residual-stream

LIVE PULSE

4.0 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.2 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.8 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src4.0 Anthropic CEO Amodei calls for slower AI development and shared safety rules — 11 src2.2 Agility Robotics unveils Digit 5 humanoid for warehouses and factories — 2 src2.0 Apple ships rebuilt Siri with Google Gemini, but not in the EU — 2 src1.8 Siri AI in macOS 27 Golden Gate: FAQ, Germany availability, privacy questions — 2 src1.4 Sam Altman says OpenAI will not go public in 2026, citing AI safety concerns — 5 src1.1 OpenAI contractors review real ChatGPT conversations to rate responses, report says — 2 src1.1 Anthropic data retention policy prompts firms to limit Claude use for sensitive work — 1 src1.1 Study examines issue bias in LLMs used as writing assistants before Swedish 2026 election — 1 src1.1 Study Audits Misalignment in Multi-Modal World Models — 1 src1.1 Retrieval-Grounded Reasoning Approach Proposed for Universal Multimodal Embeddings — 1 src

topic2 events

papersSEP 12 04:00 UTC

Paper compares diff-in-means and INLP for finding refusal directions in LLMs

A preprint revisits the finding that refusal behavior in safety-tuned chat models is controlled by a single linear direction in the residual stream, which can be recovered by taking the difference in means between harmful and harmless activations. The authors compare this diff-in-means approach with INLP, an iterative nullspace projection method, to see whether a one-direction account holds up. The work is presented as a preliminary comparison of the two techniques.

safety/interpretability/llm/alignment

papersSEP 11 04:00 UTC

RDQ quantization method targets accuracy loss below 4-bit in LLMs

A research paper proposes Residual Distribution Quantization, a post-training quantization approach for large language models. The authors attribute the sharp accuracy drop seen below 4-bit precision to distributional drift in the residual stream, where quantization error introduced at each transformer layer builds up in the shared representation. Their method aims to correct this accumulated error to preserve model quality at lower bit widths.

quantization/llm/research/efficiency