play this www.puzzlescript.net/play.html?p=...

I complain a lot about RL lately, and here we go again. The CS view of RL is wrong in how it thinks about rewards, already at the setup level. Briefly, the reward computation should be part of the agent, not part of the environment. More at length here: gist.github.com/yoavg/3eb3e7...

the fascinating (to me) quality of hard-core RL researchers (e.g Sutton) is the ability to have an all encompassing view of RL as the basis of intelligence, while at the same time working on super low level stuff like tabular TD algorithms, and yet strongly believe these are actually the same thing

what's the latest-and-greatest attempt to reverse-engineer and document the inner-working of claude-code?

lets talk about "In context learning". it is clearly NOT "learning", because its ephemeral. It IS some form of generalization from examples, which is very cool. but we need a name. how do we call this skill of generalization from example?

חשבתי שההתעלמות מסודאן היא סתם סוג-של הזנחה, אבל הנה אני קורא ספר שמתאמץ מאד להראות ש: - מה שהיה בדארפור זה לא ג'נוסייד אלא רק תוצאת לוואי של מלחמת אזרחים אכזרית - "Darfur is where the language of genocide had become an instrument" - הערבים אינם settlers באפריקה, הם הגיעו ממקומות שונים - המערב אשם

how many of you are aware of ai-2027? how many of you read it? curious what ya'll think.

"in a future in which AI assistants are being used by world leaders and influential figures, aren't you afraid of rogue AI using this access to control the world?" well maybe, but I am more worried about near future in which these influentials act on random AI advice.

LISP code does not have significantly more parentheses than other typical languages. change my mind.

this "feature" in code editors where you type an opening character and it immediately inserts the closing one for you as well -- why?? at the very very best case, you will have to skip over this character with an arrow. so why?

can someone explain "serverless backends" to me? it seems that they run functions on demand. but if these functions cannot access any persistent state, why not run them on the client? the only reason I see is to hide DBs/APIs tokens/secrets from the client, but is that really all there is to it?

i created this thingy yesterday and now I cannot stop watching it. yoavg.github.io/eternal/

a trivia fact about this paper is that we submitted it to arxiv weeks ago, and it was hanging there in limbo for quite a while. apparently because we submitted to "AI" while they moved it to "HCI".

When reading AI reasoning text (aka CoT), we (humans) form a narrative about the underlying computation process, which we take as a transparent explanation of model behavior. But what if our narratives are wrong? We measure that and find it usually is. Now on arXiv: arxiv.org/abs/2508.16599

if you REALLY want to understand DL, you should start by honing your Category Theory skills, as almost everything in DL at its core can be mapped to a functor or an endofunctor.

taking it a step further, I'd say in many cases using the algebra jargon is harmful to understanding, and its better to just describe whats really going on. ie, "we add an L2 penalty term" --> want the sum of squares to be small. "project to vocab space" --> compute similarity to each vocab item.

i'll elaborate: a common computation pattern in DL happens to coincide with a known operator in linear algebra (matmul), and so we conveniently borrow linalg notation and terminology (matrices, vectors, ranks, norms). but this is just jargon. the algebric properties arent needed.

"Modern ML is built on Linear Algebra". lol no its not.

why is "MCP" implemented as server exposing a set of endpoints, rather than as some JSON schema for defining tool descriptions and allowing these JSON files to be accessed over http? what is the purpose/benefit of the middleman server?

you know what, nah, we don't want to close it. it will be just 80% closed.

and now, we will proceed to peacefully close the strait of Hormuz. you know, for the environment.

today, during a peaceful flight over an iranian mountain, a US airplane dropped a mostly peaceful bunker buster bomb, who flew peacefully until it hit the mountain and mostly peaceful facility underneath it. there was a brief period of violent detonation on impact, then peace again.

אחד מלקחי ליל המקלטים אמש הוא שאין לי סבלנות לקרוא ספרות מקצועית, אבל לקרוא ספרות קלה זה די סבבה. מצד שני הספר שהיה לי בנייד, הוא כזה שהתחלתי לקרוא והפסקתי והיתה סיבה שהפסקתי, הוא מייגע ומעפן. בקיצור שילחו המלצות לספרים. עברית או אנגלית, אבל באנגלית ככהנ יהיה לי יותר קל להתארגן.

there are lots of "AI agents are just software around LLMs" comments recently. this is true by definition, but also misses the point. while there is no "magic" in software, there is engineering and best practices. and we dont yet have them for systems that involve LLMs.

a good gag from the other social network:

this "class 9" is such a cool idea for an LLM course! (from ufal.mff.cuni.cz/courses/npfl... via @zdenekkasner.bsky.social )

i created a gist with some non default llm courses gist.github.com/yoavg/95bbc5...

