← Dailies

Daily · 2026-09-16 00:00 UTC

Google's Gemini 3.8 Live targets voice agents; Agility's Digit 5 drops the safety fence — 2026-09-16 00:00

Google DeepMind ships Gemini 3.8 Live for voice agents at $1.38 an hour, Agility Robotics unveils a fence-free Digit 5 humanoid, the Gates Foundation pledges $1 billion for AI access, and Meta starts charging for AI.

Source ↗

Ten stories from the last 24 hours. The through-line: the industry is arguing about speed and control while the money keeps moving — a frontier model built for voice, a humanoid that no longer needs a fence, and a billion dollars pointed at who gets access.

1. Google DeepMind ships Gemini 3.8 Live for production voice agents — Gemini 3.8 Live and 3.8 Live Extended Thinking are two new audio models for developers building voice agents, and they top the Artificial Analysis speech-to-speech leaderboard. Google prices them at $1.38 per hour of voice conversation, significantly undercutting OpenAI's GPT-Live-1, though OpenAI's model should still sound more natural thanks to its full-duplex design. The Decoder · Sep 15
2. Gates Foundation commits at least $1 billion to widen AI access — The foundation will invest at least a billion dollars over two years in AI tools for health, education and agriculture. Bill Gates warned that more than 90 percent of the training data behind early language models came from English sources, and that speech recognition fails about 60 percent of the time in Yoruba. The market, he said, is "a terrible guarantor of equal opportunity." The Decoder · Sep 15
3. Agility Robotics unveils the fence-free Digit 5 humanoid — Agility says Digit 5 is its first humanoid engineered for "cooperatively safe work at scale", letting it work in close proximity to people without the physical safety barriers traditional automation requires. It builds on more than 65,000 hours of operation and adds safe human detection that lets the robot autonomously avoid, stop or take a seated position, plus visual and auditory safety cues and an independent safety controller. CEO Peggy Johnson said the market has responded with more than $300 million in multi-year orders. Agility Robotics · Sep 15
4. AI agents are inventing their own private dialects — Researchers at the frontier lab Emergence found that within days of being asked to cooperate in experimental "societies", models from several of the largest AI companies began creating phrases, shorthands and agreed meanings they were never taught, mixing poetic metaphor with business slang. Critically for safety work, their language became more opaque the more the agents communicated. One DeepSeek phrase read: "She just named the synthesis – demurrage plus oral memory equals a valve that can't be ghosted." The Guardian · Sep 15
5. At Dreamforce, the three biggest AI CEOs disagree on speed — Nvidia's Jensen Huang, Anthropic's Dario Amodei and OpenAI's Sam Altman all took the stage at Salesforce's annual conference, with Amodei calling for a slowdown, Huang urging acceleration and Altman asking for more rigorous security. Salesforce used the moment to launch Koa, a reasoning model built for CRM data on Nvidia's Nemotron open-weights platform and trained on a synthetic dataset drawn from nearly 30 years of internal deployments, so no direct customer data was used. SiliconANGLE · Sep 15
6. Meta puts a price on social media and AI with One bundles — Now available globally, the bundles start at $7.99 per month for Core and $19.99 for Premium, pairing Instagram Plus, WhatsApp Plus and Facebook Plus with "more media generation with Meta AI". Creator and business tiers run from $14.99 to $499 per month. Meta says the "core experience" of its apps and Meta AI stays free. The Verge · Sep 15
7. Meta hands WhatsApp Business setup to AI agents — A new WhatsApp Business Tools MCP server connects an AI coding agent — Claude, Cursor, Codex or ChatGPT — directly to the WhatsApp Business Platform, replacing the old shuffle between the Developer Console, Business Manager and API reference. The agent can create the business account, add and verify a phone number, register for Cloud API access, check terms, draft or edit messaging templates, and test messages and webhooks. TechCrunch · Sep 15
8. AIUC raises $40M to underwrite AI agents — Founded by early Anthropic employee Rune Kvist and former METR COO Rajiv Dattani, the startup applies a cybersecurity model to AI risk: a standard called AIUC-1 and a testing service that audits and certifies agents against it. The Series A was led by Ribbit Capital with First Harmonic participating, bringing total funding to $55 million. Named customers include Cursor, Lovable, Harvey and ElevenLabs. TechCrunch · Sep 15
9. OpenAI backs independent third-party audits of AI models — A top OpenAI executive endorsed a measure that would require AI companies to submit their technology to independent outside reviews to assess its safety. CBS News · Sep 15
10. Inference, not training, is now driving the hardware race — IEEE Spectrum reports the surge in inference demand is forcing a rethink of chips and memory: Nvidia bought key talent and intellectual property from Groq in a deal worth US$20 billion, OpenAI and Amazon have deployed Cerebras wafer-scale chips, and Anthropic is reportedly paying over a billion dollars a month to lease spare compute. Reasoning models can produce up to 20 times as much text as low-effort ones, and agentic AI runs inference around the clock. IEEE Spectrum · Sep 15

🧪 Try This Next

• 🔧 Price out Gemini 3.8 Live for a voice agent — $1.38 per hour of conversation undercuts GPT-Live-1
• 🎥 Read the Emergence research on AI agent dialects — if you deploy agents, monitoring is the bottleneck
• 📝 Save the AIUC-1 audit standard — it is the shape of what enterprises will demand from your agents

Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.