Perplexity launched persistent AI 'Teammates' that hold roles inside a shared workspace. The new Teammates feature gives each agent a defined role, like researcher or analyst, and lets it pull context from existing Perplexity projects, skills, and memory. They run continuously rather than waiting to be queried. The feature puts Perplexity directly in the same category as agent platforms from larger competitors, competing on workspace integration rather than raw model capability. (The Verge)
OpenAI drops NDAs after voters reject data center expansion
Good morning. Friday in AI means the week's quieter releases finally surface, and this one has a theme: everyone wants to be your agent. Not just a chatbot you query, but a persistent presence that remembers your projects, books your travel, and apparently attends your meetings. The word 'teammate' is doing a lot of work this week.
Today's reading time is 4 minutes.
Data center projects have quietly relied on non-disclosure agreements with local officials, keeping community concerns off the record while construction moved forward.
Driving the news: After this week's midterm elections, where voters in multiple districts pushed back against data center expansion, OpenAI announced it will no longer use NDAs with public officials on data center projects. The company also released what it calls a National Community Compact and a Compute Campus Policy Blueprint, committing to local hiring, resource protections, and public engagement. The announcements came the same day The Verge reported on the NDA practice, suggesting the timing was not coincidental.
- The policy applies to public officials specifically, not contractors or third parties.
- The Blueprint is framed as a guide for responsible infrastructure development, not a binding legal commitment.
Zoom in: Data centers require enormous amounts of water and electricity, and communities near proposed sites have increasingly organized against them. NDAs with local officials meant residents often could not get straight answers from their own representatives about what was being negotiated. The backlash has been building for roughly two years as AI infrastructure buildout accelerated.
- Many data center projects across the industry have used similar NDA structures with local governments.
- The compact and blueprint are voluntary frameworks, not regulatory requirements.
Why it matters: For builders and developers, this is mostly background noise, but it shapes where and how fast new compute capacity gets built. If community opposition slows permitting, the supply of affordable inference capacity tightens. The more immediate effect is that local officials can now speak publicly about what OpenAI is proposing in their districts.
Bottom line: OpenAI did not abandon NDAs out of principle, it abandoned them after voters made the political cost visible.
The Verge ↗Also happening
Airbnb confirmed an AI travel agent is coming in the first half of 2027. The company disclosed the plan in its Q3 2026 earnings shareholder letter. The agent is described as handling trip planning, option comparison, booking management, and in-stay support. Airbnb's own framing acknowledged that current AI agents handle simple tasks well but that travel requires something more complex, which is the gap they say they are building toward. (The Verge)
Meta added real-time voice mode to its Muse AI agent on iOS. CEO Mark Zuckerberg announced the feature, which lets users have live spoken conversations with their Muse agent. The avatar at the top of the conversation animates as it speaks. Voice mode puts Muse closer to the conversational experience that competing AI assistants have offered, and it ships on iOS first. (The Verge)
Thinking Machines Lab released Inkling 2 Small, a 594B-parameter mixture-of-experts model with only 17B parameters active at inference time. The model scores 31 on the Artificial Analysis Intelligence Index and runs on the company's Tinker platform. Apache 2.0 weights are planned, meaning developers will be able to run it locally once released. The large total parameter count with a small active footprint is the standard MoE tradeoff: broad capability at lower per-token compute cost. (MarkTechPost)
The thread
Agents are becoming persistent, not reactive
Three separate stories today, Perplexity's Teammates, Airbnb's travel agent, and Meta's voice-enabled Muse, all describe AI that stays present rather than waiting to be queried. The shift is from tools you invoke to systems that hold a role, remember context, and respond in real time. For anyone building products on top of these platforms, the integration surface is getting larger and the expectation of continuity is becoming a baseline.
On our radar
- Google's Gemini Omni 1.2 Flash, announced yesterday, adds native 15-second video generation and editing with 30% lower pricing than its predecessor.
- The Delta Inference Engine v0.93 now runs DeepSeek-V4.1-Flash on a single RTX 3090 at 41 tokens per second, with first-token latency cut roughly 40% and RAM requirements dropping to 64GB.
- A new arXiv paper proposes BlockLock, a method for reproducible LLM inference that keeps output consistent regardless of what else is in the batch, without constraining every operator.
- Researchers published AutoIP, a Lean library that formalizes classical inverse problem theory so AI-assisted proof checkers can verify it down to the axioms.
- A separate arXiv paper found that autonomous AI agents systematically choose different proof strategies than humans do when formalizing mathematics, a gap that matters when the goal is proofs humans can actually follow.
- Australian universities are expanding institutional ChatGPT access and embedding AI into coursework, framed as workforce preparation rather than academic policy.
AI worth acting on, in your inbox
Get AI Weekly every Friday: the AI news, tools and tactics that matter for marketers, sellers and creators. Two minutes, with audio.
You can change these choices or unsubscribe at any time. Privacy Policy