"oh but have you tried opus 4.5" as a response to skeptics of llm utility falls flat imo bc of the fact that obviously if one is still uncertain about the utility of llms they will not have tried a model that they'd have to sign up for a 20$/mo subscription just to try once

ppl spend so much on shampoo it's ridiculous. u can just use dawn dish soap. it's good enough for the oil spill ducklings and it's good enough for your hair. u think ur better than a duckling? u think ur hair is more precious and delicate than their downy fluff? well it's not

tokens are such a weird metric. i have basically no intuition for how much text fits in "262144 tokens"

hmm when i plug nvidia orchestrator 8b into my basic agent thinking loop it just gets stuck repeating the same thoughts over and over. rip

looks like i can get up to about 32 tokens/sec (+/- 2) locally w/ qwen3 coder next q5kxl. i have yet to really get a sense for its capabilities tho

my numactl bullshit only made perf worse 😔

vrei să pleci dar nu mă, nu mă ctl numă, numă ctl numă, numă ctl ctl ctl

gave llama.cpp more threads and it got slower

what a lovely day to benchmark every conceivable combination of llama-server command line arguments

browser devtools should give me multiple consoles to write to

put yourself in my shoes. yes, good, now the socks too this time. excellent. pants as well. good, good, the transformation is well under way. now put yourself in my shirt. eeeeexcellent, the ponderfication is nearly complete. put on my contact lenses. hmm, we may need to adjust your hair color

every programming language is exactly the same

ugh python update broke this code everything apparently has a return value in the bytecode now wtf

my wretched little javscript homunculus wants to be rewritten in python

😭

sonnet 4.5 was being way more helpful than usual w/ coding today and then i realized i have extended thinking mode turned on and usually don't lol

ok ❤️ yay ❤️

imagine running out of chili powder. imagine running out of cumin. imagine running out of coriander. imagine running out of all three of those things at the same time

i am way too tired to rebuild llama.cpp rn

tool_choice: "required" actually makes it worse at tool calling bc it fails just as often but *when* it fails it does so by filling up the entire context w/ inexplicably failed tool calls where im not even sure what's malformed about the tool call xml it's putting out

task successfully simultaneously failed and succeeded, failingly

llm that's fine-tuned not to take "instructions" or "user and system and assistant messages" but just "input data and valid output schema"

system prompt begging the llm at length to please always use tools, even during refusals, noting that there is in fact a refusal tool, but best it can do is like. even odds of using a tool or not

glm 4.7 flash q4_k_xl seems extremely unreliable at actually using its tools and doing so properly w/ llama.cpp. maybe i just haven't found the right settings to enforce it idk

hmm. maybe the technology just isnt there yet

actually the most annoying thing about llms is they use 2 space indentation for js

should i make an agent or is that too hubristic even for me

annoying thing abt composing llm outputs is everything basically has a return type that's an undiscriminated union of "string that's of the form you actually want" and "total garbage string"

ok 2026 is here time to learn how tool calling works

when the blazor app hangs at startup while process memory quickly and inexorably climbs

would be nice if llama.cpp could just decide how much stuff it can offload onto the gpu by just kinda checking how much space there is

GLM-4.5-Air-UD-Q6_K_XL too big for my computer :(

ai democratizes art

"what model should i use on [hardware setup]" "it depends, best models for everyday use are glorpo-v2.5.4-32-79-4q-d3-quant-6-instruct & blip-3 but the latter might not be that performant, reports differ. but if u need better reasoning u might wanna go for antelope-3T-72moe-4bit. but then again, --"

we take a great deal of pride in our expertise and professionalism here at Le Epic Narwhal Bacon Residential HVAC Maintenance Co.

vibe coding (verbally walking another team member thru the changes they need to make to the code)

if you are experimenting with npm, do not do it on a company device. full stop if you have already run npm packages on a work device, treat it as a potential incident and engage your security team immediately

blog posts from software engineers talking about their experimental agent harnesses are so weird to me bc like. i can understand ppl spending their off time on that but why is a company paying you to slosh ungodly amounts of tokens around to produce unclear and uncertain results? my job would never

almost every time i have run into a failing test in the test suite of the codebase at work, the fix is that the test needs to be changed, not the code being tested

if you haven't personally scrutinized the code, it needs to a thorough test suite. however unfortunately if you haven't personally scrutinized the test suite, that's just as bad and you're right back where you started

my job should just give me a personal discretionary budget for tooling

vibecoding is like relying on UB in that in many cases it's fine

with the right mindset even 5:45 pm can be morning

gmbsky

":" as a build command

was gonna complain about wrangler dev server not accurately tracking live updates to files included in the static assets as symlinks but tbh i feel like it's probably fair that they didn't anticipate this use case

omg u can just post .svg files on here