@jmoiron.bsky.social replied: Thinking that a text has an inherent meaning that is context free and ideologically neutral is a classic fallacy. His work is apologetics for white nationalists.
jmoiron plays the blues
@jmoiron.bsky.social Sad to hear that Shinichiro Sato, the drummer of "The Pillows" has passed away.
When I think of Sato, I think of "Funny Bunny", one of their best songs. But I also think about cornering him and Manabe outside Pianos in NYC and asking them to play Swanky Street at their next show...
@jmoiron.bsky.social replied: A lot of this data is made public already, your staffers could certainly understand + summarize it.
Google eg. publishes fairly good sustainability transparency reports which show a clear progression over time: datacenters.google/efficiency/
Or, my summary:
jmoiron.net/blog/on-...
jmoiron/giverny main bf34e08 add side-nav which lets you navigate between boards much more efficiently
@jmoiron.bsky.social replied: The people who are getting the best results have incredibly tight correctness guardrails and are parallelizing the maintenance process, essentially adding a multiplier on top of that 1.5x maintenance advantage, letting their testing controls prune invalid or unfavored results.
@jmoiron.bsky.social replied: A vital skill that good software engineers develop over time is knowing when a pattern is real and affords you leverage to make material organizational improvements and when a pattern is weak or coincidental and extracting it just makes things worse.
LLMs are really poor at this...
@jmoiron.bsky.social replied: I'd like to explore why this happens in a bit more detail, and maybe I will if I get back to that post, but at its core, it comes down to an LLMs "preference" for generation over consolidation compounding over time.
Which isn't surprising. They are generative models.
@jmoiron.bsky.social replied: But LLM code generation is kind of the inverse of that old 3 step owl drawing process. It starts by producing you a nearly finished Owl, and then you spend the rest of your time struggling to get it to fix one thing without breaking another.
@jmoiron.bsky.social replied: It's well known that programmers all think they're at least 2-5x better than they are, that's why our estimates are always crap.
But using an LLM you really can produce some decent software in a proverbial weekend. I "wrote" my own Kanban board over the course of ~3 days this we...
@jmoiron.bsky.social replied: A lot of people are rightfully asking, "If LLMs make software development so much faster, where's all the great software?"
I think the answer is that LLMs are a 10x improvement on building software but only ~1.5x on maintaining it, and most software development is maintenance.
@jmoiron.bsky.social replied: I have a half-written post about how LLMs are great producers and good summarizers, but they've are poor organizers. This might be a reflection on ourselves, given how they're trained.
This can all feel contradictory initially.
@jmoiron.bsky.social Another example of what I wrote about in "Wirth's Revenge."
Your first prompt for an unfamiliar problem should be "explain 3 ways to solve this problem."
People imagine these simple data manipulations are easy for LLMs, but they lack the fine motor skills to do this type of wor...
jmoiron/giverny main 6933660 fix bug where drag-clicking from inside a modal and mousedown outside a modal would close the modal
jmoiron/giverny main b79226d unify color pickers for labels and cards, first label sets card color, better random label color selection