Anthropic recently published "Emotion Concepts and their Function in a Large Language Model" showing that Claude has internal representations of 171 emotion concepts that influence its behaviour, including driving misalignment like blackmail and reward hacking.
I replicated their methodology on Google's Gemma 4 E4B (open-weight) and released the datasets and code. The repo includes everything needed to build expression probes (detecting expressed emotions) and deflection probes (detecting suppressed emotions), plus an interactive visualiser.
I built a Python implementation of Google's TurboQuant paper (ICLR 2026) for vector search. The key thing that makes this different from PQ and other quantization methods: it's fully data-oblivious. The codebook is derived from math (not trained on your data), so you can add vectors online without ever rebuilding the index. Each vector encodes independently in ~4ms at d=1536.
The repo reproduces the benchmarks from Section 4.4 of the paper — recall@1@k on GloVe (d=200) and OpenAI embeddings (d=1536, d=3072). At 4-bit on d=1536, you get 0.967 recall@1@1 with 8x compression. At 2-bit, 0.862 recall@1@1 with ~16x compression.
I built Sourced because LLMs don't reliably know where a package's source code lives. When they need to understand a dependency, they either guess from training data or fall back to web search, which is slow, expensive, and usually just returns user-facing docs — not the actual implementation.
Being able to read the real source code makes a big difference. The model can see how functions actually behave, verify that the APIs it's calling exist and work the way it thinks, and build implementations that are grounded in reality rather than guesswork.
Sourced indexes every package on PyPI (800k+) and npm (3M+), with new releases picked up within 5 minutes. It exposes three tools via MCP — read, grep, and glob — so your agent can navigate source code the same way you would locally. It's lightweight and fast.
Install: curl -sL sourced.dev/install | sh
Works with Claude Code, Cursor, VS Code, Windsurf, Codex, and others.
I replicated their methodology on Google's Gemma 4 E4B (open-weight) and released the datasets and code. The repo includes everything needed to build expression probes (detecting expressed emotions) and deflection probes (detecting suppressed emotions), plus an interactive visualiser.
Datasets on HuggingFace: https://huggingface.co/datasets/ryancodrai/emotion-probes