Dar Sbitar

[Dar Sbitar @user_settings]

@DJSnM How did it taste?

[Dar Sbitar @user_settings]

@SovereignSteak @andrew__rea I had Gemini 3.6 Flash make a grammar mistake (it wrote "an guest") a couple of weeks ago. I had to check and recheck multiple times to make sure I didn't misread anything.

[Dar Sbitar @user_settings]

@TomNavrtil17 @poppy_haze

[Dar Sbitar @user_settings]

@MapQuest You're it, @mapquest!

[Dar Sbitar @user_settings]

@Hunters_iphone @DPRK_News First time I actually laugh reading a post on X :D

[Dar Sbitar @user_settings]

@chesscom You had to go and ruin it for most of us, didn't you?

[Dar Sbitar @user_settings]

@vxunderground Is this my email by any chance? (the long ass phishing .vbs file sent a while ago)

[Dar Sbitar @user_settings]

@osanseviero An E4B class model with shorter reasoning. The current one seems extremely verbose.

[Dar Sbitar @user_settings]

@vxunderground If it can help, I saw the article but didn't open it. I usually just read your tweets, even if they're long. I rarely (if ever) have the reflex of opening an X article.
Just a personal preference, maybe this is generalized?

[Dar Sbitar @user_settings]

@pintleinjector Please amend your prayer to clarify that it shouldn't be its *last flight* period. Because that would technically answer your prayer too.

[Dar Sbitar @user_settings]

@tpimentelms Because, in principle, a thorough review is what's supposed to dissuade against publishing poor work. And if by "lazy review" you mean quick rejection, the point still stands, but I see your point.

[Dar Sbitar @user_settings]

Got this in my timeline immediately after posting. His Grace String42069 provided the best illustration to my last point (not the babbling thing, the one above).

[Dar Sbitar @user_settings]

I dreamed that arxiv began requesting that an author's institution must have publications in reputable journals before accepting to publish a preprint. Anyway, just got another citation from a preprint where my paper is referenced for a concept that has nothing to do with it.

[Dar Sbitar @user_settings]

I'm in no way against using LLMs, but people should know that these systems are "garbage in, garbage out". Some effort is needed in their prompting and in the managing of their context to produce good outputs.

Please excuse my babble, just sharing some thoughts.

[Dar Sbitar @user_settings]

@BenjDicken @t_crayford Why are you calling him a tampermonkey? He just asked a question.

[Dar Sbitar @user_settings]

@yoavgo Exactly! Em-dashes aren't just an "LLM tell", they're a corner stone of LLM writing style. <-- there's this stupid construction too which drives me crazy more than em-dashes. No idea where it came from, but but both large and small new models over use it.

[Dar Sbitar @user_settings]

@shauseth How do you backprop with clay tablets?

[Dar Sbitar @user_settings]

@AstroJake In principle, it IS a good move. But I'd like more info on how they'll flag fake refs. How many 1-year bans for false positives are tolerable? Example: you can't always export citations to bib format, so mistakes WILL happen in conversion. Will these be flagged? @tdietterich?

[Dar Sbitar @user_settings]

(the model didn't learn time travel, I only messed up the logging)

[Dar Sbitar @user_settings]

> designed a good model for my application
> training data wasn't good enough though
> found a nice way to scale it
> didn't work
> found a fix
> this happened ...