@xssfox@cloudisland.nz
she/her
xssfox or crossy for short, is a Naarm (Melbourne Australia) based vixen who enjoys hacking on cloud infrastructure, radio hardware and software. She specializes in cursed software and horrible infrastructure. Notable works include BigBuckAWS, nr-basic, sondehub.org, ft8_beep and etch-a-tv.
Replying to @xssfox@cloudisland.nz
Tiers
"We just made a resource for each of our existing rest API endpoints"
"Single resource to manage everything, and you may only have exactly one"
"Looks fine on paper but even the most trivial real world use case requires local-exec"
"Resources have a single content attribute that's a JSON string"
"No way to not hard code/ commit credentials"
Terraform provider tier list:
Cat: why would I want an out of box experience?
Next api I build, I'm going to choose violence. Requesting an object by ID will require a string not a number. You'll have to use the spoken word representation.
{"id": "forty two thousand, six hundred and twenty two"}
Replying to @camwilson@mastodon.social
@camwilson I can't see the article but tasmania is currently consuming 1.5 gigawatt including 300MW(?) aluminum smelter
Seems the latest LLM tell is seeing "why this matters"
Replying to @steadilyebbing@girlcock.club
@steadilyebbing *hugs*
Replying to @daisy@cloudisland.nz
@daisy "Login to the website to see if there's a notification about a bear shitting in the woods"
I wanted a way to visualize how enabling non deterministic output on LLMs (temperature and what you find on most public services) can have influence on the output. Being able to see "how likely was this to choose a different number and if so, what numbers" is a good way to judge trust (lots of asterisks here)
This little script takes a prompt, and then any time a number is encountered in the output reports the chances^1 of other numbers being reported, the continues on.
I'm just running a model that I can fit in my memory (qwen/qwen3.5-9b) so this isn't a demonstration of LLMs quality in general, just the concept that we could use to show/inform people of confidence(lots more asterisks here)
^1 This is using LLM Studios default Top P Sampling at 0.95
@firstyear billion fucks
@wren6991 @whitequark I've been thinking recently that nondeterministic token selection in public LLM services is part of a "Psychic Sports Picks"-like scheme^1. The randomness means that enough people get a good result to sustain the hype.
I've wanted to build a little toy LLM website where whenever a number is produced as outputs it let's you hoverover to see the other output options were possibly/likely.
@arrjay I think that removing it is probably a good idea but in general basing removal/additions based on percentages of usages is a terrible metric because a small percent of a large number of users is still a lot of users
I do wonder if there are more Linux ax25 users when it was removed than when it was originally put in
@decryption it's enforced after 200 cycles now as well https://www.pcguide.com/news/google-confirms-pixel-10-limits-battery-capacity-after-200-charging-cycles-and-you-cant-do-anything-about-it/
PC GuideGoogle confirms Pixel 10 limits battery capacity after 200 charging cycles, and you can't do anything about itGoogle Pixel 10 supports Battery health assistance, but unlike previous generations, you can't turn off this battery throttling feature.Found these on discount at Boggis, Bunce and Bean supermarket.
Life is like a box of chocolates after shrinkflation