HNHacker News
TopNewBestAskShowJobs

jasonjmcghee

4,908 karma · joined March 25, 2015

Website: https://jason.today

GitHub: https://github.com/jasonjmcghee

submissionscomments
jasonjmcghee··on Show HN: I trained a 125M model to autocomplete piano on-device
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- how large was the data? How many samples did you use to pretrain and post-train

jasonjmcghee··on GPU Offload in Rust: Portable, Safe, and Fast
> 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.

jasonjmcghee··on Show HN: Maple-Preview – Ternary 20B MoE running at 120 tok/s on a iPhone
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 compressed so much

jasonjmcghee··on Prevent cognitive debt by manually retyping LLM-generated code
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 devx)

Needing to manually touch every line of code is not scalable.

This was true 10 years ago. It's still true today.

jasonjmcghee··on The coolest use for the Vision Pro
I feel like this is supposed to be a "gotcha", but this is VR, not AR.

So you can just, move up.

jasonjmcghee··on Kimi Linear: An Expressive, Efficient Attention Architecture (2025)
(2025)

As it's 9 months old and they just had a major model release

jasonjmcghee··on Did they ghost you?
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 there was or if this is where we are
jasonjmcghee··on Claude Cookbook
You weren't kidding. That's embarrassing.

"Use a black background and this hideously wide font!"

jasonjmcghee··on Buz – A fork of Bun using modern Zig, with sub-1s incremental builds
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")
jasonjmcghee··on The Great Steam Deck Crash of 2026: How the New Pricing Decreased Demand by 80%
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.
jasonjmcghee··on A Second-Grade Teacher Revived a Beloved Video Game
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.

jasonjmcghee··on Backtrack-Free Cursive
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.

jasonjmcghee··on Your code is fast – if you're lucky
Casey Muratori's https://computerenhance.com is a good resource
jasonjmcghee··on After 7 years in production, Scarf has reluctantly moved away from Haskell
Hey Avi - didn't know you were an HNer fun to see you / your co on here :)
jasonjmcghee··on GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture [pdf]
No there was the planar unit distance problem

(Erdős problem 90)

jasonjmcghee··on Fable created novel 4D splat format
3D Gaussian splats are still images of 3D space / photogrammetry

So I'm guessing it came from representing the additional dimension

jasonjmcghee··on Fable created novel 4D splat format
time
jasonjmcghee··on Potential session/cache leakage between workspace instances or consumer accounts
Just use a symbolic link
jasonjmcghee··on GitFut – Your GitHub stats turned into a World-Cup-style player card
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?

jasonjmcghee··on How many failed startups have you launched?
For future reference, you need to put newlines between lines to preserve formatting.
jasonjmcghee··on Qwen 3.6 27B is the sweet spot for local development
(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.
jasonjmcghee··on Qwen 3.6 27B is the sweet spot for local development
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/s on low power and >100 t/s on high power.

Compared with previous message being official qwen huggingface release (non-mlx) using lm studio.

jasonjmcghee··on Qwen 3.6 27B is the sweet spot for local development
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 laptop.

If you currently don't use speculative decoding and you start using it, it can nearly offset the difference between high and low power, and it's night and day experience.

I almost always keep my laptop on low power mode.

jasonjmcghee··on Previewing GPT‑5.6 Sol: a next-generation model
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.

jasonjmcghee··on Show HN: I made Google Trends for Hacker News by indexing 18 years of comments
Terms with spaces seem to be an "or"
jasonjmcghee··on Claude: Elevated Error Rates for Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 4.6
Sonnet 5 here we come
jasonjmcghee··on Show HN: I wrote a C++ ray tracer from scratch without AI
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 series on building a ray tracer.

jasonjmcghee··on Conventional Commits encourages focus on the wrong things
Dependabot was a thing before LLMs
jasonjmcghee··on Codex for open source
They’ve been doing this since at least March
jasonjmcghee··on Show HN: Free animated icon library for Vue
Some of the animations are really good and make sense based on the icon.

Others just draw the SVG path or wiggle/bounce a bit.

If I were the creator here I might differentiate this into two buckets: custom animation or not. And then allow the generic animations to be applied to any of them.

← PreviousPage 3 of 34Next →