I asked Claude for reading recommendations based on the new domain I find myself working in and it suggested two books I already own (unread) instead of new books to buy so AI is clearly anti-publishing industry this is bullshit
They should distribute people by generation - our team crushes until a zoomer topic
Happy stranger-wishing-you-felicitations-on-the-internet day!
Here's another one :)
my dad: people today are too sensitive
me: u bitches cried when dylan went electric
My internal scorecard is how many LOC did I remove or prevent being written
inevitable conclusion will be our users simply *imagine* entering data into a form... actually with 'agentic workflows' in enterprise this is an apt description lol
we desperately need a certain demographic to stop listening to podcasts, go pick up a book, do an online uni course on literature, or history, or anything
A SKELETEIN FLICKEN OFF THERE COMPUTER "THE ONLY GPT I CARE ABOUT IS GETTING HIGH AND PANICING BECAUSE I GOT TOO HIGH " - YOU WILL NEVER CONVINCE ME THAT A PROGRAM THAT BOILS SMALL TOWNS IN ORDER TO ASSEMBLE MEANINGLESS ENGLISH LANGUAGE SENTENCES IN A STATISTICAL MANNER VOID OF MEANING IS "INTELLIGENT" ANY MORE THAN MY FUCKEN APPLE 2 WAS INTELIGENT WHEN I LOADED UP ELIZA AND IT RESPONDED TO ME USING QUESTIONS MIXED WITH FRAGMENTS OF WHAT I TYPED TO IT EARLIER , HUMANS ARE DUMB AS FUCK BUT ONE THING YOU CAN GUARANTEE FOR SURE IS THAT COMPUTERS WILL ALWAYS BE DUMBER, WE SHOULD ALL FOLLOW THE EXAMPLE OF KURT RUSSEL IN DA MOVIE "THE THING" AND DESTROY EVERY COMPUTER WHO THINKS THERE BETTER THAN HUMANS, IF THE VENTURE CAPTALISTS LIKE THIS SHIT SO MUCH THEY CAN PUT THE DATA CENTERS AT THERE HOUSE AND TALK TO IT ALL DAY AND WE CAN SEIZE THERE WEALTH WHILE THERE BUSY GETTING COMPLIMENTED BY THERE PHONE - DASHARE.ZONE ADMIN
SHIT GPT - dashare.zone ADMIN
another day, another example of where reading the damn documentation cover to cover saved me, and Claude Opus did not
(Claude proceeded to create a spike of my hypothesis in 5min which would have taken me a day or two... which also saved me)
So much about living life is getting harder because we now have to jump through hoops in systems that were designed by technologists without a very good grasp of the humanities.
People tuned into that aren't going to be very excited about stuff which makes coding easier! The coding's not the issue!
Aperture being discontinued is up there with killing Google Reader
I started my photography career with Aperture, and grumpily had to switch to Lightroom and Capture One once I started working in studios
cynic in me thinks these language/framework docos are a bit naff
BUT I actually believe they are important. The history of the present is often overlooked, and the medium of video is the right one for recording that history today
looking forward to this
actually I have radicalised myself
DELETE ALL THE TESTS
tests created by the LLM when it was doing feature implementation cannot be trusted and have no intrinsic value. the only way to make them valuable is to read and reason about them yourself
windows terminal (which amazingly is just about the best terminal I've ever used), arch linux with ghostty, macOS with ghostty
every other tui I use in every setup (nvim, tmux, lazygit, fzf , opencode et al) has zero input lag. computers are fast
just found out why lmao
remove the third frame this is goes incredibly hard
Google: "read them charlottes web"
just reminded me my mum read that book to her fifth grade class and another teacher had to come finish it for her
Everything lags (typing, invoking slash commands) - they made a tui like a gui
The input field moves around!
Scrollback is whack
claude code cli clearly built by people who don't spend significant time in the terminal
if it sucks... hit da bricks
Yea it fucking sucks (derogatory)
An actual transformative LLM usecase is meeting transcription to minutes pipeline.
I’m co-architecting a pretty hairy service and spend 6h workshopping over a couple days. tonnes of tangents. loads of unresolved questions.
in a day or two I would’ve forgotten half the things we discussed
They wouldn’t make it to the destination honestly
watching what LLMs execute has actually improved my bash skills. many “oh **thats** how you do that” moments
lovely post - when I went to mona a looong time ago there was a massive Buddha statue made of compressed ash from temple incense and it was falling apart by force of air changes from people walking through and it was just one thing I'll never forget
experience report: Java is awful
its also insanely token inefficient compared to clojure
here's how your company is rotting right this moment:
- your senior devs stopped writing code
- they ask Claude to generate it, they check that it mostly works, they ask a junior to approve the new PR
as well. My assumption is that when writing new code the weight of the training data is the driving force (and as you've noted the bulk is bog-standard imperative style)
But creating a plan to refactor does work, because its working through specific anti-patterns one by one.
Well-published CS authors are already in the training data. So if for example Ousterhout's writing really influenced you, you can ask Claude to critique existing code with that lens and come up with a refactoring plan.
Adhering to a style at original generation time however doesn't seem to work /n
I feel my biggest risk right now is the sheer volume of unit tests LLMs create (600 tests around a moderately sized module). So many I can't actually reason about correctness
resisting the urge to delete all the tests - at least if there's no tests then there's no false sense of safety!
Even when doing clojure/lisp they drift to imperative and you have to constantly course correct
