YuechenLi

5 ms·

31.
by YuechenLi 2mo ago
Yeah, it's pretty much apples to oranges, and I don't consider GPT and Claude to be interchangeable at all. From my anecdotal experience, GPTs generally codes more creatively and verbosely but Claudes tend to code more carefully a
32.
by YuechenLi 2mo ago
Fable is just way too expensive and limited compared to GPT 5.6 Sol, and the only task that requires that level of intelligence is frontier scientific research. I use GPT/Codex primarily for coding and usually keep Claude on Sonnet 5 m
33.
by YuechenLi 2mo ago
Funny the performance section just progressively turned into "don't use Python if you need performance": Remove Python loops and vectorize with NumPy => use Numba JIT'd Python that is only partially compatible with CP
34.
by YuechenLi 2mo ago
The sad reality is that having something that works, even if badly, is better than having a theoretically elegant architecture that is not implemented. See JS or Linux vs Hurd for other examples. Ask yourself this question, supposedly someb
35.
by YuechenLi 2mo ago
I agree, yeah, SQL syntax is awful. But the easier solution is what pretty much what backend has converged on, have something in your backend programming language that lowers to SQL so you never have to write any raw SQL at all except as a
36.
by YuechenLi 2mo ago
SQL is what it is today because it is battle tested and has to handles a very hard problem of handling arbitrary concurrent reads/writes, so the likely scenario is that trying to replace general SQL wholesale will just end up making a
37.
by YuechenLi 2mo ago
A lot of time I'm not very confident in what I wrote because I have a tendency to ramble and not get to the point quickly. Reading helps, but it's more about style, doesn't help me express myself more eloquently at all. Anoth
38.
by YuechenLi 2mo ago
Well, because you need a geometric modeling kernel to generate BRep CAD models, it really has nothing to do with training set as it is very, very, difficult to make one. Lucky for you, I made one that you can try out for vibe-CADing. https
39.
Show HN: BRep Geometric CAD Kernel and Parametric Code CAD(github.com)
2 points by YuechenLi 2mo ago|0 comments
40.
by YuechenLi 2mo ago
I think you misunderstood me. I did use them at first, but it took so much effort for me to set them up and config them for that particular application (Code only Skyrim modding) where I pretty much spent an entire day debugging and diagnos
41.
by YuechenLi 2mo ago
Well, Carbon would have been better off if they just made "C with `match` and sane error handling" instead of whatever it is they are building there over at Google. But still, I think C and C++ are honestly very different language
42.
by YuechenLi 2mo ago
Carbon isn't even a real language, no real language will spent 4 entire years and is still stuck in version 0.0.0-0. It feels more like bad Rust than a proper C++ successor at all. Just to give you a small example? why would a C++ succ
43.
by YuechenLi 2mo ago
In 2026, there really needs to be a REALLY good reason for people to use C++ for greenfield projects over using something like Rust, Go or C#. I mean, I've made it work, but that pretty much involves making a completely new build syste
44.
by YuechenLi 2mo ago
https://github.com/yuechen-li-dev/Aetheris So, yeah, mine is full geometry kernel that was made for CAD. I'll do a full write-up on Show HN later as there is just way too much stuff to cover for the project. The g
45.
by YuechenLi 2mo ago
@baigy Huh. Interesting. I was just doing the final clean-up for something convergent to this research that I had been working on for the past few months. I think I arrived at your thesis (code first semantics from a different direction in
46.
by YuechenLi 2mo ago
Well, I suppose fake Python is better than fake C++ at least. But yeah, I think Python's dominance in science/ML will eventually pass, just as FORTRAN and Matlab did before.
47.
by YuechenLi 2mo ago
So... why go through LLVM at all instead of having the MIR target PTX/HIP C directly then? If they really wanted a vendor neutral solution for Rust GPU, that already exists: you write the CPU side code, including buffering, allocation,
48.
by YuechenLi 2mo ago
I think you are pretty precise and your comment is appreciated, so I think I need to clarify my wording as well. I think a lot of CUDA C/C++'s difficulties are self-inflicted because they conflate two things into what they call &q
49.
by YuechenLi 2mo ago
Oh, since the topic of semantics compression via LLMs came up, here is some interesting research result that I had found earlier this year that I posted here and failed to explain properly, with a benchmark as well for you to try on your ow
50.
by YuechenLi 2mo ago
I think there is a misconception here: CUDA is not even close to being C for assembly used by Nvidia's GPUs, PTX is, and it uses JIT to compile to Nvidia's GPU assembly, SASS. It honestly easier just to have LLMs write PTX directl
51.
by YuechenLi 2mo ago
I'm curious on what you mean by "TS use-after-free", because as you obviously know, TS is GC'd. I don't have access to your codebase of course, but it seems to me that it is an FFI/native code boundary lifetime
52.
by YuechenLi 2mo ago
Sure, what I mean by that is that LLM makes different kind of mistakes than humans, they usually take the shortest direct route to accomplish their task. You can see that with the Bun Rust rewrite, I don't think any human coder would p
53.
by YuechenLi 2mo ago
C#, only because of dotnet ecosystem and tooling is great. TypeScript is a dark horse candidate, it's a great language with a great ecosystem trapped by JS tooling, and most of all, NPM. Rust is fine if you just tell LLMs to use short
54.
by YuechenLi 2mo ago
I wouldn't say Go is IDEAL for AI coding, but it certainly has the case for one of the best programming languages that currently AI uses. Go definitely has its share of problems for human authors because it's so verbose and boiler
55.
by YuechenLi 2mo ago
Nvidia's biggest advantage in AI has never been only their hardware performance but how entrenched their software is in ML research that flowed down stream. However, if you've actually used CUDA C/C++, it's pretty one of
56.
by YuechenLi 2mo ago
If you actually read the data, especially the distribution graph in the last image, the conclusion that it draws is "the run-to-run distribution variance is so big that there doesn't seem to be a correlation that can be drawn from
57.
by YuechenLi 2mo ago
Reflection is good for prototyping and get something setup quickly, but if you build your architecture around it, not only do you lose access to NativeAOT, the code becomes very hard to debug, and if you code with LLM a lot, you either have
58.
by YuechenLi 2mo ago
Go is absolutely one of the best programming languages for LLMs for the reason you say, and Python is just what LLMs like to use to write short throwaway scripts. Frontier LLMs are generally pretty good at most programming languages and can
59.
by YuechenLi 2mo ago
>But a genuinely good steak, medium-rare from edge to edge, browned properly, seasoned right, consistently delicious, is a different matter entirely. Actually it's pretty easy to get a consistently good steak with reverse searing, p
60.
by YuechenLi 2mo ago
Cobol code isn't esoteric, but it is designed in a way that is unambiguous and clear about its control flow, but very verbose and tedious to read/write. Translating from Cobol to a modern programming language is trivial for LLMs f

More ›