model checkpoints

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

Audit Finds Batch-Normalization Stats Skew Machine Unlearning Evaluations

A new audit examines 263 publicly released checkpoints that use batch normalization and finds that reported unlearning results shift depending on which version of those statistics is used. Because batch-norm statistics are not produced by gradient updates and are rarely documented in model releases, refitting them on retained data can change the numbers an evaluation relies on. The authors argue that this makes some unlearning verdicts unreliable, since apparent forgetting may reflect checkpoint bookkeeping rather than the removed data genuinely being gone.

unlearning/evaluation/privacy/benchmarks

papersSEP 10 04:00 UTC

KVShareArena: KV-Cache Reuse Across Contexts and Model Checkpoints

A new arXiv paper tackles a limitation in LLM serving systems, which typically reuse KV caches only when shared text appears at the very start of a prompt. The authors target two workloads that break this rule: retrieval-augmented generation servers that assemble different retrieved chunks per request, and requests served across different model checkpoints. KVShareArena aims to expand cache reuse in these cases, reducing redundant computation and improving serving efficiency.

kv-cache/llm-serving/rag/efficiency