A prompt posted by Twitter user eigenrobot

what's an "eigenprompt"?

i am looking for academic (or similarly "serious") courses or textbook-y texts on LLMs with a major component (lets say >66%) that is NOT about the technicalities of Transformers, optimization, RL, tokenization, etc, but instead focus more on data, eval, interpretability, etc. pointers?

"hmm i wonder what [researcher person] is thinking about these days" /checks google scholar page] [50+ papers in 2025] nope, no signal there.

an interesting observation following neurips's 25k submissions: since the main conferences in "AI" and adjacent fields grew so much, we de-facto dont have 2nd tier venues anymore. only "top tier" venues. and multiple of them. this is a weird situation.

what's really interesting to me in ML these days is the prospect of creating systems* that learn declaratively, without losses, rewards or gradients. we are not there yet, but we will be. * not models

"this studies small models on toy synthetic data in a narrow domain. the findings may not transfer to realistic settings. unclear what we can learn from this.reject" "no no you dont understand. i study spherical frictionless lms in a vacuum. you know, like physicists!" "oh, i see. fantastic work!"

(niche) This is a nice paper but it seems they don't really understand what the word "recursive" means for CFGs, and hence their results are significantly weaker than they think.

btw, even though i dont fully understand what this doc is, i HIGHLY recommend reading it. makes you realize how much product decisions and 'behavior shaping' efforts go into "the AI" that you see at the end model-spec.openai.com/2025-04-11.h...

who here understands what OpenAI's "Model Spec" actually is? Its an interesting doc and I understand the sentiment, but what is it? just a wish list for desired behavior? an annotation guideline? a set of system prompts? llm as a judge prompts? a mix? model-spec.openai.com/2025-04-11.h...

watching a youtube video and got reminded of a nostalgic assembly trick: replacing PUSH PUSH CALL PUSH PUSH PUSH CALL PUSH CALL with PUSH PUSH PUSH PUSH PUSH PUSH CALL CALL CALL to shave a few bytes thanks to better compressibility. those were the days!

I learned that @emilymbender.bsky.social suggests replacing the term "AI" with "Automation", and I think we finally agree on something in this space! I also think that "Automation Intelligence" is a nice name for describing many of these systems and abilities.

leave it to academics to take a praise phrase such as "one of the top 10% smartest researchers in their generation" and interpret it as kinda weak bc "generation" implies a small comparison group of those graduating the same year, and explicitly stating "smartest" implies excluding other qualities.

מי אתם, דור מוזר שלא מוכן לקרוא יותר מכמה עמודים אבל יצפה ב-reaction video של סטרימר באורך כחמש שעות, שמגיב לדיבייט של סטרימרים באורך ארבע וחצי שעות

"LLM on way to replace doctors" gets published in Nature. meanwhile "LLM judgement not as good as human MDs" gets a spot in "Physical Therapy and Rehabilitation Journal".

when you hear or say that "the LLM performed a search before answering", how do you imagine the search process? e.g., what was searched for, using which kind of tools/engines, how many results were retrieved, how much text was read, etc.

כשאתם שומעים או אומרים ש"הללמ חיפש באינטרנט לפני שענה", כיצד אתם מדמיינים את תהליך החיפוש? מה חיפש, כמה תוצאות קרא, וכו.

כולם עפים על השרשור טוויטר של נגה ארבל מתווכחת עם גרוק כ"איזה מטומטמת" ו"איך גרוק הראה לה", ואני מבין שזה בגלל שהיא בצד ה"טועה" של הויכוח וגרוק הוא בצד "הצודק", אבל זה ממש מבאס אותי כצופה מהצד בגלל שהטיעון שלה נגד גרוק הוא נכון: אין לו סימוכין, הוא נשמע משכנע, ואנשים מאמינים לו.

2) the entire "dilemma" in this work and similar, is starting from the premise that there are some user instructions that the models should not obey. without this premise, the "dilemmas" never arise, and the blame sits fully on the human operators. i think that's much healthier.

two thoughts following this interesting work: 1) should we talk about models "lying"? my first reaction was "NO!", but on 2nd thought it might be a way to communicate the situation to people (yes, at the cost of misrepresenting the models and anthrophormizing them). which is better? idk. discuss?

i see many demos recently about "look! we hooked an LLM to [some software] via an MCP and it managed to call it!" this is cool, but also shows little. the real question is how to enable non-trivial, long-running interactions using this software. this is much much harder.

can someone tldr the spain/portugal++ blackout to me? focus on all the known technical details and causes, no social aspects and implications, please.

i checked twice and this post is from today, not some old retweet from 2021