7:35 PM
Today in brief.
Perplexity open-sources Lily, the Mac hybrid-compute engine
Yesterday hybrid compute was a product feature. Today the local inference engine behind it is open source for Qwen on Apple silicon.
3 sources
🤖 AI
Gemini 3.8 Flash and Flash Cyber land, Fairwind gates the cyber model
Google ships its third Flash in six weeks: a same-price smarter workhorse plus a defender-only cyber model that patches near frontier quality at Flash cost.
+48 sources 2:39 PM
🔧 Tech
Runway ships Dev MCP so coding agents can drive its API
Runway exposes Model Routers, tasks, and debug hooks through an MCP server your coding agent can call without leaving the IDE.
1 source
🤖 AI
Meta Muse Spark 1.3 hits the frontier coding pack
Meta Superintelligence Labs ships Muse Spark 1.3: Wang says it is competitive with Fable 5.1 and better than GPT-5.6 Sol on coding, with fewer tokens and Contemplating multi-agent mode.
🔧 Tech
AI answer engines are citing fake best-of pages and missing numbers
Two audits land the same day: 215k manufactured software pages feeding Perplexity, and about a third of Perplexity number citations that do not contain the number.
🏛️ Policy
US government sides with OpenAI in the NYT training fight
The Trump administration backs OpenAI against the New York Times: training on copyrighted material is not infringement, which would reshape every lab that scrapes the open web.
🏛️ Policy
Six curl CVEs after OpenAI and Anthropic found zero
Aisle re-tested curl after frontier labs reported clean: six new CVEs. AI vuln hunting still misses what a focused human pass catches.
No stories in this category for this edition.