Han Xiao ✈️ ICML 2026 (@hxiao)

interesting position paper throwing cold water on autoresearch/ai scientist: LLMs can't jump. The thought experiment is this: Take an LLM with a 1905 knowledge cutoff. Feed it every paper, every dataset, every equation of that era. Could it invent general relativity? No. Discovery isn't one thing. It's three. You can induce — generalize from data, which lands you at Newton plus some epicycles to explain Mercury's weird orbit. You can deduce — derive rigorously from axioms you already have, which never gives you new axioms. Or you can jump — invent the frame itself, decide that spacetime curves. That third move is the one that matters, and it's exactly the one induction and deduction can't reach. Penrose put it as three worlds: Physical, Mental, Platonic. Data flows from the world into a mind fine. But the new law has to be discovered into the Platonic world first — and that step is the jump. LLMs are induction machines running over what already exists. Structurally, they don't take it. I think it’s a warning to AI scientists/autoresearch against collapsing two very different things into one word. Hill-climbing: LLMs are already superhuman here, and autoresearch in this sense is real and moving fast. Abduction/leap/jump: a new frame that reorganizes the field, that is a different act entirely, and nothing about scaling induction suggests you get there. Most of what Autoresearch ships today will be spectacular hill-climbing. The jump is still ours for now.

Pingcap可以说激励了近十年来一大批做开源软件公司的国人。2020年我开始做jina的时候 投资人给的建议就是对外看elastic和mongodb对内看pingcap和zilliz。6年后的今天我还经常能从各个meetup上看到pingcap zilliz在agent/memory上的各种创新，创过业的人都知道非常不容易。有时候创始人的热情可能还在，还能撸起袖子pivot几次、上场再战，但团队已经老了，不是身体老是心态老了。所以我对于P&Z两家公司挺敬佩的。我觉得文章中提到的growth确实对于创业公司非常重要。有句话不是说 “growth without revenue is a disaster; revenue without growth is boring”。硅谷经常谈论upside，说啊你加入谁谁谁 买谁谁谁的股票有很大的upside，我觉得和文中提到的故事是一个道理：别人用你的产品看中的是你比google或竞争对手的更大的那个upside。 关于文中提到的“25岁以下的人不用你产品，你就完了”我觉得也不绝对。在硅谷sf云集了一帮high agency的年轻人，每天sfmaxxing tokenmaxxing的玩儿，自然会觉得能在他们圈中混开是个成就。但很多来钱的地方在于boring/chore/public sector。这些人和sf人精神面貌和思维方式完全两码事。但是一个公司的品味在这时就显得很重要，尤其是在公司经历中年危机时，要么pretend to be young & cool, 墨镜一戴谁都不爱，装嫩到底；要么就别装直接fashionable old，走法拉利老爷老了也还是你大爷那套路线。卡在中间，嫩的学不会，老的不甘心，公司就迟早拖死。

zhipu和deepseek在25年春都曾经是jina reader的数一数二的大客户，也都是由我直接founder support。二者给我留下的印象就是非常精，对技术指标要求非常苛刻，动不动就p99 latency要降到多少。沟通基本就是一个微信群要随叫随到，从他们手里赚点钱非常难。zhipu这边买了用了我们服务两个月后，来一个清华来的实习工程师发现我们jina 官网的api token价格比给他们2b的价格便宜不少，觉得我是奸商欺骗他们，我解释说2b价格本来就不一样，其次我直接为你服务调优debug这些什么都不算么？他说不算。我这人ego比较大当时是一肚子火，一方面我觉得zhipu太小气，面对我们一个成长中的startup完全没有任何的支持。另一方面，第一性原则角度来讲他们也没做错什么，无非是想要最低价。他们也没有义务帮生态中的任何玩家。我觉得b2b销售本来就非我所长，自己ego太高放不下，即便放下我做founder sale极度缺乏耐心。再加上jina本身除reader服务之外需要我做大量技术管理，来美国后又忙并购的事情，zhipu这个2b的deal就凉了下来。

Built sth I've always wanted, using jina-v5-omni embeddings locally for multimodal file search on mac. + airgapped & fully local, indexes text, PDF, image, audio, and video + Swift-native UI + mlx-swift-transformer optimized core. No Python. + tested on M3 Pro 18G / M3 Ultra 512G / M4 Pro 48G. all work fine, no OOM + local HTTP server exposes index to agents like OpenClaw & Hermes

seeing a pattern shift in how I use agents for long-horizon tasks in 2026 vs. how i do deep research in 2025. 2025's DR: search → read → reason → repeat until done. Everything hits the web on every loop. but 2026 long-horizon tasks have two distinct phases. - Phase 1: Web IO for research & planning. Search, read, reason. but the goal isn't to produce an answer directly. It's to materialize web knowledge into local files (.md/.json/.csv). - Phase 2: Agent "mounts" the files and starts the loop. The agent reads, runs, writes against those local files only. No more web calls for grounding. Why I cut web grounding IO in Phase 2? • Determinism: local files are immutable snapshots. Web content shifts, 404s, hits paywalls • Speed: filesystem reads are ms, web fetches are seconds. Agent loops need tight iteration • Consistency: cross-checking requires operating on the same knowledge base, not fetching different versions each time • Cost: web IO burns tokens parsing HTML noise. Local files are already clean