Sampling Parameters as Substances: An Honest Analogy

lleqsnoom — Mon, 14 Sep 2026 06:11:00 +0000

Sampling Parameters as Substances: An Honest Analogy

A strange question that works: why does cannabis seem to make people more creative?

It doesn't make anyone smarter — it changes filtering. The mind lets through associations it would normally reject. Distant ideas reach awareness more easily. That feels like creativity from the inside.

Then someone asked a better question: so it's like turning up a model's temperature?

Yes. And that opens a door. Every sampling parameter in a language model sits on the same trade-off — how boldly the model picks its next word — and each one has a rough equivalent in how a mind works.

The one trade-off behind everything

A language model doesn't "know" what to say. At every step it produces a probability for every token it could write next, then picks one.

Pick the most likely word every time and the text is safe, correct, boring. It loops.
Pick something unlikely often enough and the text is surprising — sometimes brilliant, sometimes nonsense.

Every parameter below just moves the needle between those two poles.

The LLM pharmacopoeia

Max tokens is the odd one out — it's a budget, not a mood. It never touches which token gets picked, only how many are allowed.

Why this matters now

One caveat, because it's the most interesting part: the dial is disappearing. The newest reasoning models expose no temperature at all. The industry stopped turning the knob up and started spending the compute on thinking instead.

The full article develops the analogy parameter by parameter, with a careful caveat up front: these are analogies, not neuroscience. No substance has a top_p value, and a model is not a mind with loose filters. The comparison earns its place because a person and a model both face a version of the same decision — choosing among thousands of possible next steps, and deciding how much to trust the obvious one.

→ Read the full post on mielony.com.

Path-Tracing Doom on macOS: Metal Ray Tracing on Apple Silicon

lleqsnoom — Sun, 13 Sep 2026 14:30:45 +0000

There is no Vulkan ray tracing on macOS. Apple ships Metal, so a Vulkan RT app runs only through MoltenVK — and MoltenVK's ray tracing is an experimental, unmerged pull request, not an installable driver. On that missing layer, the 1993 game's path tracer runs: every pixel traced per frame, direct and bounced light, reflections, soft shadows — at ~44 FPS at 800p on an M4 Max.

The short version

I ported path-traced Doom 2 to Apple Silicon on a patched MoltenVK with Metal RT. It translates SPIR-V to Metal Shading Language and runs the same RTGL1 path tracer and Doom port as the Intel Arc article. The port ships as a self-contained zip (game + path tracer + patched MoltenVK + SDL); you supply Doom2.wad.

Download + every patch: https://mielony.com/blog/path-tracing-doom-on-macos/

Why macOS is different

On Intel Arc the bug class was "assumptions about the device." macOS goes further — there is no Vulkan loader at all. The game links MoltenVK (Vulkan-on-Metal), which stops at Vulkan 1.2 graphics/compute. Ray tracing (VK_KHR_acceleration_structure, VK_KHR_ray_tracing_pipeline) exists only on an unmerged MoltenVK PR (#2771), which maps the Vulkan RT entry points onto Metal behind MVK_CONFIG_ENABLE_EXPERIMENTAL_RAY_TRACING=1.

What broke, layer by layer

Driver. MoltenVK fails vkCreateDevice if any requested feature is unsupported, and the path tracer asks for its full set up front — the patch clears every feature bit the device does not report. SPIRV-Cross's MSL backend lacked Metal types for ray-tracing built-ins like gl_LaunchIDEXT, so every ray-generation shader failed to compile until a +21-line patch added them.
Renderer. An sRGB swapchain double-encoded the already display-ready image; a missing barrier loaded stale tiles — 8×8 and 16×16 artifacts on a tile-based Apple GPU.
Game. Metal surface from SDL_Metal_CreateView, true fullscreen, HighDPI matching the panel backing store (3456×2234), and a green FPS: n HUD counter.

Performance

Doom 2 demo1 on the M4 Max: 43.8 FPS at 800p, 51.7 FPS at 450p internal. rt_renderscale sets internal render height, and RT cost scales with it — not the window. For comparison, the Intel Arc A770 reaches 90 FPS at 450p internal on the same renderer.

Should you bother?

Only if the Mac is the point. It runs 44–52 FPS at low internal resolutions, has no music, stands on an open unmerged PR, and needs a commercial Doom 2 IWAD plus a ModDB lighting addon. If nothing ties you to macOS, Arc/Linux or an RTX machine is smoother — same renderer either way.

Full write-up with benchmarks, build scripts and every patch: https://mielony.com/blog/path-tracing-doom-on-macos/

DEV Community: lleqsnoom

Sampling Parameters as Substances: An Honest Analogy

Sampling Parameters as Substances: An Honest Analogy

The one trade-off behind everything

The LLM pharmacopoeia

Why this matters now

Path-Tracing Doom on macOS: Metal Ray Tracing on Apple Silicon

The short version

Why macOS is different

What broke, layer by layer

Performance

Should you bother?