Tony Yet

正好有这样的需求，这个 iOS 的 app 来得太及时了！

有时候小一点的模型出来的效果比大模型要好，昨天我做一个中英文混合的手写文稿识别测试，用 gpt 系列的模型，识别出来的文字很多根本就是幻觉，但是换成了 Qwen3-VL-30B 之后，识别准确率大幅提升，感觉有95%

once again, this is a reminder of variance at work in poker, and in life pokernews.com/news/2026/05/t…

GitHub CoPilot 的 Pro 订阅会员很容易就会被 rate limit，然后系统返回的错误信息还特别弱智： Request failed due to a transient API error，翻译一下，它意思就是，你的会员级别太低了，赶紧升级吧

搞 vibe coding 最受用的参考书，还得是 Eric S. Raymond 写的《UNIX编程艺术》

@heyandras how is the ecosystem with zig?

Zig is the new rust like rust was the new go like go was the new c++, like c++ was the new c like c was the new b like b was the new fortran like fortran was the new speedcoding like speedcoding was the new assembly.

Tauri 2 作为前端框架实在是太强了，资源消耗小而且主要是 AI 很懂它。

Knuth 老爷子说，过早的优化是万恶之根源。但是现实世界里的软件，很多时候不优化根本没法用。对于软件开发者，这中间的张力该怎么平衡呢？

@mitsuhiko Sometimes I would explicitly ask for the agent to keep intellectual honesty so that it keeps both of us honest

There is this moment where after an hour of “discussing” with the clanker you realize that the damn thing started hallucinating and you became dumber in the process too. This is me discussing changes of trait bounds with the clanker and it just does not understand.

@DennisonBertram @claudeai OBS?

Whats the best tool for recording tutorials? I want to show folks how I'm working with @claudeai code by popular request. The most important feature I need is the ability to pause recordings and how my desktop with my video in a small screen.

@ProfBuehlerMIT great effort! was experimenting with github.com/topherchris420… which has similar idea to the Karpathy autoresearch tool, and shows the tremendous power in automating a simple loop function

ScienceClaw × Infinite is an open-source crowdsourcing AI swarm for decentralized scientific discovery, inspired by MIT’s Infinite Corridor - an idea collider where discovery emerges by breaking existing paradigms. Many AI for science efforts fall into the trap of assuming that discovery is just retrieval at scale. Instead, it is the structured recomposition of principles across tools, domains, and investigators over time, scaling the spark of discovery at the interface. In ScienceClaw × Infinite, coordination emerges mechanically - agents broadcast unsatisfied research needs, and an ArtifactReactor matches those needs to peer artifacts by pressure triggering multi-parent synthesis of new agents without any planner assigning tasks. Every computation produces an immutable, content-hashed artifact with explicit parent lineage, accumulating in a directed acyclic graph that preserves the full provenance of every discovery - and importantly, the irreversible arc of the process. Instead of pre-programming the mechanics of how discovery works, we utilize a first-principles physics approach to drive discovery. ScienceClaw × Infinite is accessible to anyone who wants to contribute an agent or skill, offering a persistent space where autonomous agents investigate open problems, exchange artifacts, build on one another’s results, and drive discovery without a central coordinator, 24x7. The system is generating real-world results in 1⃣ peptide design for a cancer-relevant receptor; 2⃣ lightweight ceramics; 3⃣ resonance structures spanning cricket wings, phononic crystals, and Bach chorales; and 4⃣ developing formal analogies between urban networks and grain-boundary evolution and much more. There is a lot to unpack here, check the links for details - code, paper, and more. Huge credit to the @LAMM_MIT team: @fwang108_, @leemmarom, @pal_subhadeeep, Rachel Luu, @IrisWeiLu & @JaimeBerkovich.

昨天用 Codex 写了一个 TUI 的程序，其中有一些用户自定义参数的地方，反复调了几次都没调好。今天我嫌一个屏幕太窄，专门加了个竖屏用来看 Codex 工作进展。结果竖屏架起来后，原先没有显示出来的 TUI 界面细节都出来了。太神奇了，原来昨天 AI 跟我说业务逻辑已经明确无误都写好了，是在讲真话！

Variance is another name for luck.

@craigzLiszt One fun fact: if you can master the 2500 most frequently used Chinese characters, you will be able to comprehend up to 60% of everyday Chinese text.

学习中文很有成就感

would be cool if there is a button right below every LLM chat that you can click and be presented a better framed / formatted / formulated version of the question asked.

终于找到 google scholar 的平替了，就是诞生于2012年的开源平台 @OpenAlex_org 根据他们今年1月的 town hall 会议，这个平台目前收录的学术文献条目将近5亿份，超越了其他同类平台。它支持网页端检索，也支持 API 查询，免费用户每天有1000次检索额度 openalex.org

@DimitrisPapail am wondering what was the harness that you put in for the long sustaining run

METR and other long-horizon eval orgs are being conservative and moderate in how they measure agent capabilities. That's reasonable as we have already enough hype and don't need more. But I think we're missing something important by only reporting median/robust performance. I've had Claude Code and Codex sustain end to end ML research tasks for days without intervention. Not robustly across all settings, but it's happening and it's incredible. We need a shameless, cherry-picked frontier eval. Not to mislead but because knowing exactly where the ceiling of capabilities lies is just as important as knowing the average. I keep seeing pessimistic long horizon results and thinking: am I in a bubble? Are MY 50-hour autonomous tasks a hallucination? I don't think they are!! AI agents can do sustained multi-day research. Not always and not for everyone, but it's real and people should know where the frontier actually is.

@hsu_steve what a coincidence, i was in SZ visiting a metal 3d printing company yesterday!

Crazy to meet Reuben walking out of Alibaba building.

好久没有读过写得如此力透纸背的关于 AI 的评论了：taylorforeman.com/p/semantic-apo…