Next I’m working on making reviewing large AI-generated PRs easier, but haven’t gotten there yet.
It’s coming along great. I’ve only been working on it since February, and it’s very complete, functionally. I’ve been using an LLM, extensively.
The original server took over seven months to write, but with the LLM, I had it done in about a month. The PHP code is great quality.
It’s not all bunnies and flowers, though. I’m spending the next few days, rewriting a Swift viewcontroller that the LLM wrote. It works, but the code is awful. It looks like it was written by a jargon-addled high schooler on Adderall. I just can’t bear to have that level of junk in my app. I asked it to refactor the code to improve the quality, and reduce the size, and it removed all the documentation and logging.
The good news is, is that I’m so far ahead of schedule (a position that I’ve never been in, in forty years of development), that I can afford to do this.
- spec driven development workflow: https://github.com/sermakarevich/sddw
- fleet, orchestrator for running many coding agents: https://github.com/sermakarevich/fleet
- chunker, which builds smart chunks of the documents with hierarchical organization: https://github.com/sermakarevich/chunker
- ai knowledge wiki, a collection of ai papers and articles organized in hierarchical Wiki
- hierarchical organization of all passed Kaggle competition solutions (write ups and notebooks) - https://github.com/sermakarevich/kaggle_wiki
- claude code workflows, plugin to build custom claude code workflows based on ssd idea: https://github.com/sermakarevich/ccw
I sometimes need to have a quick but realistic model of an optical system without paying a few thousand for some of the well known commercial offerings, so I've been building this.
I just implemented collapsing comments and root/next/prev/context nav links.
You can read the entire changelog here: https://github.com/shawwn/sharc/tree/main/docs/agents/handof...
I'm working toward making a Hacker News simulator in the vein of https://reddit.com/r/subsimulatorgpt2 (but using HN's code instead of a subreddit).
In theory, it's really not suited for this because (at least back then) this was designed for computers with:
- relatively large amounts of RAM (uncompressed color+heightmap data takes a lot of space),
- fast CPUs with data caches (it's a 2D trapezoidal walk of a top-down projected fustrum with plenty of additions, multiplications and divisions),
- memory-mapped linear framebuffers (for software rendering).
The PlayStation has none of these things: only 2 MiB of main RAM; an in-order scalar 33 MHz MIPS III processor without a FPU or a data cache, where any CPU load from main RAM stalls the pipeline for 5 cycles, multiplications have a 6 to 13 cycles latency and divisions have a 36 cycles latency; and the only way to touch VRAM is to send commands to the 2D GPU.
What it does have is 4 KiB of I-cache (direct-mapped), 1 KiB of scratchpad with no wait states and a fixed-point GTE coprocessor which is mostly geared towards projecting 3D points onto a 2D screen and not general matrix/vector processing. Meaning that if I get my hot loop within 1024 instructions (and no function calls), fit my working set+stack within 1 KiB and can pipeline the GTE for transformations and the CPU for map scanning/GPU submissions, it might just work.
So far I'm getting decent framerates (like 10-15 FPS at 256x240) with just software projection and GPU line rendering. Right now I'm wrestling with the GTE and various ways to trick it into transforming more points than it's supposed to be able to do, by looking at the fixed-hardware math equations for the various instructions and trying to fit my equations into them (doing stuff like putting relative altitude into X/Y vector registers instead of coordinates and massaging the rest of the values to get screen Y coordinates out of it).
Kind of a hardcore topic to pick for my first homebrew ever, but I craved a simpler low-level challenge after ghidra-delinker-extension.
ifh-hn•1h ago