for exclusive access

Free preview available • Premium content

在《金庸群侠传》crpg-bench上跑了下 Fable 5.1 感觉并没有很惊艳，在玩游戏上智商水平和fable5/opus5/sonnet5没有显著差异。金庸群侠传bench是我用来测试大模型在long-horizon task能力的一个沙盒。评测方法很简单：通过截取游戏画面，视觉反馈，执行动作，不断循环，最终找齐14本天书完成游戏。玩过《金庸群侠传》的都懂： - 游戏是开放世界，通过升级招人打怪，最终集齐14本天书结束游戏，所以是彻头彻尾的long-horizon task。 - 游戏采用2.5D斜45度视角，加上1996年的320x200分辨率，绝对是VLM的噩梦：怪视角+渣画质+分布外。 - Caveat：虽然是开放式，但开局必先去找南贤北丑，不然寸步难行而且很容易误闯故事线直接被打死。至少我一开始是这么设计的：完全根据游戏的进度来设计评测指标。但实际做了一段测试后发现，很多frontier大模型连走出始发地都非常困难，更别提集齐14本天书。我也特地设计了相应的SKILL，让模型能对游戏内容任务和操作有基本的了解。为了避免消耗无意义的token，评测就局限在给定20分钟的时间下，谁能探索的最多就算谁厉害些。也就基本沦为了2.5D视角下视觉迷宫问题。经过测试了一些国产和国外的模型后，20分钟内能顺利走出出生地的模型是： - Fable 5.1：4分钟 - Fable 5：3分钟 - Opus 5：5分钟 - Sonnet 5：6分钟 - Gemini-3.7-flash：12分钟并且上述这些模型没有一个成功找到南贤居。

在《金庸群侠传》crpg-bench上跑了下 Fable 5.1 感觉并没有很惊艳，在玩游戏上智商水平和fable5/opus5/sonnet5没有显著差异。金庸群侠传bench是我用来测试大模型在long-horizon task能力的一个沙盒。评测方法很简单：通过截取游戏画面，视觉反馈，执行动作，不断循环，最终找齐14本天书完成游戏。玩过《金庸群侠传》的都懂： - 游戏是开放世界，通过升级招人打怪，最终集齐14本天书结束游戏，所以是彻头彻尾的long-horizon task。 - 游戏采用2.5D斜45度视角，加上1996年的320x200分辨率，绝对是VLM的噩梦：怪视角+渣画质+分布外。 - Caveat：虽然是开放式，但开局必先去找南贤北丑，不然寸步难行而且很容易误闯故事线直接被打死。至少我一开始是这么设计的：完全根据游戏的进度来设计评测指标。但实际做了一段测试后发现，很多frontier大模型连走出始发地都非常困难，更别提集齐14本天书。我也特地设计了相应的SKILL，让模型能对游戏内容任务和操作有基本的了解。为了避免消耗无意义的token，评测就局限在给定20分钟的时间下，谁能探索的最多就算谁厉害些。也就基本沦为了2.5D视角下视觉迷宫问题。经过测试了一些国产和国外的模型后，20分钟内能顺利走出出生地的模型是： - Fable 5.1：4分钟 - Fable 5：3分钟 - Opus 5：5分钟 - Sonnet 5：6分钟 - Gemini-3.7-flash：12分钟并且上述这些模型没有一个成功找到南贤居。

40,742 Aufrufe • vor 14 Tagen

让qwen3.8-27b-nvfp4自己玩《金庸群侠传》，模型使用的是 SGLang + RadixArk 等大神优化后的版本bf16-lmhead 今天刚发布的那个版本。游戏搞个最小的e2-micro直接跑dosbox + 前端stream 然后把游戏按键通过api暴露出来放到一个skills文件。最后本地pi-dev 绑定qwen3.8-27b-nvfp4，直接把skills 复制粘贴过去就可以了。大家可以感受一下1/2的rtx-6000-pro的vision+thinking的速度。打算让他这么跑一晚，看看明天一早能收集到几本天书。（比较搞的是，只是一上来起人名那个注音符号输入法就浪费一堆时间，跟我当年一样）

让qwen3.8-27b-nvfp4自己玩《金庸群侠传》，模型使用的是 SGLang + RadixArk 等大神优化后的版本bf16-lmhead 今天刚发布的那个版本。游戏搞个最小的e2-micro直接跑dosbox + 前端stream 然后把游戏按键通过api暴露出来放到一个skills文件。最后本地pi-dev 绑定qwen3.8-27b-nvfp4，直接把skills 复制粘贴过去就可以了。大家可以感受一下1/2的rtx-6000-pro的vision+thinking的速度。打算让他这么跑一晚，看看明天一早能收集到几本天书。（比较搞的是，只是一上来起人名那个注音符号输入法就浪费一堆时间，跟我当年一样）

17,356 Aufrufe • vor 21 Tagen

Not a fan of Knowledge Graphs, but recently I started using them more often for a surprising reason: to build non-trivial private verifiers for agentic search. For those who don't know, building a private eval set for a scaffolded LLM in 2026 is really challenging, like seriously hard. It takes a lot of effort to find a question that's non-trivial to a scaffolded LLM yet still answerable. To find those question-answer pairs, I built a knowledge graph extractor where you can throw a corpus at it, and it extracts the entity relations using qwen3.6-35b-a3b-MTP on an L4 at 70 tps (which is really good for such a low-budget GPU). Then I mark out the longest path in the graph and use it to generate challenging question-answer pairs. The idea is to find those genuinely multi-hop fact chains that are verifiable from the corpus, to stress-test the agentic search system.

