this last one is exactly the "hide keys" scenario right? because otherwise the client could also read/write from s3.
this last one is exactly the "hide keys" scenario right? because otherwise the client could also read/write from s3.
got it. i guess it can be done with views / stored procedures, but maybe people will prefer using a non-db language.
but in the early days the PHP/CGI did also the rendering. now the rendering is done on the client, so why not call the SQL from the client as well?
i guess i also don't understand the majority of the "backend" thing, assuming it is stateless and only orchestrates calls to other services.
i guess i also don't understand the majority of the "backend" thing, assuming it is stateless and only orchestrates calls to other services.
how do you ensure it happens in a serverless function? always call some authentication server before calling the other server(s)? if so, why can't a token-based solution work just as well? (the auth servers gives the client a token which is then used with the other servers)?
wdym by "maintain code on the client"? its a browser and it runs whatever code you put in your static html/js files.
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?
its perhaps THE most fascinating thing about LLMs in my view.
how do you know the ground truth for human steps?
(or we can see it as a metaphor maybe)
yup, happened to me too.. thats a bug. will attempt to fix tonight.
not early - but i did get both balls on the same side after a day+, a bug on my part which should be fixed at some point...
cest ne une rosh shel dag
maybe. should be easy enough to check.
the suspense is like 80% of the fun!!
it kinda looked like a pokeball at some point.
wdym by "similar"? the outer circle cannot change. but it is quite diverse within that constraint.
interesting. i'd say "take the code and try it out", but i suspect it is too brittle and will just trash the hacky collision mechanism and get you some very weird results... (so, "yes", i guess, but for the wrong reasons)
it is currently fully deterministic. it should be easy to introduce randomness in the starting directions of each "ball".
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".
the appealing thing for me about this particular test is that we managed to somewhat-robustly measure something meaningful about the model's "actual" process, *and* sort-of measure how humans perceive it.
i mean, i think about these topics a lot and i think the levels in which we don't understand these things are very much multi-faceted ;) this is just one of them.
I was not familiar, and it looks very interesting. thanks!
what this means it that we don't have a good mental model of how different steps in the LLM generation process depend on each other, and what are the causal relations that make it tick. which in turn means that LLM's explanations might be "transparent", but also that we fail to interpret them.
the test is very simple: given an AI reasoning text broken into steps, we highlight one step as the target, and then ask you to identify the one-out-of-four preceding steps that we choose, which, if removed, will change the target step. turns out this is really hard.
how well do *you* understand how AI reasoning works? test yourself here: do-you-understand-ai.com
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
"attention" duh. also "zoneout", though it was admittedly short-lived.
to be more precise, I did not say they were not "based on" linear algebra, only that linear algebra is for the most part not important to understand how they work, and hardly any improvements came about because someone "knew linear algebra". but yes, it's very similar.
It's live and kicking! paperfinder.allen.ai/chat/ If it was down it was down it was a temporary technical issue.
לג'יט למרות שאתפלא אם השימוש בנקדן מקנה להם בעלות על החומר המנוקד
גם התורמים אותם תורמים, ככל הידוע לי. אבל הארגונים נפרדים זה מזה.
אני הפסקתי להקשיב לרוג'ר ווטרס אבל בעיקר כי הוא די משעמם תכלס.
הי! אני לא שם כבר תקופה אך מדבר איתם מדי פעם. צוות טוב.
wdym by "enable manipulation"? the metaphors i think are only "nice to have" are the "projections between spaces" kind of things.
also, there ARE benefits to taking algebra in undergrad. you learn stuff, and the notions of proofs and abstraction are important to acquire. it just not central to DL.
and there is also this: bsky.app/profile/yoav...
currently it is pre-req also here. but ideally there will be removed and replaced with a more suitable class.
the algebraic terminology is here to stay unfortunately, and we should use it. it doesnt mean we DL is "built on" linear algebra, nor that a linear algebra class should be pre-req.
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.
OR you could do what people actually do in ML these days and associate each symbolic token with a random list of numbers, and let the optimization take care of it.
ah, likely. idk
("any" is a bit extreme because the purists will come and say "ohh but you use commutativity and associativity of addition! thats group theory!!". but i agree with you of course)
it seems to me that a large chunk of ML can actually be characterized as doing dim reduction numerically
ok, but where is PCA important as a building block in ML?
idk the sociology of it. tensors are popular because hardware/software support them, and its convenient to implement batches this way i guess.
where is it used that is central / important?
btw, re the hardware example: the hatdware is not built to perform "norm". it is built to approximate norm over a floating point representation of real numbers. this is highly specific and doesnt enjoy the generality of algebra at all. i think it may not even be commutative and associative.
bsky.app/profile/yoav...
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.
(and the useful parts of the metaphor is also much more geometric than algebric, imo)
i think these are nice metaphors, and i use them daily. i am not at all convinced that they are essential, except for very few select places.
because when we use a term such as "a norm" we get the definition, which is nice, but also a bunch of properties that hold for items of this kind. and if we dont actually rely on these properties later on, then its a waste in terms of what we communicated.
so i would argue that efficiency of computation is nice but also somewhat accidental and not that important. but lets focus on efficiency of communication: my point is kinda exactly around this. in a field that is actually built around algebra, the communication efficiency would be MUCH higher.
