lexi

my only real gripe with coder-next so far is that it occasionally has a fit of pique and tries to nuke entire files because of a compile error. had to set an explicit permission requirement for rm

but how will I know if 30-50 feral hogs break into the pantry

no satire is as hamhandedly unsubtle as real life. Aesop himself would keel over from sheer cringe

say what you will about carcinogens or whatever, but when you turn the corner and smell a campfire, someone burning birch or fir in the winter-spring sunset breeze, gosh if that sweet scent doesn't hit you right in your mammalian core

seconded! i've been wondering about this a lot

that artificial contained parasitic fungus spewing out spores into the cosmos is entirely american politics

i'm particularly curious if there are discernible trends on the lines of sex, generation, or profession

wondering what kind of variation we see in terms of how people would prefer to interact with an AI. how many of us want our bots to have distinct personalities? which human behaviors do we want them to imitate? do we want them to have persistent long-term memory, or ephemeral minds? etc

why is everyone getting so hot and bothered about the Superb Owl

i get it, it's superb, but like. it's still an owl

maybe with qwen-coder-next i'll finally be able to build an SLD resolver library

I really like that I can just give the thing instructions, go to sleep, and wake up in the morning with finished code

coder-next confirmed for ASI

almost nothing that uses the chat completions API ever seems willing to report prompt processing progress. the llama-cpp built-in client does at least try these but it always goes directly from 0% to 100% so the information content is nil. aggravating

"captain! the periodic oscillations -- they're dephasing!"
"they must have been perturbed -- but how?"
"i don't know, ma'am, but if we can identify an excitation frequency region in the microtubule nanowires, i can reconfigure the deflector dish to break the time symmetry harmoni

This is the future liberals want.

many such cases

trying a thing

yeah the default fitting algorithm is bizarre. if it can't fit the entire file in memory it immediately goes for the jugular and cuts context length down to 4096, before it starts trying offloads. i don't know why anyone thought that was a good idea. i still p much have to tune everything by hand

well i was able to get tool calls working by overriding the chat template with this abomination github.com/ggml-org/lla...

oh for pity's sake

apparently in the autoparser branch they completely axed support for lfm2-style function calls and hacked around them with, this

OH

no i see the problem

lfm2-vl-3b apparently doesn't support tool calls in the first place

9b!!!

lcpp is never going to support this, is it 🙃

omg huggingface.co/openbmb/Mini...

apparently something broke lfm2-3b tool calling

no idea if it was in master or in one of all the PRs i've pulled 🙃

current status: debugging opencode with wireshark

the Warp is empty and all the daemons are here

This is, however, a perfect paragraph

literally since i was a kid i've been relying on `sudo echo` or `sudo -v` followed `sudo do-thing &disown`

somehow it never occurred to me until now that `sudo` might have a better way of handling this

well today i finally learned about the `sudo` `-b` flag and i feel like such a fucking dumbass

oh HELL YES

it's mostly a vibes thing but Anthropic kinda scares me in a way ClosedAI's tepid corporate floundering doesn't

better late than never!! any particular features you're looking forward to?