Zimeng Xiong's WeblogRecent
somehow still passed USACO despite their servers going down. my last two submissions were graded ~15 minutes AFTER contest end. one 100% passed the other TLE'd on 50% of the them, estimated ~833/1000
USACO needs better servers. Cannot access or submit any problems at all. Website does not load. Even when I could, grading was taking 10-15 minutes per submission. Not passing this one 😦
last day of Claude Max sub (expiring midnight today)... time to give ChatGPT Pro a shot? not sure how Opus 4.5 compares to 5.2 high/xhigh. ChatGPT Pro might be worth it just for access to 5.2 Pro though. Will consider options after usaco tomorrow, have to get up earlyish.
my linear algebra class often makes us name theorem numbers by theorem - i often need a fast way to search semantically 70+ PDFs and look up theorem numbers and proofs, might put claude onto this and see how it does, make a quick website
have yet to find an LLM that can accurately replicate style and voice given a large corpus of text. why is noone making the equivalent of voice cloning for TTS models for LLMs? seems obvious. tried gemini 3 pro, opus 4.5 and kimi 2.5 thinking. I would say kimi was the closest, but still far from passable.
based on some searches, it seems Trutone is doing this, but they are still on a waitlist basis.
i hate how github doesnt send you any emails for new issues created on your repositories. i understand for large repos they only want assigned memebers tagged, but i feel the default behavior for all repos (most on Github are small) should be to send emails on issue creation. that aside, let's see if i can close all 7 issues + 5 PRs (excluding 2 dependabot) today
it is totally possible to get addicted to programming. past week i have not done any schoolwork before 11PM, just coding. as a result i have too many assignments due tomorrow, including expiring extensions that i asked for so i could setup my sever 😦
14h39m on Ghostty today alone...
someone should make a screen time tracker for the terminal and see how long you spend on certain commands, not sure how you would hook into it though? might be cool to explore
just setup backup for the blog and feed to export convex data and commit to github 🙂. if only...if only.
opus made light work of integrating convex. now both the frontend and CMS interface with convex directly using the `userQuery` hooks so that changes in convex are reflected *immediately* in the frontend, including post creation. absolutely sick and saves compute time.
one bug that i inadvertently found out is that unpublished posts are still accessible via their slugs. it's fine i guess, likelihood of guessing a slug is low and so are the stakes.
love not having to wait for vercel to build after each git push or bandwidth limits on convex for serving media
thinking about the new CMS system. i jokingly teased github's handling of images in markdown in my blog post about it. but I think that is the optimal solution. i really dont care where my files are and this is the closest to containing them within a markdown file as encoding them in b64. i think i will implement such a solution - uploading them to convex file storage and just replacing it with an URL in the editor.
nothing i can do about the blog getting deleted, i have all the images so I guess i will try to recreate the blog posts retroactively?? but thats a weekend project.
just migrated the blog (blog.zimengxiong.com) to use a local convex instance instead of serving a static site from vercel. this means no need to rebuild the site when i make edits to the content since it fetches from convex each time. live at blogv2.zimengxiong.com for now. working on a better cms that uses authentik for auth instead of its own github oauth implementation that blog.zimengxiong.com/admin uses.
already seeing performance improvements from faster image loading. would love to do this on the hosted SaaS convex but their bandwidth and storage limits are too low (0.5/1GB) respectively for multimedia.
wanted to give gemini 3 pro another chance, asked it to do some edits on a simple convex project. it somehow managed to deploy new schema (overwriting) a COMPLETELY UNRELATED convex project convex-feed.zimengxiong.com (sound familiar??) 4+ years of posts are just gone... completely on me for not having backups for stuff like this, but yeah, back to Opus 4.5 it is
will see what I can do to fix this mess
if you are not using `--dangerously-skip-permissions` your wasting untapped potential. letting agents run in the background is the entire point of them. if you are worried setup a sandbox with Mac's `sandbox-exec` or some form of seatbelt, most linux systems have one.
Porting WeDLM to MLX
Tencent's WeDLM paper reports 3x speedup over vLLM using diffusion-style parallel decoding—multiple tokens generated per forward pass while maintaining KV cache compatibility. This post documents an independent port of WeDLM to Apple's MLX framework, covering the architecture decisions, performance optimizations, and implementation details from a week of development on an M4 Max.
[... 6,963 words]Porting Hunyuan3D-2 Texture Generation to MPS
This post documents the process of porting Tencent's Hunyuan3D-2 texture generation pipeline from CUDA to Apple Silicon's Metal Performance Shaders (MPS). The work involved debugging stochastic scheduler behavior, implementing Metal compute kernels for custom rasterization, and integrating with PyTorch's MPS backend.
[... 5,073 words]Porting Meta SAM-3D to Apple Silicon: Custom Metal Kernels and Memory Magic
Tl;DR This post documents the process of porting Meta's SAM-3D Objects (a 12GB foundation model for single-image 3D reconstruction) from CUDA/Linux to Apple Silicon macOS. The work involved rebuilding sparse convolution backends, implementing custom Metal compute shaders, and engineering a sequential model loading strategy that reduced peak memory from 61GB to 17GB.
[... 2,804 words]Excalidash, a whiteboard organizer
The robot was done, the mechanisms were functioning smoothly, and our next match wasn't for another hour. I sat in the pit, laptop open, refining our engineering notebook and staring at yet another Excalidraw tab with a drawing I'd probably never see again.
[... 5,086 words]Unifying my computing
Every day I shuffle between a 15" MacBook Air for school, a Linux workstation at my desk, and a 5080 server for anything that might make my laptop break a sweat.
[... 660 words]Markdown Editors Suck at Handling Files
What has stopped me from using ANY mainstream markdown editors for my blog is because they SUCK at handling images and renaming. They typically fall into a few categories:
[... 1,102 words]Solving picoMini by CMU-Africa
A friend recommended picoCTF to me. I had played similar games like banditOverTheWire before, so I thought it would be pretty fun to try out.
[... 4,734 words]The Macbook Air throttles a lot
Without a fan and with summer slowly ramping up, I often find my M2 MacBook Air 15" heating up and becoming hot to the touch while building Xcode projects or running CAD software. I wanted to measure its actual performance impact and see how effective a simple solution would be.