On math or programming problems, they are overfit to solving the entire thing end to end (presumably for benchmarks). I have had very poor results asking for pointers and hints that don't give away key insights. This has been the case across models I have tested.
An architecture with a "judge" that gates responses and ensures a lack of spoilers would probably work better. But this is a simple thing that they keep messing up.
I haven't had this experience at all. I've used Cursor+Opus on homework e.g. to understand algorithms, but I usually prompt it with something like "DO NOT GIVE ME THE ANSWER, I care about understanding and solving this myself".
If you use Firefox (and will pardon me plugging my own project), you might be interested in a version of this I made a while ago as a browser extension:
My project is somewhat different from the userscript in that mine uses Bloom filters to check whether a link has been submitted, whereas this userscript sends every page you visit to the Algolia search API. Using Bloom filters means that there are no external requests unless you actually press the button. My project also doesn't show the comments in a sidebar.
Having a link to past discussions, whether using this userscript or my extension, is useful not only to read the comments for a page you're browsing, but also to find good sites that haven't been submitted yet!
I like the bloom filter idea - I agree I don't want every URL I visit to be sent to Algolia. I really like this project's UI though - automatically opening it in a pane so I don't have to leave the site to see comments is really nice.
Not the author, but this prevents leaking all URLs you are visiting to agolia. Instead of asking Was this URL submitted to Hacker News? you ask Which URLs have been submitted to Hecker News?, store that in a Bloom filter, and then use the Bloom filter to answer the question Was this URL [likely] submitted to Hacker News?
Is the story being submitted a common feature like dupe detection or maybe unsubmitted link detection?
How does the filter know anyways?
If it is provided by the package, one could imagine adversarial scenarios where "high interest" submissions are excluded from the filter(maybe even make a custom one each time that somehow lands them in a collision) so Algolia knows exactly what it wants to know about you.
Probably, I should have read more into both of the projects before asking.
Wow, I'm seriously impressed! Questions for the author:
I haven't had a chance to read the code yet, but how does this actually adjust the text? I get that it uses Knuth-Plass to calculate layout, but what does it do in CSS (or SVG?) to apply the layout to the DOM?
Looks like the first commit was sometime last week, which is pretty tight turnaround for something like this. For my own sanity when reading the code (no judgment either way), did you use LLMs to help write this? Knowing is helpful for me to calibrate what I pay attention to when reading through the code.
The layout is done entirely in DOM. The layout engine calculates the line breaks, tracking, expansion, etc. then applies the layout using HTML spans with the appropriate attributes applied. There’s generally one span per line, but that can change for more complicated lines! I invite you to inspect the demo page in your browser’s dev console if you’re particularly curious.
And yes, I did use LLMs for the implementation, both Fable 5 and 5.6 Sol specifically. I would not have been able to achieve anywhere close to the current result unaided. My patience and expertise would’ve been exhausted long before getting anywhere especially interesting.
My philosophy with LLMs is to try to use them to achieve levels of quality and polish that would’ve been previously unreasonable. I hope this project is a demonstration of that.
> My philosophy with LLMs is to try to use them to achieve levels of quality and polish that would’ve been previously unreasonable. I hope this project is a demonstration of that.
This philosophy has been growing on me - a coworker framed it as “nuance melts away with LLMs”
I recently wanted to render a 100+ page document in an interactive WYSIWYG editor using a single stroke font and join the letters to form single stroke cursive words that appeared handwritten.
The permutations of rulesets that governed connections between letters (I.e. how does t connect to c, and do you need to lift the pen to make that connection) amplified by the randomness injected into the strokes and line width/angle (handwriting doesn’t have uniform letters or lines!) would have been intractable previously.
But, with an agent harness, I could do 3 or 4 examples of what “right” looked like - give it the ability to generate and screenshot test cases to see the output of its work - and it could grind through a reasonable ruleset for about $20 in spend.
I agree with the spirit of what you're saying: that many aspects of this post seem unnecessary. But I do think there are reasonable answers to your (admittedly probably rhetorical) questions.
Zig is a relatively young language with a small community, and Oven/Bun is one of few places that someone could previously have written Zig code professionally. It's therefore Zig's business to make sure that either it's a good place to refer community members for work, or that they don't explicitly encourage people to work there. Likewise, as one of the highest-profile Zig projects, the community's leaders were understandably invested in making sure it represented the language well.
