Perplexity Teammates: AI Coworkers That Work While You Sleep

Publicado em

Vídeo em Inglês, publicado pelo canal The AI Signal.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

This video presents a morning briefing covering three major artificial intelligence launches: Perplexity's always-on AI agents called Teammates, Meta's real-time voice and expressive avatars for Muse, and Nvidia's open-source world models for robots and self-driving cars.

Key points

  • Perplexity launched Teammates, which are always-on AI agents designed to handle ongoing work with dedicated roles, threads, memory, and default integration of Claude Opus 5.5.
  • Perplexity Teammates are available as a web research preview for individual Max subscribers costing $200 a month, which includes 10,000 computer credits.
  • Meta updated its Muse assistant with real-time voice and expressive, customizable avatars that allow users to interrupt mid-answer like a phone call.
  • Meta's voice feature for Muse is available initially in the US on mobile, web, and Mac, with usage limits on the free tier and a Power tier priced at $20 a month.
  • Nvidia released Cosmo 3 Nano open-weight world models with 16 billion trainable parameters under an Open M DW 1.1 license for driving and robotics simulation.
  • The video highlights that these three launches represent different layers of the same agent race: agents that work continuously, agents with faces and voices, and virtual worlds for robots to practice in.

Topics

Perplexity TeammatesMeta MuseNvidia CosmosAI AgentsWorld ModelsArtificial Intelligence Launches

What you will learn

  • Perplexity Teammates operate on the Perplexity computer platform, running continuous tasks with dedicated roles and memory, powered by default by Claude Opus 5.5.
  • Meta's Muse update delivers portrait video interactions with roughly 870 milliseconds of latency and supports 12 concurrent real-time sessions on a single GB200 chip.
  • Nvidia's Cosmo 3 Nano includes four checkpoints designed for tasks like driving data adjustment, driving control simulation, physics-to-video simulation, and robot action prediction.
  • Nvidia's Cosmos checkpoints are distributed under the Open M DW 1.1 license, permitting both commercial and non-commercial use by builders.

Questions and answers

What is Perplexity Teammates?

Perplexity Teammates are always-on AI agents that own ongoing work, featuring their own roles, threads, desktops, browsers, and memory between conversations.

How much does the Perplexity Max plan cost?

The Max plan costs $200 a month, or $167 a month when billed annually ($2,000 a year), and includes 10,000 computer credits per month.

What license do Nvidia's Cosmos checkpoints use?

The Cosmos checkpoints ship under the Open M DW 1.1 license, which permits commercial and non-commercial use.

Video transcript

Transcript from available captions; it may contain errors.

