btrettel

4,167 karmajoined 9 anni fa

I am a mechanical engineer working in computational fluid dynamics.

Personal website (including contact information): http://trettel.us/

comments

13 ore fadiscuss

I'm a mechanical engineer who has written similar tools for work and hobbies. Producing pretty pictures does not mean that the model is physically accurate. Unfortunately, such tools seem be evaluated much more on flashiness and not on more reliable and objective criteria like physical accuracy based on verification and validation test suites. I'm seeing that in the comments here. I don't think LLMs make what I do irrelevant, but I have thought that I'm going to have to improve how flashy my simulations look to compete better with non-experts who use LLMs.

15 ore fadiscuss

Unfortunately, running an online forum has been a pain for a long time. You have to promote the forum to keep it active, deal with various bad actors (historically spammers, trolls, and black hats), maintain the forum, deal with drama, etc. It adds up and largely isn't appreciated. People act like a forum exists by itself, but it doesn't.

And in the last 5 years, running an online forum has become more of a pain given how badly behaved some bots are. Just recently, I installed Anubis on some online forums I run, and I've been amazed by how much traffic dropped. Before, server load was becoming a problem to the point where all the forums I ran were taken offline by their hosts! I have been thinking about how there's a need for a forum software which produces static HTML for the content, while all dynamic components are behind a login. Bots won't increase server load much in this case. If the forum administrator decides to end the forum, they can easily keep hosting the content without any future maintenance beyond paying the bills. Two of the forums I run are just archives at this point, and I'd love to be able to flip a switch and make them static HTML... (I probably will adapt some script I found on GitHub do to this in the future, maybe with help from LLMs.)

10 giorni fadiscuss

Location: United States (Open to any US location)

Remote: Yes, open to remote, hybrid, or in office

Willing to relocate: Yes

Technologies: Fortran, Python (Matplotlib, Numpy, Pandas, Scipy), OpenMP, Git/GitHub, Linux, Bash, others...

Résumé/CV: Available on request

Email: [email protected]

GitHub: https://github.com/btrettel

Personal website: http://trettel.us/

I'm Ben Trettel, an experienced mechanical engineer with a PhD, specializing in computational fluid dynamics, design optimization, and verification & validation of computer simulations.

I am particularly interested in opportunities to build cutting-edge physical products where computational simulation and design optimization are key.

11 giorni fadiscuss

A spell checker, grammar checker, and tutor change a relatively small fraction of the writing, preserve the writer's style/voice, and rarely introduce errors that are hard to detect like hallucinations.

A translation app changes nearly 100% of the content, often changes the writer's style/voice, and can introduce hard to detect errors. But there's a far closer correspondence to what was written by the original writer. The basic ideas are still from the writer. A translation app is not expanding a short idea into something longer, and including some things the original writer never thought in the process.

***

Pre-LLMs, I did in fact disclose when I was using a translation app in some translations of scientific articles I produced. It would be weird to disclose the use of spell checking, grammar checking, or who previously taught me writing as these things are ubiquitous. I will also acknowledge people who were influential in my thinking. If a LLM is doing a lot of the thinking for me then I do think disclosing LLM use is appropriate.

13 giorni fadiscuss

Recommender systems for papers tend to be pretty bad, so there's a lot of room for improvement. I'll use Semantic Scholar as an example. I have a bunch of folders in what they call a "Library" with recommendations turned on. Semantic Scholar tends to recommend things that are in the same general area but not specific enough. So I guess that Semantic Scholar seems to interpret adding a paper to a folder as expanding the scope of the folder, but it could be narrowing. There's no way to distinguish between the two. Their recommender system is supposed to magically figure it out. Some way to add additional context like relevant keywords or a way to select which parts of the papers are relevant would be helpful. As it stands, I have to repeatedly thumbs down recommendations, and Semantic Scholar doesn't figure out what I mean from that vague signal and instead stops recommending much anything. It's not that there are no additional papers to go into these folders either as I've added more over time that I've found through other means.

