Daily digest
6 items · ~6 min · Week 2026-W28
Worth knowing (2)
Vidu S1: A Real-Time Interactive Video Generation Model
ShengShu TechnologyVidu S1 is a real-time interactive video generation model using an autoregressive-diffusion hybrid architecture that produces 540p video at up to 42 FPS on consumer GPUs with no drift or visual artifacts. Users control an AI avatar through voice input, and the system generates synchronized facial expressions, gestures, and full-body motion from a single uploaded image, with infinite-length output.
Claude Code Desktop gains built-in browser in Week 28 update
AnthropicAnthropic's weekly Claude Code digest for July 6–10 (Week 28) introduces an in-app browser for the Claude Code desktop app, letting Claude pull up documentation, designs, or any website and interact with pages the same way it does with local dev server previews. The release also upgrades the /doctor command into a full setup checkup that can diagnose and auto-fix configuration issues, and tightens auto mode with safeguards against transcript tampering and destructive rm -rf commands on unresolved variables.
For reference (4)
Super Weights in LLMs and the Failure of Selective Training
Amazon Web ServicesThis paper challenges the hypothesis that 'super weights' — a small set of parameters whose removal catastrophically degrades model performance — can be selectively trained to update model behavior. Targeting even 100–8,192 such parameters in isolation causes accuracy to collapse to random-guessing on OLMo-1B and OLMo-7B, while training an equal number of random parameters in the same layers actually improves performance. Accepted at COLM 2026.
IdeaGene-Bench: Benchmarking Scientific Lineage Reasoning and Idea Generation
Shanghai Jiao Tong UniversityIdeaGene-Bench introduces a benchmark representing each scientific paper as typed 'Idea Genome' objects with inheritance, repair, and recombination relations to prior work. The benchmark contains nearly 2,000 lineage traces across 10 scientific domains. The best-performing system reaches only 27.3% exact accuracy on lineage reasoning, revealing a major gap in current LLMs' ability to track how ideas evolve.
llama.cpp b9967–b9969: Adreno GPU Acceleration and OpenAI-Compatible Null Sampling
Three llama.cpp builds landed on July 12: b9968 adds OpenCL int8 dp4 dense-matmul and MoE prefill kernels for Qualcomm Adreno GPUs, improving inference throughput on Snapdragon-powered Android and Windows ARM devices; b9969 fixes a Vulkan path that routed large matmuls to the wrong tile size on Adreno, causing crashes with long prompts on quantized models; b9967 allows null values in sampling parameters so clients can explicitly request server defaults, aligning with the OpenAI sampling spec.
GitHub Copilot Adds Prompt-Injection Detection in CodeQL 2.26.0
GitHubCodeQL 2.26.0, shipped July 10, adds Kotlin 2.4.0 language support and introduces new query coverage targeting AI-specific security threats, specifically detecting prompt injection vulnerabilities in code that passes untrusted input to LLM APIs. GitHub Copilot's Mobile sessions interface also received filtering and sorting controls.