david tolpin

"simple software smart, complex software dumb" chatted the sheep at the animal farm. pi.dev's extensibility is a design artifact rather than a goal. @badlogicgames

a better model degrades performance: dtolpin.github.io/2026/07/be… Back in the good old days of Yoom I had trained a neural density estimator on human body poses, and Michal Heker, Sefy Kagarlitskyand yours truly uploaded an arXiv paper (arxiv.org/abs/2507.12138) about this. One trick described in the paper is that for proper density estimation in 6D rotation space, one needs to “de-Gram-Schmidt” the training data, similarly to the dequantization trick, but for points on a lower-dimensional manifold. However, the adoption of this trick was not smooth.

A coding agent is a dumb program that does what an LLM tells it to do.  The LLM that drives the agent was trained on 1) linguistic correctness as an autoregressive text generator 2) human feedback. For human feedback, it was not trained to do "the right thing", there is no way to train like that. It was trained for confirmation bias. Agent team managers always produce slop. Chief surgeons of agent surgical teams have a chance: dtolpin.github.io/2026/07/cl…

Just caught myself checking for AGENTS.md, CLAUDE.md and friends in every new repo I browse on GitHub. If present, mostly a sign of poor design, sloppy code, and extreme unit tests overfit. A rare exception is a repo that is a harness by itself. Agentic scaffolding for development, as opposite to runtime tool-assisted harnesses, should go into a separate repo and be included as a development dependency, rather than pollute the project's source tree.

In its holy war on Israel, Iran has a new mighty proxy - the United States of America. Iran gives orders to USA, USA prevents the IDF from defending Israel against terrorist attacks. This is a greater win of the Ramadan war for Iran than Hormuz tolls. #IranWar‌ #hezbolla

I thought that the relevant book for our times was Orwell's 1984. I see now that the relevant book for our times is Orwell's Animal Farm.

זה לא טראמפ מחר אותנו, זה ביבי. הבטיח סחורה תמורת פרוטקשן, ולא סיפק. השאלה היא האם החשבון סגור כבר או עדיין פתוח, והאמירקאים ימשיכו לגבות מביבי ריבית. כי זה על חשבונינו, אין לביבי ארנק ואין כרטיס אשראי.

כשאנחנו שוב מזדעזעים מרצח ופשיעה אצל בני דודנו הערבים, לא צריכה להדאיג אותנו השאלה "האם זה מגיע להם?". צריכה להטריד אותנו התובננה שזה עוד מעט יגיע אלינו. הרצח והפשיעה. של יהודים ביהודים.

Whoever presents Orban's loss as a victory of good over evil is a hypocrite; Orban lost and Magyar won as a result of democratic elections, it is how democracy works, not a victory over Putin or over Lord Voldemort. #hungary #Orbán

Had bad experience with submitting to @RL_Conference. #RLC has nice reviewer guidelines: docs.google.com/document/d/1… Got the review that basically said "the paper is fine according to the guidelines, but I didn't like it, therefore judged against the guidelines." Withdrew the paper, will submit elsewhere. This paper: arxiv.org/abs/2602.17375

מדינת ישראל חייבת לחוקק חוק שכל אדם ששירת בצבא והגיע לדרגה עם מילה "אלוף" בה מנוע מלהיבחר לכנסת, או אייש כל משרה ציבורית בכירה. זה לא יפתור את כל הבעיות אבל את חלקם המשמעותי כן. קצונה בכירה מעוותת שכל ללא תקנה.

pair-program with the coding agent instead of managing it. you, the coding agent, and the code will benefit.

Claude code is expensive to run. Most code written with claude is in enterprise environments. Enterprise licenses imply 'no learning'. Does it mean that claude cannot learn from most of the new code produced with claude?

A cs/ml paper is mentioned im AGENTS.md. The prompt is "implement the algorithm and reproduce the case studies". No supplementary, no code in the appendix (or ignored). If the agent fails, the paper is desk-rejected.

LLMs are the best thing happened to academic publishing in decades. For a long time, writing convincingly had been the key to success. Now, with LLM's aid, everyone can produce extremely convincing bullshit, and the differentiator is the quality of scientific results, again.

If plain English is better than the programming language you use than you use the wrong programming language.

הלכתי לעשות בדיקות דם. חמש אחיות ערביות מוסלמיות, חייכניות וחינניות. מדברות בינן לבין עצמן בעברית, מתלוצצות מי מהן יותר אשכנזיה. שאלתי מה הסיפור - אמרה לי שמחייבים אותן לדבר בעברית בנוכחות מטופלים. שאלתי מה עם מטופלים ערבים? צחקקה. הן עושות מזה צחוק. אבל זה לא תקין בעיניי.