Tianjiao Ding

Glad to see the renaissance/revival of sparse structures brought by AI bigheads, from Anthropic to OpenAI! Instead of training extra AEs and manually interpreting AE features, our latest paper decomposes the activations along concept vectors that have semantic meanings by design via classic compressed sensing. Not only is the decomposition fast, but it also reveals the extent of concepts in each activation, allowing for effectively steering LLMs’ behavior. SOTA performance on detoxification and other alignment tasks. Paper: arxiv.org/abs/2406.04331 TLDR/Code/Dataset: t.co/zXPA1YluiJ

@ziv_ravid 1. This policy was only applied to people who agreed not to use LLMs for their reviews, but then used LLMs anyways. It's not an anti-LLM policy, it's a rule-following policy. 2. More sophisticated methods were used than AI detectors. Post from @icmlconf tomorrow.

To ensure compliance w peer-review policies, ICML has removed 795 reviews (1% of total) by reviewers who used LLMs when they explicitly agreed to not. Consequently, 497 papers (2% of all submissions) of these (reciprocal) reviewers have been desk rejected Details in blog post 👇

@alfcnz Quick Q: where do you define each agent, skills? Sub agents? Some openclaw?

I got to unfuck myself by creating a team of agents for my book. We have: Editor (root), Typesetter and Illustrator (latex), Cyberneticist and Plumber (python), Examiner (exercises). They cannot edit, but tell me what to work on next. I’m literally flying! Blog coming out soon.

@Chiuchiyin What a steal. I regret not packing all of these to the U.S.

Should I?

@Chiuchiyin Pretty sure someone were on them when building NNs.

Can we build neural networks with mushrooms?

How do you change directory?

Isn't China banned from Claude.ai at least officially? 🤣

Someone asked me why the new HPC/AI clusters at @Penn is named Betty. 🤣 Here you go!

Symmetry is one of the things that are elegant in the mathematical form and applications!

@simon_zhai Appreciate it man this is fire

@RickTDing Updated, check again bro

Introducing another fun project: Financial Independent Retire Early (FIRE🔥) calculator -- yx-s-z.github.io/fire/

@simon_zhai Why is Shanghai not in the default list haiyaaaa

This calculator will give you a rough estimate of how many asset you will need to retire early in the selected city.

@Chiuchiyin There’s also 伐 呢 吧 喔 唷 撒 if you are interested…

The HR staff who handle my paperwork and support me often end her messages with “哈” or “呗” is this Northern thing?

@simon_zhai @Yuchenj_UW For mysterious reasons my roommate’s Claude harnessed by gh copilot often responds in Chinese even tho he never code / ask anything in Chinese

@Yuchenj_UW I hope one day when the Chinese open source model catches up, ppl will start coding in Chinese 🤪

I’ve noticed something: When Claude is down, no software engineer says, “Fine, I’ll just write code myself.” They complain, then speed-run to Codex or OpenCode. We’ve lost that ancient skill of manual coding. English is now the only programming language.

@alfcnz Hold on - what? 🤣

*Alf sees student, starts speaking Chinese* *Student doesn't realise that, answers in English* *Alf asks if student is Chinese* *Student notices, switches to Chinese, confirms* *Alf reveals his nationality* *Student, actually born in Italy, switches to Italian!* 🤯🤯🤯

Wow. What is impressive is that these aren't just one-off tricks but look like they are being done reliably by many robots at the same time in front of a big audience.

I've recently started handing off llm generated proofs to Aristotle with some success. It takes more time than reading the proof myself, but I can at least avoid the effort / disappointment of reading something incorrect and do something else in the meantime.

🚀 We present the first large-scale Lean 4 formulation of Statistical learning theory from scratch! Led by my student @yuanhezhang6 and collaborated with @jasondeanlee 📄 Paper: arxiv.org/abs/2602.02285 💻 GitHub: github.com/YuanheZ/lean-s… 🤗 Dataset: huggingface.co/collections/li…

@JacquesThibs @MillionInt I came to feel similarly: without a plan I'd be overly committed to something and miss something more important. How do you do this one hour - do you split the time equally among projects? Or from planning high levels to low levels?

@MillionInt Every morning I devote at least an hour for planning the tasks for the day, if not more. Definitely what sets things apart these days.

Run fewer experiments and think about them more

A device that visualizes how a computer performs calculations

Made possible by Claude Cowork (Anthropic please send me $$, or even better, job offers 😉)