Peter pszemraj

Semble — local code search (hybrid keyword + semantic)

semble is a search index over this repo. It splits the code into chunks and ranks them by a blend of keyword (BM25) and semantic-embedding similarity, then returns the top matches with their file_path and line. It is a retrieval tool, not an analysis tool: it finds existing code that resembles what you describe. It does not reason, judge, or summarize. Think "show me the code most like this description," never "answer this question about the codebase."

Invocation: semble is installed only in the misc conda env. Always run it as conda run -n misc --live-stream semble <args>. Never call semble bare; never use uvx. The short query examples below drop the prefix for readability — prepend it whenever you actually run one.

How to query

Applying Mamba-3 Ideas to an OLMo-Hybrid-Style Gated DeltaNet

Recommendation

The best target architecture is not a Mamba-3 hybrid. It is an OLMo-Hybrid-style interleaved Gated DeltaNet model with Mamba-3's hardware-aware recurrence upgrades ported into the GDN mixer.

The highest-value variant is:

Hybrid MIMO-HGDN: [GDN, GDN, GDN, Attention] × N, with negative-eigenvalue GDN preserved, rank-R MIMO updates added to the GDN state update, short convolutions retained for the first ablation, and optional data-dependent q/k rotations as a lower-priority expressivity experiment.

Decorative Monospace

A formatting failure mode in modern LLM outputs where code blocks are used as a rhetorical device for signaling rigor rather than as semantic markup for content where whitespace, syntax, or literal reproduction matters.

The model isn't trying to communicate "this is code." It's trying to communicate "this is structured thinking" by borrowing the visual texture of a technical document. The texture is decorative; the underlying content is prose.

	// ==UserScript==
	// @name Quiet media alongside Dark Reader
	// @namespace local.quiet-media
	// @version 1.0.0
	// @description Show still images instead of animated images and looping previews; dim mostly-white images.
	// @match https:///
	// @match http:///
	// @run-at document-start
	// @sandbox DOM
	// @grant none

	# /// script
	# requires-python = ">=3.10"
	# dependencies = ["joblib", "rich"]
	# ///

	"""Measure Codex generation speed (output tokens per second) from local session logs.

	Reads the rollout JSONL files that Codex CLI and Codex Desktop write under
	$CODEX_HOME/sessions and $CODEX_HOME/archived_sessions. Nothing is sent
	anywhere and no message content is kept.

	"""Back up your Weights & Biases projects to local disk, read-only.

	Nothing here creates, changes, or deletes anything on the W&B side. It only uses read calls from the public API, authenticated with your existing `wandb login` (or WANDB_API_KEY). The script is a single self-contained file: the parts of termfmt it uses for terminal output are inlined near the top.

	Subcommands:
	probe Show the account, its teams, and every project each entity can see, testing that runs are readable through the API (the web UI's plan lockout does not apply to the API).
	size Count runs, artifact versions and artifact bytes per project before committing to an export. Add --files to also total run-file sizes (one extra request per run).
	export Save everything: configs, summaries, full unsampled metric history and system metrics for every run, plus run files (logs, media, wandb-metadata.json) and artifact contents. --no-files and --no-artifacts leave those two out.

	What export does, in order:

	#!/usr/bin/env python3
	# /// script
	# requires-python = ">=3.10"
	# dependencies = ["mlx-audio==0.5.1", "numpy", "soundfile>=0.13.1", "soxr"]
	# ///
	"""Transcribe a long recording with MLX VibeVoice-ASR into grouped Markdown.

	uv run --python 3.12 vibetranscribe.py meeting.flac --dry-run
	uv run --python 3.12 vibetranscribe.py meeting.flac -o meeting.md
	uv run --python 3.12 vibetranscribe.py meeting.flac \

	#!/usr/bin/env python3
	"""
	find_flash_attn_wheel.py — find the right prebuilt flash-attention wheel from
	https://github.com/mjun0812/flash-attention-prebuild-wheels

	Detects (or takes via flags) your Python / PyTorch / CUDA / platform, fetches the
	repo's full wheel index (doc/packages.md), and prints every matching wheel with
	a recommendation and a ready-to-run `pip install <url>` command.

	Usage:

	<#
	.SYNOPSIS
	Finds every file type and protocol currently assigned to Microsoft Edge
	and reassigns them all to your preferred browser using SetUserFTA.

	.DESCRIPTION
	Windows protects file associations with a hash -- you can't just edit the
	registry. SetUserFTA reverse-engineers that hash so it can set associations
	from the command line.