I feel like I am exactly the target audience for this post: someone who uses Zig regularly, but hasn't touched Bun, despite being aware of it. While I would have proceeded differently than Andrew Kelley here in terms of framing and phrasing (and leaving out some parts entirely), I do think reading this gave me new information about Zig's relationship to Bun. The specific dry, professional post you suggest wouldn't have given me any new information at all.
The best exploration of this that I have seen in media is one of my favorite
movies: Nightcrawler (2014), starring Jake Gyllenhaal. The movie doesn't touch
on the government/democracy aspect of the article, but it very much captures
the notion that desparate people can be pressured to do horrible things when
their job is at stake.
In Nightcrawler, some characters are trying to get ahead, and others are
desperate not to fall behind, but their opportunism (driven by the necessity to
make money in order to survive in our capitalist society) makes all of them
vulnerable to exploitation by an ambitious psychopath. In that case, he is
profit-motivated, whereas the article here is about dictators retaining power,
but the same principles apply. The movie does an amazing job of exploring how
these individuals can wield power irresponsibly, poison everyone who gives them
an inch, and sound almost reasonable while they do it. It is a masterful
portrayal of how much some people can be willing to compromise on their morals
for their job.
If you haven't seen it, you should watch it. If you have seen it, but don't
remember it being deeply critical of capitalist society, you should re-watch
it. (It's easy to get so engrossed by the truly suspenseful and thrilling
moment-to-moment action that you miss the big picture.) The deterioration of
American news media is a more overt theme in the movie, but in my opinion, that
serves as a complementary backdrop to the anticapitalist message, which is the
engine that drives the movie inexorably onward. Also the acting, directing, and
writing are great.
Don't spoil it by reading the plot summary, just watch it.
I like currying because it's fun and cool, but found myself nodding along throughout the whole article. I've taken for granted that declaring and using curried functions with nice associativity (i.e., avoiding lots of parentheses) is as ergonomic as partial application syntax gets, but I'm glad to have that assumption challenged.
The "hole" syntax for partial application with dollar signs is a really creative alternative that seems much nicer. Does anyone know of any languages that actually do it that way? I'd love to try it out and see if it's actually nicer in practice.
Clojure CL as well have macros that let you thread results from call to call, but you could argue that's cheating because of how flexible Lisp syntax is.
Clojure also has the anonymous function syntax with #(foo a b %) where you essentially get exactly this hole functionality (but with % instead of $). Additionally there’s partial that does partial application, so you could also do (partial foo a b).
Thanks! I also told Aga via email in the thread where I submitted my article.
Worth noting that the HTML tag in the title was stripped from the PDF table of contents as well, so the title for that article in the contents is missing a word. No big deal, but good to know for future submissions!
Wow, this is jam-packed with interesting information. Thanks for writing it! (Also thanks for all of your other great open source work!)
Are there plans to upstream this into the Zig std library? Seems like it could be useful for more than just the cryptography package, since the benchmarks at the end have it often being faster than std pdqsort. I just checked the issue trackers on Codeberg and GitHub, and didn't see anything mentioning djbsort or NTRU Prime, which leads me to believe there aren't (official) plans to upstream this (yet).
pdqsort is a generic comparison sort. Want to sort employee names, customer email addresses, JSON blobs, or Zebras? No problem, pdqsort just needs an ordering, in both Zig and C++ you write this as a single boolean "less" predicate.
DJB's speed-up relies on vectorization, which works great for integers or things you can squint at and see an integer - but obviously can't sort your employee names, customer email addresses, JSON blobs or Zebras. You could write these branchless network designs anyway but I'm pretty sure they'd be markedly slower, at least for some common inputs.
It’s different every time, but basically “marketing”. No matter where you are showing your stuff, it’s in a subset of the population, chances are HN won’t be buying your app subscription. You need to get it in front of your actual audience.
As someone who has a small bit of skill in marketing nowadays, where would you advice me to deeply learn it? I think I have enough knowledge now to have developed a taste of what is actually good but don't know where to find a good curriculum to get started.
See also "Product Hunt". Oddly it's been about a year since I've noticed anybody who mistakes a Product Hunt launch for a marketing plan but that used to be endemic.
On math or programming problems, they are overfit to solving the entire thing end to end (presumably for benchmarks). I have had very poor results asking for pointers and hints that don't give away key insights. This has been the case across models I have tested.
An architecture with a "judge" that gates responses and ensures a lack of spoilers would probably work better. But this is a simple thing that they keep messing up.