kida2002.bsky.social

kida2002.bsky.social

@kida2002.bsky.social

I am not a robot, I just don't post that much

105 followers323 following397 posts

kida2002.bsky.social reposted

“babe what’s posthog? what’s datadog? What’s firecrawl? what’s what’s gnomefuckr? what’s cuckdb? what’s slutcon?” do you have any idea how humiliating being a software engineer is

kida2002.bsky.social reposted

i actually don't think most people can imagine a better future. i can, but i'm built different.

hailey@hailey.at

we can imagine a better future everyone. its possible. and frankly, its going to be needed!

The velocity claims are less interesting to me than seeing a nice validation of some engineering ideas pushed to scale wrt verification & bug-repo tools being extremely powerful, software architecture still mattering quite a lot, and low-hanging fruit like lint rules and static analysis being good

Oh I guess the author is also on bsky: @no.lol - this was a great presentation, thanks! It would be really great if you could cross-publish the talk on Youtube or another platform if you wouldn't mind please

kida2002.bsky.social reposted

The TTC is finally considering increasing streetcar stop spacing, active gap supervision on subway, and introducing mid-route timing points on surface network. Please consider filling this out to help bring TTC operations into the 21st century.

surveymonkey.comTTC Service Standards Review - Round One SurveyWe are reviewing and updating our Service Standards, and we want your feedback on the options we’re exploring. Learn about the options by taking this survey.

Taking a five hour train ride, this is so dope so far. It'd be nice if the train was a real high-speed line but still a very nice experience so far. I'm glad I didn't elect to fly instead this once since I had the extra time to spare.

I find it funny how much Canadian folk music resonates with me as an immigrant over a shared reluctance to say goodbye to our homes and move west for better opportunities. I *also* hate cowboy clothes, Stan

For a while I think the meme was that the median Google engineer was a lazy coaster, while the median netflix/meta/amazon engineer was a cutthroat amoral mercenary. Excited that OpenAI is creating a new meme category of bumbling incompetent boobs who simultaneously try way too hard

I find it difficult to keep straight in my mind that agentic tools are both designed and capable of completing tasks only loosely specified in natural language and also that there are huge differences in success rate depending on how the tools are used.

Claude keeps asking me recently if I want to "spike" new features that I have been sketching out - I've usually only seen the models echo slang of similar tone/era back to me, this is the first time I've seen the model start to use slang I didn't recognize unprompted. How odd.

This was Opus 4.8 [1m], nothing new there on vanilla claude code. Long running multi-turn session, exploring code-base and iterating on a plan for a feature - nothing particularly notable in prompts. Asking if they should proceed is pretty standard, but the unprompted slang usage is novel (to me)

Claude code web is really, really close to exactly what I want for 90% of tasks and would almost obviate my custom VPS+tmux+docker session harness, but still feels incomplete - I wish there was a way to better customize the environment (ideally even being able to just give them a docker image)

I picked up and worked through the little learner recently by Friedman & Mendhekar. It was really fun! I really liked "the little schemer from Friedman and its very much in the same vein. I wonder what a good follow-up text might be. Considering diving into Chollet, but UDL by Prince also looks good

I find giving cc+opus optimization and profiling tasks and pprof/strace instructions still works rather poorly if left fully unattended. Lots of time spent in unproductive directions before, elementary mistakes in reasoning traces around interactions with ulimits, gomemlimit, RSS, etc. Skill issue?

I check in on it occasionally and the auto-research-esque harness I jerry-rigged is able to invalidate concrete hypotheses, so with occasional steering, the harness does what I want, it's just a lot of time and token spend.