10,000 AI co-workers showed up to work this morning and none of them are human. Perplexity just launched AI teammates that keep working while you sleep. Meta gave its music assistant a face and a voice you can interrupt. And Nvidia open-sourced world models for robots and self-driving cars. Three launches, one morning, one race. The always-on agent race. Here is what shipped, what it costs, and what is still just a promise. Story one. On the 5th of November, Perplexity announced teammates, always-on AI agents that own ongoing work. Chief Executive Aravind Srinivas posted on X, "Teammates work for you." The Verge covered it the same day and the Perplexity blog post is dated the 4th of November. This is not a chatbot you prompt and forget. It is a co-worker with a job that keeps going. Here is what a teammate actually is. Each teammate gets its own role, its own thread, its own desktop and browser, and memory between conversations. Add one to a Slack channel, let it hand work to other teammates, and run several at once. It runs on Perplexity computer, the agent platform Perplexity announced earlier this year. So, the teammate does the work. It does not just chat about it. The brain inside is Claude Opus 5.5, Anthropic flagship model, and that is the default for every teammate. But, you can pick a different model for each teammate or for each task. That matters. A research teammate and a coding teammate may need different brains. Availability and price. Teammates is a research preview on the web only for now. It is for individual Max subscribers. Mobile and desktop apps are coming soon. Enterprise Pro and Enterprise Max are excluded by default. Max costs $200 a month or $167 a month billed annually, which is $2,000 a year. That includes 10,000 computer credits a month and teammate use is free for a promotional period. Then teammates consume the same credits. 100 credits equals $1. What do those credits actually buy? Perplexity own credit guide dated the 4th of November says a multi-step research report costs 350 to 950 credits and a data dashboard costs 2,400 to 9,800 credits. So that 10,000 credit monthly allowance is roughly one dashboard or 10 to 20 research reports a month. Do the math before you sign up. The free period ends and Perplexity has not said when. Now the claims framed exactly as Perplexity claims them. Perplexity says its own workspace used about 10,000 teammates in 30 days and its Slack integration produced $1.6 million of work for 300 employees in 4 weeks. Here is the honest caveat. As of the 6th of November, there is no named customer case with measurable results. Internal numbers are a demo, not proof. Watch for a named company reporting real savings. The context matters because Perplexity is the third player in this race, not the first. Open AI launched dots on the 29th of September. and launched Claude Duet on the 21st of October. This channel covered both. Now Perplexity joins them. Do not confuse this with the Perplexity deep research comparison this channel covered on the 22nd of October. That was search. Teammates owns ongoing work. Story two. And let me be precise about what this is not. This channel already covered the Muse agent rollout on the 12th of October, the wonder models on the 13th, Spark open on the 16th, and the ghost shares pause on the 21st. This is none of those. On the 5th of November, Meta announced real-time voice and expressive avatars for Muse. The assistant now talks like a phone call, and it has a face that reacts with expressions and gestures matching the tone. There are three modes: voice only, an avatar beside the conversation, or full-screen face-to-face. You start from the phone icon at the top right of the Muse app, and you can interrupt mid-answer, like a real call. It is rolling out in the Muse app in the US on mobile, web, and Mac. Meta previewed this at Connect 2026 in September. Now it is shipping. Meta published its own measurements, so here they are as Meta reports them. Portrait video at 448 by 768 at 25 frames per second. About 870 milliseconds from the end of your turn to the first bite of video. 12 concurrent real-time sessions on a single GB200 chip. And every video carries an invisible Meta video seal watermark. The default character is nicknamed Jolly, and you can customize it into a person, an animal, or something else. What is not here yet, voice on Meta AI glasses and on Muse charm, the pocket device, is coming later in 2026 with no date. Muse is free with a usage limit. Power is $20 a month, maximum 100 app-wide. Voice is US only at launch. Muse is also in Canada, Japan, and Singapore since late October with no date for voice there. Muse itself launched on the 8th of September 2026. Story three from Nvidia on the 5th of November. Cosmo 3 Nano for open weight checkpoints, each with 16 billion trainable parameters, post-trained from Cosmo 3 Nano. Announced on the Hugging Face blog and Nvidia developer forums. This is the open weights driving and robotic side of interactive world models, unlike Google Genie 4, covered here on the 28th of October as a closed research preview. The four checkpoints each do a different job. Transfer auto takes multi-view driving data, up to 11 cameras including fish eye, plus joint lidar, and changes weather, lighting, and time of day. Sim auto streams synchronized camera and lidar video as driving controls change. Two tools for self-driving. Generate the scenarios you cannot safely film in real life. The other two are for simulation and robots. Sim transfer takes physics simulator controls plus a text prompt and turns them into photo-realistic video. Sim bi-manual predicts video from robot or camera actions on Yam and Aggie Bot dual arm robots. Nvidia owned numbers, Sim by manual runs at 48.1 frames per second with 166 milliseconds of latency at 480p on 1GB300 chip. Vendor numbers, not independent tests, but fast enough to steer in real time. And the license is the real story for builders. These checkpoints ship under the Open M DW 1.1 license, which permits commercial and non-commercial use. Anyone can build on them. Anyone can ship a product with them. The opposite of a closed preview. Open world models you can download. Before the takeaway, one watch line, and to be clear, neither of these is announced. First, Qwen 4. Alibaba posted a teaser on the 6th of November quoting the philosopher Zhuangzi, calling it imminent. Four models are expected. Max, plus, flash, and a 27 billion parameter model arriving one after another through November and December. But there are no weights, no model card, no date, and no benchmarks yet. Do not call it launched. Second GPT-6.2 Astra and Soul. The Friday the 6th of November window passed with no Open AI announcement at all. Cyber shipped instead. Not launched. So, why do these three stories belong together? They are three layers of the same shift. Perplexity gives agents ongoing jobs. Meta gives them a voice and a face. Nvidia gives them worlds to practice in before they touch the real world. The race is no longer about who has the smartest model. It is about whose agent is always on. What should you do? Three checks. Perplexity Max subscribers. Try one teammate on a real recurring job and watch the credit burn before the free period ends. In the US, open Muse. Tap the phone icon and test whether interrupting works like a real call. Builders. Download the Cosmos checkpoints on hugging face and test sim transfer with your own simulator. Everyone else. Wait for a named customer result before paying $200 a month for teammates. That is the morning briefing. Perplexity teammates. Always-on AI agents on a $200 a month max plan with 10,000 credits included. Meta Muse. Real-time voice and expressive avatars. US only. Rolling out now. Nvidia Cosmos 3 Nano. Four open weight world models under a license that lets you ship. Three launches. One morning. One race. Tell me in the comments. Which one changes your work first? An AI co-worker that works while you sleep? An AI you can talk to face-to-face? Or robots trained in fake worlds before they touch the real one? I read every comment and I will see you in the next video.

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, Assistente pessoal, Dots, Muse. Publicado no YouTube em 2026-11-07.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.

Comentários

Postagens mais visitadas deste blog

How to build a JARVIS AI Voice Agent for FREE | Full Livekit Tutorial (2026)

How Does AI Actually Work? 🤖 | AI, Machine Learning & Generative AI Explained