טוב גם הטענה "משין לרנינג בנויה על תורת הקבוצות" היא הזויה בעיני. אנחנו כנראה לא מסכימים על ההגדרה של "בנויה על".
אם הטענה היא שאין כלום במתמטיקה בלי אלגברה אז אוקיי, אבל זה קצת טיעון ריק בעיני
איפה אנחנו משתמשים בקונספט הזה? (ואנחנו קצת דוחקים פה את ההגדרה של אלגברה לדעתי, כי אני גם יכול להגיד כך גם שכל אנליזה פונקציונאלית זה בעצם אלגברה, אבל בוא נזניח את זה לרגע)
תרחיב על לפרמל ייצוג וקירוב? למה זה מעניין אותנו ומה האספקטים האלגבריים שם?
what efficiency are we discussing here? efficiency of communication, or efficiency of computation?
סליחה התכוונתי '' "משין לרנינג על אופטימיזציה" זו טענה שאני בכיף ''
"משין לרנינג בנוי על אופטימיזציה" זו טענה שאני מקבל בכייף
i agree that the language is being used. but the language is pretty much the only part that is being meaningfully used.
linear algebra is also concerned with solving systems of linear equations, the representations of linear equations as matrices/vectors, and related objects like spans and bases.
for me, the "algebra" part is the realization that the real numbers and addition/multiplication over them are just a special case of a "group" or a "field", and that many other kinds of groups and fields exist, and can be manipulated similarly, and share many properties.
LoRA can be described without discussing ranks at all, and it would be just as effective. it doesn't rely on any property of rank to work.
i would say maybe geometric intuitions and how matrix operators relate to geometry? distance functions? JL lemma? not sure. this sounds like a fun class to have, i must say! happy to learn what you *do* teach.
what do you teach?
i think you need to learn the terminology. you do not need a traditional linear algebra course, which will be a waste of time from the perspective of ML (of course, its a beautiful topic on its own).
well if you consider the commutativity and associativity of addition to be linear algebra, then sure, its in. what about convolutions? what mathematical/algebraic properties of them are important?
i do agree that some amount of LA creep in through optimization, yes. but i don't see how the perceived motivations for representation learning is relevant here.
what intuitions would these be, for the case of linalg use in ML? matmul is defined the way it is in linalg for a reason, but i dont see how these reasons matter for the ML use.
why is this statement useful / needed for ML?
there are many results and concepts in linalg beyond determinants. they are not used in ML. even norms arent really used as norms (what properties of norms are needed, beyond the definitions?)
good thing that we have GPUs now, then.
writing matmul by hand in C wouldnt be that bad in terms of performance, and also adding SIMD support to this code. it is nice that we had BLAS routines available, but i dont think it was a deal breaker.
we could probably invent DL without any linalg concepts if these didnt exist. all we need is scalar addition and multiplication (and associated calculus rules). yes, this includes regularization and things like LoRA. figures like the one below kinda proves the point:
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.
it is wrong.
prove me wrong by listing linear algebra topics / results that are central (or even just important) for ML, modern or otherwise.
"Modern ML is built on Linear Algebra". lol no its not.
yes, but this can be achieved also if i provide a json schema for describing tools and point to files containing such descriptions, with the implementations being done externally to the protocol. what's the benefit of coupling in the implementation?
can you elaborate on these "more than API calls"? what else does it provide?
the issue is that the MCP couples two things: tool discovery (an LLM friendly description of what tools are available and how to invoke them), and tool implementation. but "implementation" both takes effort, and doesn't need new standards. so why not focus only on discovery?
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.
for some reason, iran saw this mostly peaceful operation as an act of aggression. they will retaliate in a peaceful manner by causing chaos.
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.
אולי זה תלוי את מי קוראים ראשון
אחרי קבליר וקליי הכל מתגמד
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.
לא יודע אם זה שונה (אני לא חזק בעולמות הטקסונומיה של הדמגוגיה). אני רק אומר שזה שהשמאל "לא ממציא דברים כמו הימין", אז זה אולי נכון אבל זה רק כי יש לו טכניקות אחרות לאותה מטרה
טוב זה לא ממש מקרי קיצון, כי המקרים שונים זה מזה, אבל יש פה מגוון של טכניקות שלא ממציאות עובדות שגויות ועדיין יוצרות דמגוגיה אפקטיבית
במקרי הקיצון יותר יש את מקס בלומנטל עם 'הישראלים הרגו את עצמם בשביעי באוקטובר' ו(להבדיל, בכל זאת..) את הכתבות 'בינה מלאכותית בצהל' של יובל אברהם בבמה מקומית ו972.
(וספציפית פה, זה לא שהם המציאו את הסיפור. הם מהדהדים נרטיב שמישהו אחר דוחף, עד כמה שאני רואה)
כן זה בדיוק מה שאני אומר. הטכניקה אחרת.
שונרא