13 giorni fadiscuss

I also use a unique email address for Hacker News, and I get no spam to that address.

The (different) address I use for the HN jobs threads gets only spam, though.

18 giorni fadiscuss

Are you loosely paraphrasing here? The closest thing I could find by Scott Adams was "Whenever you have a lot of money in play, combined with the ability to hide misbehavior behind complexity, you should expect widespread fraud to happen."

https://www.goodreads.com/quotes/10213582-whenever-you-have-...

(I didn't check the book, though.)

20 giorni fadiscuss

I had a similar experience before [1]. I fully agree that too many interview tests select for dishonest people.

[1] https://news.ycombinator.com/item?id=35496976

21 giorni fadiscuss

Unfortunately, when managers expect a certain throughput, solutions like this often appear. I saw similar systems when I was a patent examiner at the USPTO. From what I recall being told, the USPTO's "form paragraphs" started out in the DOS era as some WordPerfect macros developed by an examiner, not management. Management defined the quota and examiners came up with creative solutions to meet the quota.

These solutions are symptoms of a broken system. I would not judge the people working within the system for using these solutions (edit: unless someone's quality is exceptionally bad like academic fraudsters as a academics work in a similar metrics-driven environment). Management (and politicians in the case of the USPTO) created the incentives that are the real problem.

Automation like this can be useful to enhance quality, but in my experience at the USPTO, there was a lot of automation they could have used to enhance quality, but they didn't. The incentives to improve speed are far stronger than the incentives to improve quality.

22 giorni fadiscuss

Getting Western researchers to not ignore papers published in languages other than English is a serious uphill battle.

During my PhD, I tried to do a truly global literature review of all languages. I tried to be a lot more comprehensive than others before. I got quite good at locating existing English translations of non-English papers [1], and even published around a dozen English translations I produced using Google Translate, DeepL, and Yandex Translate [2]. My PhD advisor clearly disliked this aspect of my research and didn't seem to consider it to be research. Despite that, I remember when one of my papers was being reviewed that a reviewer commented that X was the most valuable contribution of my paper. Problem was that X arguably wasn't a contribution because it came from a Russian paper published back in 1963, as I stated in my paper! Yes, I did a better job at X than they did back in 1963, but it's only a "contribution" because people ignored non-English papers.

Ultimately, I think ignoring non-English papers is one aspect of the larger problem that literature reviews tend to be non-comprehensive. The literature is rarely well organized, and no, LLMs don't solve this problem, though I do think they help and will get better over time in this aspect.

[1] https://academia.stackexchange.com/a/93209

[2] For example, here's one that I find was way ahead of its time (original was published in 1938) and still arguably had publishable elements back in 2020: https://repositories.lib.utexas.edu/items/ca7fc7d3-cf16-4859...

24 giorni fadiscuss

I don't think I experienced discrimination during admissions either. Off the top of my head, I don't know any US citizens who told me that they wanted to go to grad school but were unable to be admitted to a school.

24 giorni fadiscuss

As a US citizen with a PhD, I didn't experience any clear discrimination in favor of foreign students during grad school.

I think the main reason so few US citizens get PhDs is because PhD "student" (they're actually workers) positions pay so poorly. Make PhD student positions have non-poverty wages and you'll see a lot more interest from US citizens.

