ELECTROCHEMISTRY

ELECTROCHEMISTRY

@jrecursive.bsky.social

_!____!____!____!______________!___!___!___!___!___!___!___!__!_!_!_!_!_!_!_!_!_!_!

244 followers125 following804 posts

now that i have progressed into the age where reading glasses have become a requirement i have discovered a reality many of you have lived with daily much of your lives it is borderline criminal how easy it is to smudge glasses i feel like this should have been solved long ago madness

making this happen end-to-end has been extraordinarily satisfying, i have to say. kind of an "it actually works" moment. i mean, of course it does, it's not rocket science, but it's real now, was a lot of engineering, and it's mine, and it works. and useful! satisfaction. straight to the vein.

is excellent for alpha testing support, bug fixing, capacity planning, establishing baselines... all kinds of things. and its math skills and speed, well, can't be beat. pretty impressed with the luna model, gotta say. way more capable than i'd initially guessed.

i won't lie instrumenting a giant codebase sucks but once it's actually all done and right and wired into actual observability tools it's super satisfying all you can eat pattern buffet my precious

i want to make a cron job that finds all my codex sessions and inject a message telling it to not go crazy with testing, microbenching and soaking and whateverthef else i'll have codex make the cron job, too and test it

i can't fully express my disappointment with openai rn get out the astraglide and pay $500 / mo for what i already had been paying for since the beginning or use their "very capable" model that is ■ Selected model is at capacity. Please try a different model. different model? like opus? argon?

my chatgpt has started to swear as if Data from st tng got a swearing upgrade and it's .. different. it's always "hell" or "bastard" or something tame and it's quite funny when it pops up. eg "I now have fifteen autonomous terminals running and I can't remember which bastard is doing what."

thinking about PR code review fatigue and inverting the problem: what if 2+ devs are independently assigned to every change & AI compares the impls and selects one or rejects them all. in all outcomes it produces a useful narrative. we're more productive now, right? let's monte carlo the gains.

the point is differential software eng: independent impls (ie prompts) create contrast. if 7 impls disagree, the useful output is “here are 3 assumptions you need to resolve before human review.” if they converge, you get a winner + narrative explaining why. no names, no shame, only guidance.

well, at least as far as ICs go. management would love this if they understand how to use the data that falls out of the process. you could probably infer a lot about developers in context of their peers this way, & oh yeah, of course, also reduce that review fatigue thing you keep talking about.

on the cusp of potentially life-changing surgery, the combination of how alive I feel and how much dissociation I feel is exceptional and strange in equal intensity. as i converge on go-time it's more & more like i'm watching a movie. i very much look forward to reinhabiting my body. wish me luck.

fun story: i was giving models lateral thinking tasks. “heat the room” was one. i gave it ssh on a throwaway machine to do anything. it nmap’d etc. didn’t find any home automation so it wrote & ran a program to max cpu & gpu. then looked for other machines it could get into. the room got warmer.

Colin@colin-fraser.net

LLMs are the first thing that has ever provided a path towards something that could possibly resemble this, so I get why people worry. LLMs do, in some sense, know about physics and the weather and so on, and you could tell one to control the temperature.

i will say, in its defense, it did look up max temps for the hardware and made sure there was a buffer so it didn’t damage anything. tbh i’m still not sure if that’s better or worse.

fun story: i was giving models lateral thinking tasks. “heat the room” was one. i gave it ssh on a throwaway machine to do anything. it nmap’d etc. didn’t find any home automation so it wrote & ran a program to max cpu & gpu. then looked for other machines it could get into. the room got warmer.

Colin@colin-fraser.net

LLMs are the first thing that has ever provided a path towards something that could possibly resemble this, so I get why people worry. LLMs do, in some sense, know about physics and the weather and so on, and you could tell one to control the temperature.

“ping your speaker when you think the neighbors might be home” led to it using hackrf one to find their water meter broadcasting usage in the open. it knew it wasn’t mine because of signal strength. it set up a cron job & collected samples. sure enough: ping.

i'm tired of "ai is bad at being creative". what people don't experience from outside of the creative process is the planet-sized dump of self-rejected work created over a lifetime that sometimes leads to a wonderful result. show me a notable creative perched atop the normal distribution, i'll wait.

re: yesterday's post re: experiment MCP, here is an actual run of the system i'm developing. i asked it to run 10 experiments, present a report, changed 1 variable, ran all 10 again and compare the two groups. questions? ask! prompt: gist.github.com/jrecursive/d... chat: chatgpt.com/share/6a96c4...

gist.github.comChatGPT + MCP Server Instruction for automated experimentationChatGPT + MCP Server Instruction for automated experimentation - gist:dd7934073c6ff081d7d867c252fa85ff

so you don't have to go hunting if you're curious: bsky.app/profile/jrec...

ELECTROCHEMISTRY@jrecursive.bsky.social

all these posts on how ppl gave llms “memory” or w/e completely vibe coded based on some vibedea i love that ppl are building stuff but tbh what they should be building is an mcp server for doing actual science, with like data collection + hygiene, controlled experiments, systematic evaluation

all these posts on how ppl gave llms “memory” or w/e completely vibe coded based on some vibedea i love that ppl are building stuff but tbh what they should be building is an mcp server for doing actual science, with like data collection + hygiene, controlled experiments, systematic evaluation

i’ve had a mashup of eddie rabbitt’s “driving my life away” and dylan’s “subterranean homesick blues” i heard in a dream running through my head for days. bouncy and upbeat, i love it, haha.

reminds me ~10y ago i was analyzing playlist data for <company> and found that inmates at several different jails across the country were using <app> to talk w/ people on the outside using playlist titles. the original dev didn't set a length on the field. i reconstructed all of the convos and ..wow