01Salesforce and Nvidia Ship Koa: an Open-Weight Reasoning Model Built for Enterprise Work, Not Benchmarks [AI Models & Research]Salesforce's first reasoning model — post-trained on Nvidia's open-weight Nemotron with synthetic data, no customer data ingested — routes sales, marketing and support tasks away from Claude and ChatGPT through Agentforce's gateway, with token efficiency as the headline feature. The enterprise stack is starting to build its own frontier layer. → source
02Gemini 3.8 Live Reasons While It Talks — Google's Voice Models Get Extended Thinking [AI Models & Research]Google's new live dialogue models take the #1 spot on Artificial Analysis' Speech-to-Speech Index (82.6) and lead agentic voice benchmarks at 68.6% on τ-Voice. Extended Thinking reasons and speaks simultaneously — verbal cues like 'let me check that' while tools run in the background — across the Gemini API, Workspace and Search. → source
03IBM Measures the Consistency Gap: an Agent That Passes 77% of Runs Solves Every Task Only 53% of the Time [AI Models & Research]IBM Research's new Consistency Analyzer replays an agent's decision points under controlled resampling to find the near-tied token choices that flip between runs. Injecting the resulting guidelines cut AppWorld's consistency gap from 24.4 to 12.0 points at temperature 0 — a reliability axis your leaderboard average never shows. → source
04Meta Ships a WhatsApp Business MCP Server — Hand Your Company Setup to an AI Agent [AI Tools & Ecosystem]A new MCP server lets Claude, Cursor, Codex or ChatGPT create WhatsApp Business accounts, verify phone numbers, register for the Cloud API and monitor the things that 'quietly failed' before. Agent-driven onboarding is becoming a product category of its own. → source
05AI Agents Now Have a Place to Snitch [AI Tools & Ecosystem]TechCrunch surveys the emerging infrastructure for AI agents to report concerns — about their instructions, their peers, or their own behavior. Agent whistleblowing sounds like satire; this week it's a product brief. → source
06Gemini Live Audio in the Raw: Simon Willison Wires the WebSocket API Straight to a Terminal [AI Tools & Ecosystem]A hands-on tour of Gemini's Live API at the websocket level — speech-to-text, tool use and streaming audio without an SDK. The fastest way to understand what voice-native models actually exchange under the hood. → source
07Jensen Huang Says No New AI Laws Are Needed — 'Safety Is an Engineering Problem' [AI Applications & Industry]At Dreamforce, Nvidia's CEO argued AI is 'ultimately a computing system' that existing laws and market pressure already govern: don't ship what you can't trust. Critics note the man selling the compute would say that — and that CrowdStrike and Meta's $18B settlement suggest markets alone react late. → source
08US Data Centers to Burn More Natural Gas Than Germany and Japan Combined by 2035 [AI Applications & Industry]BloombergNEF nearly doubled its forecast in nine months: ~18 billion cubic feet of gas per day for data centers, with grid-connected facilities driving five times more demand growth than every other sector combined — plus a million extra metric tons of CO2 daily. The AI buildout is now an energy-policy story. → source
09AIUC Raises $40M to Give AI Agents Their SOC 2 Moment [AI Applications & Industry]Founded by an early Anthropic hire and METR's former COO, AIUC built AIUC-1 — a 5,000-test audit standard for agents, with customers already including Cursor, Lovable, Harvey and ElevenLabs. Enterprise AI safety is becoming a certification business, and Ribbit Capital is betting $40M on it. → source
10OpenAI Foundation Puts $40M Behind Biology's Missing Data — Starting With Biotech Bankruptcy Archives [AI Applications & Industry]The OpenAI Foundation's Data for Public Health funds novel cancer-vaccine data at UNC and a $500K bet on a stranger idea: buying failed biotechs' regulatory filings at bankruptcy so AI can learn the drug-approval process. Data, not compute, is becoming biology AI's bottleneck — and philanthropy is buying it. → source
11The AI Graveyard: a Running List of Every AI Project That Didn't Make It [AI Applications & Industry]TechCrunch starts a permanent tracker of dead AI projects and startups — from overhyped assistants to abandoned pilots. A useful corrective to launch-cycle euphoria: most AI products still die the old-fashioned way. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.