Opening

Hacker News 150 points, 67 comments, someone actually statically recompiled three 32-bit Windows executables from Black Ops II to native ARM64 code, hitting 51–60 FPS on M1 with zero Wine, zero Rosetta 2, zero JIT. I’ll explain how it hands the copy-protection to the Unicorn interpreter for instruction-by-instruction execution, plus a tutorial-grade project that implements the full DeepSeek-V3 architecture in 190 lines of pure JAX.

Today’s Open Source Gems

1. bo2-mac: Static recompilation runs Black Ops II natively on Apple Silicon

• What it is: Statically recompiles the three Call of Duty: Black Ops II executables (campaign, zombies, multiplayer), translating x86-32 machine code to native ARM64 at compile time. Custom Win32 runtime converts Direct3D 11 to Metal, XAudio2 to CoreAudio, video via libvpx decode. Zero reliance on Wine, Rosetta 2, or JIT. M1 real-world testing shows all three modes above 50 FPS.
• Recent buzz: Hacker News 150 points, 67 comments, GitHub 252 stars, GPL-3.0 license.
• Standout features: Copy protection not bypassed, instead executed instruction-by-instruction via Unicorn x86 emulator. Translation faults fall back to the same emulator, then return to native code. DXBC shaders converted to Metal via HLSLcc. Guest memory mapped as a single 4 GiB region. Threads start as coroutines then migrate to real threads with memory barriers.
• Quick start: Apple Silicon Mac, macOS 14+, 20 GB free space, Homebrew, Steam account owning the game. git clone https://github.com/adrbn/bo2-mac.git && cd bo2-mac && brew install llvm@17 cmake ninja libvpx unicorn && pip3 install capstone && scripts/make_app.sh --from-source, _{1 hour 45 minutes total; campaign playable in} 45 minutes.
• vs. alternatives: Unlike CrossOver, Whisky, Parallels (runtime x86→ARM translation), DXVK, DXMT (D3D11→Vulkan/Metal on Wine), this project compiles once to native ARM64 at build time. Zero Wine. Custom D3D11→Metal layer running on 32-bit guest memory.
• For whom: Gamers wanting native Black Ops II on Apple Silicon; engineers researching static recompilation, compilers, graphics, reverse engineering.
• Link: https://github.com/adrbn/bo2-mac

2. deepseek-v3-jax: Full DeepSeek-V3 architecture in 190 lines of pure JAX

• What it is: Complete DeepSeek-V3 architecture implemented in 190 non-empty, non-comment lines of pure JAX. Includes Multi-head Latent Attention (cache reduced from 480 to 144 numbers/token/layer), Decoupled RoPE, 1 shared + 16 routed experts (top-4), aux-loss-free bias balancing, sequence-level balance loss, Multi-token prediction (λ=0.3). Parameters as pure dicts. All modules as top-to-bottom readable functions.
• Recent buzz: Hacker News 237 points, 23 comments, GitHub 167 stars, MIT license.
• Standout features: Core model deepseek_v3.py in just 190 lines. bf16 compute with f32 master weights. Prefill and decode share MLA cache. 39 tests included (causal mask, cache consistency, routing, balancing, MTP target, NaN handling, bf16 master weights). Trainable from scratch on a MacBook Air.
• Quick start: Apple Silicon Mac + Python 3.12 + uv. git clone https://github.com/Vaibhaav-Tiwari/deepseek-v3-jax.git && cd deepseek-v3-jax && uv venv -p 3.12 .venv && uv pip install --python .venv/bin/python -r requirements.txt && JAX_PLATFORMS=cpu .venv/bin/pytest tests.py -q verifies 39 tests pass. Download released ckpt.pkl, tokenizer.json then .venv/bin/python sample.py --prompt "The fog came in off the" --num 3 to generate.
• vs. alternatives: Unlike nanoGPT, minGPT (educational dense Transformers), Hugging Face Transformers (production-grade but thick abstraction layers), mlx-examples (Apple MLX only), this project uses minimal deps (pure JAX, no Flax/Equinox/PyTorch) to fully expose DeepSeek-V3’s unique MLA, MoE balancing, and MTP. Suited for architecture research, not production inference.
• For whom: ML engineers wanting deep understanding of next-gen LLM architecture (MLA, MoE balancing, MTP).
• Link: https://github.com/Vaibhaav-Tiwari/deepseek-v3-jax

Closing

Both projects today practice "minimal dependencies, maximum transparency" to dissect brutally hard systems problems for engineers: bo2-mac compiles a 32-bit Windows game entirely to native ARM64; deepseek-v3-jax explains the core architecture of a 671B-parameter model in 190 lines of JAX. If you want to see real systems engineering chewing through hardware abstraction layers with no source code and no docs, bo2-mac’s thirteen Findings are mandatory reading. See you tomorrow.