Third-year Ph.D. student at CMU, working on programming languages and formal verification; CS undergrad at Caltech (BS '23, Venerable); ex-professional software developer.
https://github.com/jgrosso
Submissions
ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Cases
arxiv.org
2 points·by BalinKing··0 comments
NTSB subpoenas American Airlines flight crew in JFK runway incursion
I've never really understood this argument. If someone's a manager of an incompetent team, no amount of management skill will save the quality of the resulting software. I don't think "just treat LLMs like smart junior developers" fixes this, because well-functioning teams usually also have senior developers to keep things on track. Like, if we handed a team of genius-but-junior developers to the best "people person" manager in the world (i.e. who doesn't actually read the code), would we really expect decent results? Even if the manager tested the code by hand? I honestly don't think so, at least once the software gets past a certain (fairly low) threshold of size/complexity.
The general sense of “liking” something is usually 好き (suki) in Japanese, AFAIK. Depending on the context (romantic, etc.), “love” could be 愛, 恋愛, 大好き, and probably others.
I ran into that issue too w/ sentence-based flashcards, where I almost immediately memorize the sentence itself and the whole thing becomes self-defeating. Similarly, I thought to use LLMs to generate fresh sentences on-the-fly, but the output was never reliable enough for my use-case... e.g. I came across some grammatical construction that the LLMs refused to use correctly. But, this probably depends on the language in question, since it sounds like you had success with it! Maybe I should try again at some point, now that the consumer models are a lot beefier than they were even six months ago.
On the subject of nostalgic paint apps, I distinctly remember one of the Kid Pix drawing tools/games that I had growing up... what a weird piece of software that was!
IIRC linguists (like, actual academic researchers) prefer to “break up” kana and analyze the consonant and vowel separately when dealing with conjugations. Treating kana as indivisible units is AFAIK only really a thing in Japanese linguistics historically done in Japan. All this to say, I’m pretty sure this “just change the vowel” approach is perfectly fine from a theoretical perspective (and as a fellow learner, also very aesthetically satisfying :-) )
The lead says "how I approach IIS targets during bug bounty" (emphasis mine), so (assuming the author is being truthful) I'm guessing the tone of the title is just for fun.
Excluding supergeniuses, pure mathematics—even at a very basic, undergraduate level—simply can't be understood passively. Even with an infinitely patient AI teacher who could answer any question on-demand, it'd still require a massive amount of work to actually understand anything in research-level mathematics. Basically every single word in a mathematical definition is a term of art, and (IME) if one doesn't grok each of those words at a fairly deep level, the new definition never really makes too much sense. And this applies recursively: each of the words has some thoroughly inscrutable definition of their own.
Of course it'd be super helpful to have, say, a teacher who could tailor explanations to anyone's precise background (e.g. where possible, using examples that come from the student's field of study when explaining some abstract concept). Or, if some definition comes with some precondition that has no obvious purpose, perhaps an omniscient teacher could explain why it's there with concrete counterexamples.[0] But even granting all this, I think that mathematical intuition is necessarily based on a lot of hard work actually exploring definitions on one's own, with pencil-and-paper and a lot of thought. That is to say, even though the process could probably be sped up a lot with a nigh-omniscient teacher[1], I doubt that a student wouldn't still need years of training to even have a clue what's going on.
(I'm saying all this, by the way, as someone who is terrible at all this and has very little mathematical maturity[2]—I'm speaking from my own frustrating experience....)
[0] c.f. Lakatos' excellent book Proofs and Refutations
[1] without the "curse of knowledge," or else we're back to square one of "answers that are correct but useless"
The advantage(?) of take-home exams à la Caltech is that they can be open everything and 3–5 hours long :-P (For what it's worth, being able to listen to music during an exam, ctrl+F a digital textbook, etc. was super awesome; it would deeply sadden me if that becomes infeasible in the future once enough students stop caring about the Honor Code....)
Things may have changed, but I don't recall any group exams during my time at Caltech, and conversely I do recall a strong sense of pride in the Honor Code. Also, if your professor allows collaboration, then it's definitionally not cheating: There is a vast moral difference between "the professor made the assignments difficult with the specific expectation that people will collaborate" and "the professor doesn't want collaboration but people did it anyway".
Frankly, this comment feels almost entirely foreign to my experience—I suppose things could've changed over the years (although my impression is that things have gotten much worse recently, not better), or it could be major-specific, or I just got lucky with the specific people I happened to hang out with?
I'm not the other commenter (and I believe you that it's not AI), but I'd guess it's mostly the first line: a short affirmation followed by "The problem is ...." feels like the sort of formula the LLMs love to use. (Not trying to imply that there's anything inherently wrong with it, of course.)
While we're at it, I'm under the impression that the recent LLMs have also co-opted "genuinely", which I'll never forgive them for—first they stole my em-dashes, and now they're stealing my adverbs too?!
It's just that, in my (uninformed) opinion, Anthropic is incentivized a priori to claim things like this about their models. Like, it's probably really good marketing to say "our product is so smart, and we're so concerned about ethics, that made sure a psychiatrist talked to it". I guess it's ultimately a judgment call, but to me the conflict of interest seems big enough that I'm really wary of this sort of argument. (I'm reminded of when OpenAI claimed GPT-5(?) was "PhD-level"—I can personally attest that, at least in my field, this is totally inaccurate.)
I think a steelman interpretation of the parent is that entirely LLM-generated projects should be disallowed. There's a lot of submissions on Show HN that seem completely vibe-coded to me (like, including the README), which is a very different situation IMO from someone who simply used Claude to write some—or even most—of the code. When even the human-facing portion of a submission is LLM-generated, it bothers a lot of people (myself included).
https://github.com/jgrosso