The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
Nokia’s AnyJev turns open LLMs into typed decision models with no training — Apache-2.0 on PyPI — cutting order-flip rates and lifting how much traffic you can auto-decide at a fixed error budget.
Kyutai’s Voice of Reason applies reinforcement learning to speech-native math — lifting GLM-4-Voice from 27.3% to 77.1% on spoken GSM8K without a separate text LLM in the loop.
OpenAI fills the GPT-6 mid-tier with Sol and Luna — half the prior API rates versus 5.6 promo pricing, stronger caching, and a cost-per-task pitch aimed squarely at Opus and Fable.
NVIDIA’s Singapore AI Day is less a keynote and more a regional map: Nemotron localization, public-sector agents, and Sea Limited’s first ASEAN Vera Rubin adoption.
Airbnb expands engineering and product access to GPT-6 Astra and OpenAI frontier models via API and Bedrock — building on Codex and an internal assistant already shipping more features.
After math PR disasters, OpenAI backs a nine-person elite advisory group — independence claimed, influence unproven, and 100-plus claimed results waiting in the wings.
OpenAI and Grab’s GO Forward with AI aims to train 30,000 driver, delivery, and merchant partners — practical skills, not another chatbot demo.
Opus 5.5 and GPT-6 Sol/Luna aren’t moonshots — they’re price cuts with modest gains as enterprises shop the frontier like a catalog.
OpenAI’s GPT-6 caching upgrades are infrastructure for agents that run for hours — cheaper tokens, faster turns, and tools to see when reuse breaks.
Snapdragon 8 Elite Gen 6 and Extreme Gen 6 push on-device agents — local scribes, speaker ID, and a 30B MoE on the Extreme SKU — as phones stay the default AI surface.
EvilTokens turned inbox compromise into a subscription product — AI for target picking and lure drafting — until Microsoft and partners tore down the shop.
British Columbia wants OpenAI to fund Tumbler Ridge’s new school and force chat-log disclosure — arguing apology isn’t a remedy after ChatGPT-linked planning.