Wednesday, June 10, 2026
6 stories covered by gentic.news intelligence
Stanford and Meta's "Code as Agent Harness" paper proposes code-driven AI agent orchestration, potentially improving reliability over natural language prompts.
Nvidia acquired Kumo AI for $400M+ to bring foundation model predictions to enterprise relational data, filling a gap left by LLMs.
Anthropic's Mythos Preview AI built 8 working exploits from Firefox and Windows kernel patches within hours. The first exploit was ready 18 days before the patched Firefox shipped.
Claude Desktop spawns a 1.8 GB Hyper-V VM on every Windows launch due to 2,689 stale session files, consuming 11% of RAM.
KV cache quantization silently breaks LLM safety alignment, with Mistral-7B losing 15.2% refusals at 1.03x perplexity. PCR diagnostic recovers up to 97% alignment in 35 GPU-minutes.
Visual-SDPO uses visual-feedback self-distillation to improve code-generated visual artifacts by >10 points on ChartMimic, Design2Code, and AeSlides, with no added inference cost.