Not a fan of Knowledge Graphs, but recently I started using them more often for a surprising reason: to build non-trivial private verifiers for agentic search. For those who don't know, building a private eval set for a scaffolded LLM in 2026 is really challenging, like seriously hard. It takes a lot of effort to find a question that's non-trivial to a scaffolded LLM yet still answerable. To find those question-answer pairs, I built a knowledge graph extractor where you can throw a corpus at it, and it extracts the entity relations using qwen3.6-35b-a3b-MTP on an L4 at 70 tps (which is really good for such a low-budget GPU). Then I mark out the longest path in the graph and use it to generate challenging question-answer pairs. The idea is to find those genuinely multi-hop fact chains that are verifiable from the corpus, to stress-test the agentic search system.

58,118 Aufrufe • vor 3 Monaten

If you only have 60s of attention for Kimi's Attention Residuals paper, watch this.

If you only have 60s of attention for Kimi's Attention Residuals paper, watch this.

84,755 Aufrufe • vor 6 Monaten

low quant weights make the embedding model lose all discriminative power. I plotted the cosine correlation matrix of jina-v5, and one can see that low quant makes the model really blind. The off-diagonal similarities are pretty high on Q1/2/3, meaning everything looks similar in the semantic space. Q4 is a sweet spot where model quality becomes acceptable.

low quant weights make the embedding model lose all discriminative power. I plotted the cosine correlation matrix of jina-v5, and one can see that low quant makes the model really blind. The off-diagonal similarities are pretty high on Q1/2/3, meaning everything looks similar in the semantic space. Q4 is a sweet spot where model quality becomes acceptable.

Han Xiao ✈️ ICML 2026

63,042 Aufrufe • vor 5 Monaten

after turboquant and qwen3.5-35b-a3b, i got curious: how realistic is it to use kv cache as a document store today? to have vectorless, RAG-less search. so i prefilled 258K out of 262K context window on L4 (a budget GPU popular in prod). ~99% of the slot is pre-computed and stored, users load it on the fly in ~1s. system prompt + query append to the end, generation takes ~3K tokens, enough for search. at 99% fill rate, decoding runs ~20 tps on L4. i prepared some ego datasets (jina papers, which i know best), plus popular novels in chinese and english. the results are actually pretty good. some hallucination, but most answers are solid and well-grounded. what's more interesting is the cost: ~$0.26/h on L4 spot. single LLM. no vector database, no embedding model, no workflow/pipeline engineering. using kv cache as document store is nothing new, like the old CAG paper. but with quantized kv cache and modern attention (hybrid SSM-attention, GQA, MQA, MLA), the economics are changing fast. if we solve cold-prefill speed and decoding speed, and budget GPU costs keep dropping, the future of search could be vectorless. radical, but possible.

after turboquant and qwen3.5-35b-a3b, i got curious: how realistic is it to use kv cache as a document store today? to have vectorless, RAG-less search. so i prefilled 258K out of 262K context window on L4 (a budget GPU popular in prod). ~99% of the slot is pre-computed and stored, users load it on the fly in ~1s. system prompt + query append to the end, generation takes ~3K tokens, enough for search. at 99% fill rate, decoding runs ~20 tps on L4. i prepared some ego datasets (jina papers, which i know best), plus popular novels in chinese and english. the results are actually pretty good. some hallucination, but most answers are solid and well-grounded. what's more interesting is the cost: ~$0.26/h on L4 spot. single LLM. no vector database, no embedding model, no workflow/pipeline engineering. using kv cache as document store is nothing new, like the old CAG paper. but with quantized kv cache and modern attention (hybrid SSM-attention, GQA, MQA, MLA), the economics are changing fast. if we solve cold-prefill speed and decoding speed, and budget GPU costs keep dropping, the future of search could be vectorless. radical, but possible.

42,451 Aufrufe • vor 5 Monaten

Gave my autoresearch x TTC x retrieval talk at AIE yesterday, ran it again at home today. things moving so fast that after ICML next week I'll probably be onto something new, and these ideas will already be disposable. so before that happens, I want to drop a summary of the last 2 months of experiments across several projects, mostly so future me remembers they existed.

Gave my autoresearch x TTC x retrieval talk at AIE yesterday, ran it again at home today. things moving so fast that after ICML next week I'll probably be onto something new, and these ideas will already be disposable. so before that happens, I want to drop a summary of the last 2 months of experiments across several projects, mostly so future me remembers they existed.

Han Xiao ✈️ ICML 2026

17,877 Aufrufe • vor 2 Monaten

Rationale是我司推出的一款专为管理者和决策者打造的分析成效工具，它集成了最新的GPT3.x和上下文学习(in-context learning)技术，能够快速生成Pros & Cons和SWOT分析报告，帮助管理者和个人做出明智的决策。 Michael Anti

Rationale是我司推出的一款专为管理者和决策者打造的分析成效工具，它集成了最新的GPT3.x和上下文学习(in-context learning)技术，能够快速生成Pros & Cons和SWOT分析报告，帮助管理者和个人做出明智的决策。 Michael Anti

149,353 Aufrufe • vor 3 Jahren

Given an embedding vector, you can tell which model produced it. I trained a 0.8M transformer that fingerprints embedding models by reading raw float digits (vocab size: 15). Full end-to-end, zero feature engineering.

Given an embedding vector, you can tell which model produced it. I trained a 0.8M transformer that fingerprints embedding models by reading raw float digits (vocab size: 15). Full end-to-end, zero feature engineering.

19,695 Aufrufe • vor 6 Monaten

Keine weiteren Inhalte verfügbar