In partnership with
🧠 THAT ONE AI - This Week’s Signal
Here’s what’s shaping AI right now - without the noise:
🗣️ GPT-Live-1.1 brings new voices to ChatGPT, with the API next
🚫 The US bans DeepSeek in federal agencies as V4.1-Flash ships
🔧 Perplexity Computer now rebuilds apps without their source code
🧠 Stop being the bottleneck for your agent
🧰 Tools worth testing
🗣️ GPT-Live-1.1 Brings New Voices to ChatGPT, With the API Next
OpenAI released GPT-Live-1.1, its voice model for ChatGPT Voice, Codex and dots. It adds new voices in French, German, Hindi, Mexican Spanish, Korean and Japanese. It also has better accents and fewer "checking..." pauses.
It's rolling out to all ChatGPT users now, and OpenAI says the API comes soon. See the announcement.
The bigger signal: 👉 Voice apps in those six languages get a better default voice soon. If you build a voice agent for those markets, plan your tests for the day the API opens.
285 Free Prompts to Save Time
Your inbox isn't the problem. Neither are your meetings. It's the hours you spend on work AI could do in minutes.
285 copy-paste prompts: built for email, meetings, admin, and delegation. Each one has a clear role, a specific task, and a defined output.
No setup. No prompt engineering. Just copy, paste, and get your time back. Free when you subscribe to Mindstream.
🚫 The US Bans DeepSeek in Federal Agencies as DeepSeek Ships a Faster V4.1-Flash
The Department of Homeland Security banned DeepSeek from federal civilian agencies. It's Washington's latest step to keep Chinese AI out of critical systems.
On the same weekend, DeepSeek updated V4.1-Flash. DeepSeek says the model now finishes tasks faster with shorter answers, and it opened its in-house web search tools to developers. Read about the ban.
The bigger signal: 👉 If you sell to government or regulated buyers, your model list is now part of the deal. Keep a non-Chinese fallback ready for every DeepSeek call in your stack.
🔧 Perplexity Computer Now Rebuilds Apps Without Their Source Code
Perplexity added REA, an open-source reverse-engineering skill, to Computer. You point Computer at an app. It studies how the app behaves, then builds a working version.
Perplexity showed five demo apps, including copies of Figma, Photopea, Audacity and Harvey's contract review. Tasks that use /reverse-engineer are free until October 31, up to 10,000 credits per user. See the demos.
The bigger signal: 👉 If your app's only defense is its features, an agent can now copy them in an afternoon. The data your users store in it, and the habits they've built around it, are what protect you.
🧠 That One AI Tip: Stop Being the Bottleneck for Your Agent
Anthropic's 80% number matches what builders on Reddit keep reporting. The agent works for two minutes, asks a small question, then waits twenty minutes for you to come back. Many of them also admit they approve almost every permission prompt without reading it.
So most of the waiting buys you nothing. Here's how to cut it.
Step 1. Answer the usual questions before the agent asks
Most interruptions are choices you've already made in your head. Write them down at the top of the task.
DECISIONS ALREADY MADE:
- Stack and style: follow the patterns in /src/api, no new libraries
- Test command: npm test (must pass before you report done)
- Naming: if you need a new file or function name, pick one and list it in your summary
- Small choices (copy text, field order, log messages): decide yourself and note itStep 2. Pre-approve the safe actions
Let the agent read files, search the code, run tests and edit inside the project folder without asking. Keep prompts for actions that are hard to undo, like deploys, deletes, payments and anything that sends data out. In Claude Code, the allow list goes in your project's .claude/settings.json:
json
{
"permissions": {
"allow": ["Bash(npm test:*)", "Bash(npm run lint:*)", "Edit(src/**)", "Read(**)"],
"deny": ["Bash(git push:*)", "Bash(rm -rf:*)"]
}
}When the agent only asks about risky steps, you'll actually read the prompts again.
Step 3. Have the agent queue questions instead of stopping
Paste this into your task, or into your CLAUDE.md:
If a question does not block the task, do not stop.
Write it in QUESTIONS.md with the assumption you chose, and keep working.
Stop and ask only if you cannot continue without my answer.Then you answer all the questions in one sitting, and the work keeps moving between your answers.
Step 4. Review at checkpoints, and count your interruptions
Ask for a short summary and a diff at each milestone, and review the batch. For one week, count how many times a session waited on you. If the number doesn't fall by half, add the questions that keep coming back to your Step 1 block.
The bigger signal: 👉 Models keep getting faster, but the human is still the slowest step in the loop. A clear task spec gets you more of your agent's time than a model upgrade does.
State of Product reveals what AI still hasn’t solved
80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.
That One AI 🧰 TOOLBOX
A few tools quietly worth exploring:
🖥️ Muse Spark Open → Meta's open-weight model for coding and long agent tasks, built to run on your own hardware.
🎞️ Popcraft → An MCP server that lets any agent make images, audio and video.
📬 Grok Bot → Gives an agent its own email address so it can contact businesses for you.
🧱 NitroStack → Build, test and host MCP servers and AI apps with TypeScript and Python SDKs.