Good Morning! Here's what I have for you in today's newsletter:
Google adds Deep Research to Gemini Live, start reports by voice
ElevenLabs MCP generates voice, music, and video inside your assistant
Perplexity Portable Computer runs agents locally on Windows RTX PCs
Build a real second brain with GPT-6 Astra's long context
4 new AI tools worth trying today
AI TOOLS
Google connected Deep Research to Gemini Live, bringing its multi-step research report tool into the Gemini app's voice mode as part of a wider set of updates rolling out to all Gemini users.
Users can ask Gemini Live to run a Deep Research report on any topic, and Gemini continues the research in the background even after the chat is closed or the screen is locked.
A notification lands when the report is ready, and the findings stay in context as you switch between voice and text for follow-ups.
Free users can run reports with the Thinking model, while Google AI Pro and Ultra subscribers get higher daily limits and the option to generate with Pro.
Research becomes something you can kick off on a walk or between meetings. Say the topic, put the phone away, and come back to a cited report you can interrogate by voice. For anyone who thinks out loud better than they type, this makes Deep Research usable away from the desk.
AI AGENTS
ElevenLabs expanded its MCP with voice, music, image, and video generation, so a single install inside ChatGPT, Claude, Cursor, Grok Bot, or Hermes gives your assistant the full ElevenCreative stack.
Text to Speech generates a voiceover in any library voice directly inside the conversation, while Scribe converts recordings into clean transcripts for scripts, captions, or dubbing.
More than 50 image and video models are available through the same connector, allowing an image to be generated, edited, animated into video, and lip-synced within a single thread.
Installation requires one OAuth sign-in from the connectors directory, with no server to run or API keys to manage, and all output is saved to the ElevenCreative workspace for editing in Studio.
A voiceover, a music bed, a dubbed version, and a short video can all come out of the chat window where you wrote the script. You brief the assistant once and it picks the models. The output opens in Studio for the fine cuts, so the creative work stays in your hands.
AI MODELS
Perplexity released Portable Computer for Windows, running its agent harness, orchestrator, scheduler, and a local model on NVIDIA GeForce RTX and RTX PRO GPUs with 24GB or more of VRAM.
Pick a local model such as PPLX 27B or Qwen 3.8 27B from the dropdown and it downloads in one click, post-trained for Computer and tuned for RTX.
The agent works across local files, connected apps like Gmail, Slack, and GitHub, and the web, with sensitive data staying on device and no Computer credits spent on local runs.
When a task needs heavier reasoning, it asks permission before escalating to a frontier cloud model, and a built-in scheduler runs recurring jobs while you are away.
If you own a 24GB RTX card, you get an agent that reads your files, drives your apps, and runs scheduled tasks without a per-step cloud bill or sending documents off your machine. It is available to Pro and Max subscribers through the Windows app in the Microsoft Store.
HOW TO AI
Capture everything, organize nothing by hand, retrieve it all later.
A second brain only works if capturing something is easier than remembering it yourself, and finding it later is easier than digging through old notes. ChatGPT Projects, paired with GPT-6 Astra's long context, fixes the retrieval half specifically. Everything lives in one workspace, and Astra can read across all of it at once.
Step 1: Set up
Open ChatGPT, click Projects, click New. Two real decisions to make right here. First, project-only memory, which can only be set at creation and can't be reversed.
Should I use project-only memory here? I want this project's context to stay isolated from my other chats and not leak into unrelated conversations.Second, custom instructions, rules you write yourself that never change on their own.
Set these as this project's custom instructions: always cite which uploaded file a claim came from, keep responses under 300 words unless I ask for more detail, and never invent a source that isn't actually in this project.Step 2: Capture
Drag in whatever you already have, without sorting first.
Here are notes from the last three weeks, unsorted, just read them for now, don't organize anything yetStep 3: Organize
This is the step people skip, and the one that makes retrieval genuinely work later.
Go through everything I've uploaded to this project and organize it into clear categories. Tell me what the categories are and which files fall into each. Flag anything that doesn't fit cleanly anywhere.
Step 4: Distill and retrieve
A distilled brief is shorter than a summary and sharper.
Take everything in the [real category name] group and distill it into a one-page brief: the actual decisions made, what's still open, and the one thing that would change if a key assumption turned out wrong.The real payoff is pulling from this months later without re-reading everything yourself.
Using everything in this project, draft [a real piece of writing you'd actually produce] based on what's genuinely established here, not generic advice
Step 5: Audit
Memory outside the project is separate and worth checking on a real schedule.
What do you remember about me?Deleting a chat doesn't remove a saved memory that came from it. Correct it in Manage Memories directly.
P.S. You can access all the AI trainings (including the full version of this one), prompts and workflows if you upgrade.
Meta released Muse Voice Transcribe as a plugin in LiveKit Agents, giving developers its streaming speech-to-text model with real-time diarization for 20+ speakers, so voice agents can tell who said what during a live call.
Nous Research opened Hermes Business accounts in Nous Portal, letting you invite colleagues so the team runs agents across channels on one shared balance with per-member caps and shared skills, with Hermes Enterprise available for on-prem or your own cloud.
Cline launched Cline Desktop, a native app for working with open-weight models, paired with ClinePass and free models like DeepSeek-V4.1-Flash and Musespark-1.3, or bring your own key to any provider.
⚡ Bolt: Forge adds GLM, DeepSeek, and Kimi to the model picker with up to 50x more usage and zero usage charges, free until mid-October.
🎬 Motion: the new Motion MCP lets Codex take your assets and build scenes, animations, and transitions directly inside Motion for launch videos.
🔄 Linear: Loops can trigger from project, initiative, cycle, and issue changes, edit documents automatically, and send tailored Slack updates.
🚀 Raycast: 2.0 is out of beta on Mac and Windows with Screen Awareness, Memory, built-in dictation, Automations, and a bring-your-own ChatGPT or Claude account.
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial