Mastodon Feed: Posts

Boosted by jwz:
zzt@mas.to ("[object Object]") wrote:

extremely upvoted hacker news comment about how actually if you think about it it’s extremely kind of google to tell you it’s raining when they piss on you

dysfun@treehouse.systems ("gaytabase") wrote:

@glyph it is however the future of comedy. i've just spent 3 days getting it sandboxed on my local machine and i'm unimpressed with the payoff.

Boosted by glyph ("Glyph"):
abucci@buc.ci ("Anthony") wrote:

@davidgerard@circumstances.run @xgranade@wandering.shop @mttaggart@infosec.exchange @cwebber@social.coop

Incidentally, as a bit of an aside since it touches on my own CS research a bit:

high output generation with review

seems to be load bearing in the push to widely deploy vibe coding or agentic coding or whatever they're calling it today. "It's OK to have LLMs produce code as long as there is thorough human review" is an argument I've seen trotted out countless times, but should not be given any credence.

I am here to tell you that this is a misapprehension that ignores the substantive difference between:

(a) Competent human beings producing X by, in part, avoiding producing Y
(b) Less-competent human beings with machine help producing Y, catching that Y is bad with a test or review, and lather-rinse-repeating until X is produced.

If you like, (a) is gradient ascent while (b) is trial-and-error (generate-and-test, or hillclimbing, to use the GOFAI jargon). Everyone who works with such algorithms knows that (a) is many orders of magnitude faster, more reliable, and more robust than (b) when a good gradient is available. Most of machine learning is based on this observation! When it comes to producing code, competent human software developers provide such a gradient (that's what we mean when we think of them as "competent").

glyph ("Glyph") wrote:

RE: https://mastodon.social/@glyph/116546111998207690

It occurs to me that I don't actually understand how model-context protocol works, even at a high level. Given that emitting structured requests for tool use was like, a new thing, not present in the big training datasets, how do LLMs' output result in tool invocations?

Boosted by glyph ("Glyph"):
grimalkina ("Cat Hicks") wrote:

Perhaps you have thought to yourself, *I'm* on board the Psychology of Software Teams train, but how can I subtly let others know? How can I identify fellow POST fans in the wild?? What if we need to organize around the radical notions of thriving, motivation, self-compassion and psychological safety??

Well if you find me at a 2026 conference or book reading you can get a TINY PROPAGANDA STICKER

Boosted by baldur@toot.cafe ("Baldur Bjarnason"):
hejsna@ruby.social ("Johan Halse") wrote:

React apps are, overwhelmingly, just straight-up worse than plain old server-rendered markup. I don’t understand why everyone keeps building with it. The DX isn’t even that good.

Boosted by glyph ("Glyph"):
ericjames@booping.synth.download ("ericjames* 🐺‼️ (they/it)") wrote:

stolen for alt text

dysfun@treehouse.systems ("gaytabase") wrote:

well gemma was enormous fun, next up i have to try a qwen model. if you'll excuse me, i now have to go find a good one and then wait for it to download.

dysfun@treehouse.systems ("gaytabase") wrote:

the french government should make their own fork of firefox and call it firepoodle.

but they can't because they wouldn't let the fire muck up their pretty doggo.

glyph ("Glyph") wrote:

it’s kind of fascinating watching the vote count jump as this hits pockets of new networks, and yet the proportion has not moved at all since it got a few hundred votes

glyph ("Glyph") wrote:

(nevertheless please boost the heck out of this, the more it breaks containment the more I get a broader sense of how folks are feeling)

glyph ("Glyph") wrote:

I appreciate everyone voting but this is definitely one of those posts where I find myself wishing I got a dime for each reply https://mastodon.social/@glyph/116542583024883535

dysfun@treehouse.systems ("gaytabase") wrote:

it never found the answer. it gaslit me repeatedly about it, then i found something that looked odd ($SHELL not being set because i set the user up by hand and didn't copy over /etc/skel) and when i set it, it worked. i told it that it had to be something else and offered that i'd just noticed $SHELL wasn't set.

it gleefully told me it absolutely definitely wasn't this and that the only solution was to execute the full path. and also that it was a bug in busybox's PATH logic.

dysfun@treehouse.systems ("gaytabase") wrote:

i thought i'd start it off with an easy task - figuring out why my attempt to install it in ~/bin isn't working.

it's very special. i got this after several messages of it gaslighting me that it was my PATH (amazingly, i did actually do everything i could think of to get it working before i asked it...) https://gist.github.com/jjl/2c3895aa1bf0ca74f146f1588c804f19

Boosted by dysfun@treehouse.systems ("gaytabase"):
ricci@discuss.systems ("Rob Ricci") wrote:

#3dprinting question: my experience is that filament spools that are tidily wound (ie. nicely packed layers, tight side by side rows) tangle much more easily than the ones that *appear* sloppily wound

Am I imagining this or do others have this experience too?

Boosted by soatok@furry.engineer ("Soatok Dreamseeker"):
MonotoneofBill@mastodon.world ("M❍n❍t❍ne❍fBill™") wrote:

Will transparent coffins catch on? Remains to be seen.

Boosted by soatok@furry.engineer ("Soatok Dreamseeker"):
erkhyan@yiff.life ("Erkhyan") wrote:

All right, I might as well announce it.

Even if it’s just a limited-audience, one-day event, as of 2026-05-16, I’ll have reached the art milestone of “got a piece of mine into an actual exposition”.

dysfun@treehouse.systems ("gaytabase") wrote:

it's okay, i've taught it. i am confident it will now correctly interpret my sarcasm

https://gist.github.com/jjl/98a544e8e67895b3bd6ce852f816446d

dysfun@treehouse.systems ("gaytabase") wrote:

oh dear, it doesn't understand british english 😂 https://gist.github.com/jjl/16c6fd0cb74dfe5eee4f64f703ab012c

dysfun@treehouse.systems ("gaytabase") wrote:

look i'm sorry but if you genuinely believe this is a serious competitor to humans thinking about stuff you are beyond help.

dysfun@treehouse.systems ("gaytabase") wrote:

LOL look at this "helpful" shell script it wrote me https://gist.github.com/jjl/985a07ab2dfd090afc60b79b01e0a408

chipotle@mstdn.social ("Watts Martin") wrote:

This flu, or perhaps Covid that has managed to return three negative tests, has really messed with my appetite. You know what? I wish I still had some damn Soylent. (I’ve ordered vegan Ensure that’ll get here tomorrow morning.)

Boosted by glyph ("Glyph"):
glyph ("Glyph") wrote:

Is agentic coding the future of software development?

Boosted by soatok@furry.engineer ("Soatok Dreamseeker"):
AutumnWyvern@dragon.style ("Autumn :AutBabie:") wrote:

@saphire

what if someone doesnt have a smartphone..? or is currently using their phone to access something with this new captcha and doesnt have a random second device to scan QR codes off of their phone with?
even in a sunshine rainbows world where nobody did scam QRs it seems like a pretty bad idea...

Boosted by soatok@furry.engineer ("Soatok Dreamseeker"):
saphire@dragon.style ("Saphire Lattice") wrote:

Well, the new Google ReCaptcha is awful, sheesh

It's a QR code you have to scan with a "proper" device - aka with Google Services installed

Goodbye last 10 years of phishing awareness, time to scan random QRs without a thought while you are purchasing things, woo! Seriously what were they thinking?

And because it's recommended to be put in "high risk" places, people will expect them to be seen there, and so a scam/phishing QR will be so much easier to slip in.

https://cloud.google.com/blog/products/identity-security/introducing-google-cloud-fraud-defense-the-next-evolution-of-recaptcha/

#google #captcha #recaptcha #phishing #infosec #cybersecurity

baldur@toot.cafe ("Baldur Bjarnason") wrote:

Oystercatcher! On a lawn! #photo #bird #birds #iceland

dysfun@treehouse.systems ("gaytabase") wrote:

my god, look at this mess https://gist.github.com/jjl/c44f8b4f565861e29415c61d12b7b4f0

baldur@toot.cafe ("Baldur Bjarnason") wrote:

This male northern wheatear was hanging out at a construction site this week. #photo #bird #birds #iceland

dysfun@treehouse.systems ("gaytabase") wrote:

on the other hand, my token rate is much slower when i do it through llama.cpp's ui myself. presumably this is because they have filled the context buffer with some instructions for agents to ignore.

dysfun@treehouse.systems ("gaytabase") wrote:

holy crap, it's taken 3 days but i've got a sandboxed slopcoding setup running.

truly this is the future, what with me needing to rely on [redacted] years of knowledge to get here.