HNHacker News
TopNewBestAskShowJobs

currymj

3,822 karma · joined March 27, 2016

submissionscomments
currymj··on Cormac McCarthy's personal library
I don't think it diminishes him.

"Stella Maris" is a great novel that could only be written by someone who was very knowledgeable about math. As far as art that engages deeply with math and science, I don't know of anything comparable. Most artists would focus only on the human drama of discovery, without being able to engage with the subject matter.

However, I would consider "mastery" of a math textbook to be you have worked through almost all the chapters, can do a reasonable chunk of the exercises, and could TA the course without too much trouble.

While I don't know for sure, I doubt McCarthy achieved that level of understanding for all the yellow books he owned. I think buying a math textbook on an interesting topic and then not making it very far is a very common and human experience.

currymj··on Cormac McCarthy's personal library
If you read his last two novels he was clearly well-informed about mathematics.

However I have to assume that McCarthy didn't actually master all the material in the math books mentioned here, I think the reporter may be a little too credulous about that. I suspect he had the very common experience of buying a yellow book and being defeated in the first couple chapters.

currymj··on Sora 2
if you're maximizing accessibility you'd have both. often in broadcasts with closed captioning, there will also be a video of the sign language interpreter.
currymj··on Sora 2
sign languages are completely different languages from spoken languages, with their own grammar etc.

subtitles can work but it's basically a second language. perhaps comparable to many countries where people speak a dialect that's very different from the "standard" written language.

this is why you sometimes have sign language interpreters at events, rather than just captions.

there's not really a widely accepted written form of sign language.

currymj··on Sora 2
I also can't think of a reason why I would ever want to look at an AI generated video.

however as they hint at a little in the announcement, if video generation becomes good enough at simulating physics and environments realistically, that's very interesting for robotics.

currymj··on How has mathematics gotten so abstract?
two interesting cases: convex analysis and linear algebra are both relatively easy, concrete areas of mathematics. also beautiful and unbelievably useful. yet they didn't develop until the 19th century and didn't mature until the 20th.
currymj··on A lifetime of social ties adds up to healthy aging
from a certain perspective, they've found that if I tell you nothing more about a person than their lab test results on some biomarkers, you can make a better-than-random educated guess about how much social connection they have in their lives.

that's kind of interesting, though very far from the causal claims in the university press release.

currymj··on ChatGPT Pulse
they've already had that exact feature for a while, scheduled tasks are available in the settings menu. if you just tell the chat to schedule a task it will also make one automatically.
currymj··on You’re a slow thinker. Now what?
i think it just doesn’t matter much.

answering a question in 30 seconds vs. 5 minutes of thought is a very salient distinction in school and in some job interviews. it can impress people. so it feels painful to be slow.

all else equal you’d rather be quick. but it’s only a very small advantage for doing most real work.

currymj··on Claude now has access to a server-side container environment
the way it edits powerpoints is by launching a command line environment, and then editing the OOXML directly using command line tools. this takes several minutes to do even simple changes.

to me it seems miraculous that it even "sort of" works, but also it's not a reliable product yet. OOXML is very complex and the formatting can get mangled.

On the other hand, if you use LaTeX/Beamer slides, LLMs can reliably make a lot of formatting tweaks etc. and it is an actual time saver. But only weird academics use Beamer.

I agree with Simon Willison that this feature is really about writing code in a container, using that capability to edit PPT presentations as if they were markup is an odd thing to make the primary selling point.

currymj··on All vibe coding tools are selling a get rich quick scheme
i think a lot of social media influencers are selling a get rich quick scheme around other people's AI products.

the actual AI companies marketing hype seems to claim:

1) you can make fun non-serious toy apps

2) any day now massive productivity increases for software companies. this is different than anyone can launch a product with no software skills.

currymj··on GPT-5 Thinking in ChatGPT (a.k.a. Research Goblin) is good at search
the thing about students who cheat is most of them are (at least in the context of schoolwork) very lazy and don't care if their work is high quality. i would guess waiting multiple minutes for Thinking mode to give thorough results is very unappealing. 4o or 4o-mini was already good enough for their purposes.
currymj··on 996
not only does a founder capture more upside, they also pay less of a cost. autonomy prevents burnout even under very intense working conditions.

plus the nature of a founder's day to day work is very different. 12 hours a day of management, pitches, meetings, and snap decisions is doable for a long time if you can endure the pain.

12 hours a day of complex technical work under sleep deprivation is just not possible, after a few weeks of this your cognitive function will decline to the point you can't do the job right.

currymj··on If you have a Claude account, they're going to train on your data moving forward
i think skepticism is healthy, but they've handled this in a fairer way than any other online product i've used before.

they gave me a popup to agree to the ToS change, but I can ignore it for a month and still use the product. In the popup, they clearly explained the opt-out switch, which is available in the popup itself as well as in the settings.

currymj··on Updates to Consumer Terms and Privacy Policy
i found the ToS change to be surprisingly not a dark pattern.

in fact, i haven’t agreed to it yet, and was able to close the popup and continue using Claude. they also made it extremely clear how to opt out, providing the switch right in the popup and reminding me it’s also in settings.

when i eventually do have to agree to the ToS changes, i’ll probably just stay opted out.

currymj··on Training language models to be warm and empathetic makes them less reliable
in ChatGPT settings now there is a question "What personality should ChatGPT have?". you can set it to "Robot". highly recommended.
currymj··on GPT-5: It just does stuff
if the router is working properly, it should send questions like this to thinking mode which can handle it well. unfortunately the router doesn't always seem to make the right call.
currymj··on Scientific fraud has become an 'industry,' analysis finds
I do think that you can tell apart these cases. Outright fabricated data is very different from p-hacking, which is very different from a meaningless garbage paper.