On the flip side, I think foreign students experienced a lot of abusive conditions that I could more easily say no to because I didn't have a visa that required me to work at the university. I've seen some of that first hand. I don't mean to imply that there would be no cost to me saying no, just that I wouldn't have to leave the country if I said no.

mese scorsodiscuss

Even if it wasn't a large size, it likely wouldn't be great. During my PhD on sprays, I did some (unpublished) experiments using isopropyl alcohol to reduce the surface tension. The nozzles I used were around 1 mm in diameter as I recall. I did not anticipate that the room would fill up with isopropyl alcohol vapor and (probably) tiny droplets. I wore a mask and maybe left the room while each trial was running. Breathing that likely wasn't great for my lungs.

mese scorsodiscuss

Fluid dynamicist here. The word "compressible" has multiple meanings and this might be confusing you. You don't need compressible flows in the sense of high Mach numbers. There are other models where the flow is variable density, but thermodynamic and hydrodynamic pressure are decoupled to remove the pressure waves that make high Mach number flows hard. There's also the Boussinesq approximation for buoyancy when the density varies only a small amount. I'm not particularly familiar with atmospheric models, but I'm sure they don't use the high Mach number form. "Incompressible" methods are common for the second class of model I mentioned, though how to use them so might not be obvious.

mese scorsodiscuss

Location: United States (Open to any US location)

Remote: Yes, open to remote, hybrid, or in office

Willing to relocate: Yes

Technologies: Fortran, Python (Matplotlib, Numpy, Pandas, Scipy), OpenMP, Git/GitHub, Linux, Bash, others...

Résumé/CV: Available on request

Email: [email protected]

GitHub: https://github.com/btrettel

Personal website: http://trettel.us/

I'm Ben Trettel, an experienced mechanical engineer with a PhD, specializing in computational fluid dynamics, design optimization, and verification & validation of computer simulations.

I am particularly interested in opportunities to build cutting-edge physical products where computational simulation and design optimization are key.

mese scorsodiscuss

I think the type of speed that you're referring to is different from the type of speed the linked article is referring to. I'm not sure what the best way to distinguish the two is. With slow software, you're presumably getting a right answer, just slower. In my job, people who work quickly often produce a wrong answer.

2 mesi fadiscuss

One approach for testing with multiple compilers that I use on some Fortran projects (where testing against multiple compilers seems more common than in C) is to use a variable from the command line to specify the compiler, for example:

    make FC=ifx check

On my Fortran projects, that will run the tests with Intel's Fortran compiler. The Makefile has logic to automatically change compiler flags as appropriate. I default to the GNU Fortran compiler, so `FC` isn't required.

I have made a script to run through a series of compilers by alternating between `make check` and `make clean`.

I have separate Makefiles for GNU Make and NMAKE/jom. My Fortran code works fine on various Linux distributions and Windows, though I'll add that achieving that is probably easier with Fortran than C. I've also tried a BSD Make that worked (on Ubuntu at least). My Makefiles are pretty close to the intersection of POSIX and NMAKE, so the main differences between the different Make versions are the conditional statements needed to handle the different compiler flags and the include statements (as I put the compiler flags in separate files).

2 mesi fadiscuss

I had a similar setup in 2023, but the computer was reformatted after I moved. I wrote a HN comment about the setup before: https://news.ycombinator.com/item?id=37792204

I liked it and intend to use a similar setup in the future. There were quite a few "rough edges", unfortunately. In retrospect, a tiling window manager would have been a better choice.

I found Midnight Command to be great for this, with its integrated file manager, file viewer (mcview), editor (mcedit), and diff (mcdiff).

I didn't realize how much I relied on a unified clipboard until I didn't have one any longer. mcedit's clipboard was a file (or one of them was?), so I had to adjust some workflows.

The biggest problem came from my need to view a lot of PDF files. I had a framebuffer PDF viewer that was pretty clunky. It did not work with tmux and PDF files could not be opened directly from Midnight Commander as I recall. This specifically is why I'm thinking about a tiling window manager as I won't have to pick a clunky PDF viewer and the remainder will just work.

2 mesi fadiscuss

My PhD was in mechanical engineering. I don't want to get more specific than that on the subject in case it identifies the lecturer. I guess I used the phrase "lecturer" to distinguish him from my PhD advisor.

2 mesi fadiscuss

I had a similar experience when I was a TA at UT Austin that I wrote about on HN years ago: https://news.ycombinator.com/item?id=23163472