jasonjmcghee

5 ms·

31.
by jasonjmcghee 2mo ago
M5 prefill is much faster than M4. I've seen benchmarks that show 4-5x faster of M5 Max vs. M4 Max. For local models you're likely using M5 Max, prefill is low thousands of tokens per second, as opposed to, say high hundreds with
32.
by jasonjmcghee 2mo ago
It's like tongue-in-cheek intentional slop though. The written text is good lol
33.
by jasonjmcghee 2mo ago
But even then, it's targeting like 300ms not 30ms, right?
34.
by jasonjmcghee 2mo ago
I think this is a great project and very HN. Not sure why the comments are so focused on the deliverable- you learned way more and had a much more interesting experience. One think I didn't see mentioned in the post- maybe I missed it-
35.
by jasonjmcghee 2mo ago
> the rust-gpu project has to emulate pointers[8], which we consider a blocking issue for most HPC benchmarks. Why is it a blocking issue? I feel like this is very aligned with the goals of rust-gpu.
36.
by jasonjmcghee 2mo ago
It's far more likely basic consumer devices will advance such that much larger models efficiently run than finding novel ways to compress all of human experience to fit on today's mobile hardware. Information can only be compresse
37.
by jasonjmcghee 2mo ago
I don't manually retype the code from pull requests of other engineers. It's important to retain the what why and how as a team to the degree that development can be efficient, extensible, and maintainable. (And ideally a good dev
38.
by jasonjmcghee 3mo ago
I feel like this is supposed to be a "gotcha", but this is VR, not AR. So you can just, move up.
39.
by jasonjmcghee 3mo ago
(2025) As it's 9 months old and they just had a major model release
40.
by jasonjmcghee 3mo ago
Based on the strange emoji use in place of some buttons, I'm guessing this is vibe coded- and a lot of this is really nice looking front end - great color, topography, border styles... Would be interested to know how much intervention
41.
by jasonjmcghee 3mo ago
You weren't kidding. That's embarrassing. "Use a black background and this hideously wide font!"
42.
by jasonjmcghee 3mo ago
LLMs from 1+ years ago approached solving every problem by adding additional code and not cleaning up or adding strange unnecessary backwards compatibility (and often still do this kind of thing - "all tests green")
43.
by jasonjmcghee 3mo ago
Not trying to be an apologist for anyone, but if the epic store could charge 30%, it's hard for me to believe they wouldn't.
44.
by jasonjmcghee 3mo ago
I saw backyard baseball on steam the other day and was blown away by the $40 price tag. I expected $10 maybe $20. I think $40 is too far outside of impulse / nostalgia purchase.
45.
by jasonjmcghee 3mo ago
Weird amount of negativity here. Some person recognized a problem they wanted solved and put a bunch of effort into creating a unique solution. I love this stuff. This post is perfect for HN.
46.
by jasonjmcghee 3mo ago
Casey Muratori's https://computerenhance.com is a good resource
47.
by jasonjmcghee 3mo ago
Hey Avi - didn't know you were an HNer fun to see you / your co on here :)
48.
by jasonjmcghee 3mo ago
No there was the planar unit distance problem (Erdős problem 90)
49.
by jasonjmcghee 3mo ago
3D Gaussian splats are still images of 3D space / photogrammetry So I'm guessing it came from representing the additional dimension
50.
by jasonjmcghee 3mo ago
time
51.
by jasonjmcghee 3mo ago
Just use a symbolic link
52.
by jasonjmcghee 3mo ago
It says I’ve done 8 PRs- I’ve done a few thousand at least lol - you could get a ballpark by looking at “Pull Shark” count. Commits also look like it’s this year only?
53.
by jasonjmcghee 3mo ago
For future reference, you need to put newlines between lines to preserve formatting.
54.
by jasonjmcghee 3mo ago
(see above reply to myself) I misattributed the gain from dflash - it was dflash-mlx library + mlx model, not dflash itself giving me the speedup.
55.
by jasonjmcghee 3mo ago
I usually get mlx versions of models and didn't realize I was using non-mlx - misattributing the performance boost to speculative decoding. With dflash-mlx library, with dflash disabled 3.6-35b-a3b mlx model I'm getting ~60 t&#x
56.
by jasonjmcghee 4mo ago
I'm surprised no one has else has mentioned - low power mode. With no speculative decoding, using high power mode, I get 80 t/s on 35B A3B - and it gets hot and spins up. On low power mode I get 38 t/s - no fans, cool to warm
57.
by jasonjmcghee 4mo ago
Try gpt-5.3-codex-spark - it's 1000 TPS and from my experience more capable than 5.4 mini. If you have a subscription it's a different pool of usage.
58.
by jasonjmcghee 4mo ago
Terms with spaces seem to be an "or"
59.
by jasonjmcghee 4mo ago
Sonnet 5 here we come
60.
by jasonjmcghee 4mo ago
This is (or at least used to be) a right of passage in the graphics world. I think many people go through the very popular https://raytracing.github.io/ There was a big influx of this when Sebastian Lague did his video seri

More ›