davidcrespo.bsky.social
web dev + hot dad. enjoy charts, unions, conputer games, philosophy. chicago crespo.business
2,523 posts
912 followers
488 following
Regular Contributor
Active Commenter
comment in response to
post
I had missed that he was at least allowed to hold his baby a month ago. good lord
comment in response to
post
any argument against the possibility of judgment is self-undermining because if it's true then there's no way anyone could judge that the argument is good
comment in response to
post
the followup distillation makes clear how the concept of bias is being utterly abused
comment in response to
post
from a selfish POV, something I'm more worried about than code quality is the erosion of software engineering as a hard-won distinct discipline with corporate orgs. but it is not at all clear to me that the status quo is globally optimal, it's just good for engineers
comment in response to
post
this is also why I don't really buy "vibe coding is payday loans for technical debt" — I have such a dim view of existing code that I have a hard time imagining LLMs making it worse. I think the trajectory we're on is that LLMs take a wider view of our codebases and actually make some decent choices
comment in response to
post
one reason I might have an idiosyncratic POV on junior eng productivity is that I suspect that all software engineers are building an at least halfway wrong thing at least half the time, regardless of experience level. so far, LLMs don't help with this, thought I don't rule out that they could soon
comment in response to
post
a cause for uncertainty: big difference between carefully selected hires mentored by great mentors and whatever must be the average quality of hires and mentors
comment in response to
post
more likely to me than the below: people are incorrectly interpreting a short-term hiring slowdown (holding pattern as LLMs shake out) as a secular trend. labor demand is a continuum across experience levels. eventually we will have so much LLM-produced software that we need more engineers again
comment in response to
post
everyone seems to think it is! as in, it seems to me that the only way everyone stops hiring low-experience engineers (as a secular trend rather than a short term downturn) is if they cannot do anything worth their salaries
comment in response to
post
very unfair to Fanon to be cited by a guy claiming that knowledge of any kind is impossible
comment in response to
post
comment in response to
post
it's clearly true that chatgpt is not sentient in any familiar sense. the problem for Bender is that it is now equally clear that that doesn't matter!
comment in response to
post
more pointedly: people are talking about agentic editing (zed, cursor, claude code, aider) but I still do the vast majority of my LLM use in through piping files and diffs to a CLI to ask questions about it or fill in a test I've stubbed out. the barrier to entry (for a coder) could not be lower
comment in response to
post
there are quite easy ways to use local models (I don't bother, but you can). just use ollama.com and point an OpenAI-compatible CLI client at it. you can also point agentic editing CLIs like Aider at it. qwen3 32b does well on aider's benchmark aider.chat/docs/leaderb...
comment in response to
post
the button actions feel a bit Universal Paperclips
comment in response to
post
very fitting reaction given that they made it an Uncharted game
comment in response to
post
it falls apart the second you make it concrete. oh, so one prompt costs how many picoseconds of showering?
comment in response to
post
I spy a balance bike mom
comment in response to
post
what I would REALLY like to see is an empirical study of how a wide range of people are actually prompting these models. that would be incredible
comment in response to
post
good framing. I do think prompting resembles communication, even when the output isn't intended as communication. certainly the more you approach prompting like you would writing, the better you are at it. and the better the model, the more subtle cues it can take
comment in response to
post
right, the failure to sell it (to the public, not the real left, which is tiny) was his failure!
comment in response to
post
butterfly meme, wapo pointing at a dumpster. "is this [substack's value prop]?"
comment in response to
post
literally just "we want to be substack". yet it's unclear if substack makes any money!
comment in response to
post
seems clear chatgpt still has this problem under certain conditions, though!
comment in response to
post
I think people assume this voice is intrinsic to LLMs because it's so common, but OpenAI really do have to actively make them talk that way, between post-training with human feedback and the chat UI's system prompt. which is, if anything, worse than if it was like that on its own
comment in response to
post
or maybe it thinks it can read links because they tell it about how it's hooked up to web search and other tools
comment in response to
post
I wonder if the prompting in the chat UI makes it more eager to please. I get clear "I can't follow links" responses from everybody
comment in response to
post
this post by @stdlib.bsky.social is the best thing I've read about why we must stack