On the other hand, convenient laboratory errors might be hard to tell apart from fabrication (though not always, sometimes people get caught using photoshop), and statistical incompetence might be hard to tell apart from p-hacking. So I agree this is why it's hard to punish fraud.

currymj··on Scientific fraud has become an 'industry,' analysis finds
probably not. the writing quality will improve but LLM-generated papers will still be ignored.

to the extent they aren't ignored, but seem so plausible that they are taken seriously, eventually people will want to talk to the researchers about their results, invite them to give talks, and so on. at which point it becomes problem 1.

currymj··on Scientific fraud has become an 'industry,' analysis finds
there are many separate problems of scientific fraud. I think these issues sometimes get confused which is unhelpful.

1. apparently-legitimate papers in prestigious journals with fraudulent data. extremely bad.

2. legitimate papers in legitimate journals which, innocently or not, just used bad methods and have wrong conclusions. this is "the replication crisis".

3. totally fake papers in paper mills with no meaningful peer review. it's really easy to spot these, no one is individually getting taken in by the results, but...

3a. sometimes they wind up in a meta-analysis, which is really bad because people might trust the meta analysis.

Problem 1 is morally worst and much more common than one would hope. Outright fabricated data in a Nature or NEJM publication (as has happened) is a disaster.

Problem 2 is amenable to reform for the most part (fields are already doing this).

Problem 3 isn't a problem at all for scientific knowledge per se, although universities and funding bodies might not be pleased their scientists are buying fake papers. You can just ignore the paper mills.

But Problem 3a can actually alter policy, which is pretty serious.

currymj··on At 17, Hannah Cairo solved a major math mystery
i can understand a university administrator not feeling they can give a good educational experience to a 14 year old.

but she'd be the same age as all the first years on campus, any university must be institutionally equipped to deal with a 17/18 year old. seems like an odd choice to personally intervene to cancel an admission.

like presumably if she enrolled in a bachelor's program, and then signed up for grad math classes while doing independent research, this would be fine. but if she takes the same classes and does the same research for PhD credits instead, no good?

currymj··on Study mode
this basically functions as a switch you can press that says "more effort please". after every response it makes you solve a little comprehension check problem before moving on. you can try to weasel out of it but it does push back a bit.

unavoidably, people who don't want to work, won't push the "work harder" button.

currymj··on Study mode
as a professor i'm tentatively pleased. it's hard to shake out the edge cases for anything like this so it may break down or cause annoying problems, who knows.

but even with this feature in this very early state, it seems quite useful. i dropped in some slides from a class and pretended to be a student, and it handled questions reasonably. Right now it seems I will be happy for my students to use this.

taking a wider perspective, I think it is a good sign that OpenAI is culturally capable of making a high-friction product that challenges and frustrates, yet benefits, the user. hopefully this can help with the broader problem of sycophancy.

currymj··on Claude Code weekly rate limits
a lot of "all you can eat" restaurants have to charge extra for uneaten food. there are people who just enjoy the feeling of abundance they get from paying a flat fee and then wasting something.

This is clearly what was happening with the most extreme Claude Code users, because it's not actually that smart yet and still requires a human to often be in the loop.

However, Anthropic can't really identify "wasted code".

currymj··on The 1970s psychology experiment behind 'Star Wars' special effects (2023)
Those came from Larry Cuba, who worked at the Electronic Visualization Laboratory at the University of Illinois Chicago.

https://www.chicagotribune.com/2017/05/23/blueprints-for-sta...

currymj··on TikTok is being flooded with racist AI videos generated by Google's Veo 3
from an information theory perspective, predicting and efficiently representing data is so closely tied to generation that it is unavoidable.

if you want to use ML to do anything at all with image and video, you will usually wind up creating the capability to generate image and video one way or another.

however building a polished consumer product is a choice, and probably a mistake. every technology has good and bad uses, but there seem to be few and trivial good uses for image/video generation, with many severe bad uses.

currymj··on The role of the University is to resist AI
this article contains some reasonable points about problems caused both by concrete advances in AI and by rhetoric about AI used to justify the decisions of the powerful. the focus is especially on AI-based end-user technology being sold to consumers and institutions.

it seems to miss the point that some members of the university community may actually do research on AI, and may not feel inclined to ask for approval from the “People’s Council” to get permission to think about concepts in statistics, computer science, or decision theory.

currymj··on The bitter lesson is coming for tokenization
thanks to training data + this being a popular benchmark, they're pretty good at grinding through symbolic mathematical derivations, which is often useful if you want an explanation of a mathematical concept. there's not really a better tool for this job, except for "a textbook which answers the exact question you have".

but from time to time, doing this does require doing arithmetic correctly (to correctly add two exponents or whatever). so it would be nice to be able to trust that.

i imagine there are other uses for basic arithmetic too, QA applications over data that quotes statistics and such.

currymj··on Researchers are now vacuuming DNA from the air
everyone already leaves DNA everywhere, so it doesn’t seem like a step change.

genetic privacy is a good thing but is utterly artificial, we have to create it if we want it.

currymj··on GitHub MCP exploited: Accessing private repositories via MCP
even in the much simpler world of image classifiers, avoiding both adversarial inputs and data poisoning attacks on the training data is extremely hard. when it can be done, it comes at a cost to performance. I don't expect it to be much easier for LLMs, although I hope people can make some progress.
← PreviousPage 2 of 34Next →