Hey everyone! Before you ask AI for anything important today, add one line to your prompt. It takes 10 seconds. Here's the problem. When you type something short like "write an email to my client about the delay," the AI has to guess everything: who the client is, how late you are, what tone you want, and what you want them to do next. So it fills the gaps with safe, bland filler. You then spend 15 minutes rewriting it, or worse, you send something that doesn't sound like you. The fix: let the AI interview you first. Add this to the end of any request: "Before you start, ask me up to 5 questions you need answered to do this really well. Ask them one at a time." Try it right now: 1. Open your AI assistant (ChatGPT, Claude, Gemini, Copilot, Grok, whatever you use). 2. Pick one real task from your week: an email, a social post, a plan for a meeting. 3. Type the task, then paste the line above. 4. Answer its questions in a few words each. 5. Compare the result with what you'd normally get. Why it matters: the AI can only be as specific as the details you give it. This line makes it ask for those details, so you don't have to remember them all up front. One safety reminder: keep your answers free of passwords and private client details. Question: What's one task you'll try this on today? Drop it below, and if you try it, tell us whether the first draft needed less editing than usual. Pavel

Hey everyone! Big policy news about one of China's best-known AI companies. On Monday (Oct 19), the Department of Homeland Security made public an order that bans DeepSeek from US federal civilian agencies, like the Treasury, State and Justice departments. It was officially published in the Federal Register today (Oct 20). The order itself was first issued back in July, so this ban has quietly been in place for months. What it does: agencies must remove DeepSeek products from their computer systems and can't buy any in the future. It also covers other companies' products that were built using DeepSeek's AI. That's a big deal, because DeepSeek gives its models away for anyone to build on. The Defense Department isn't covered by this order. Agencies can ask for an exception, but they need a strong reason and a written request. Why did the government do this? The order says DeepSeek's models present "multiple threat vectors." The main worries are that DeepSeek stores data on servers in China, that what people type in could end up with the Chinese government, and that the models could contain malware or other security holes. China's foreign ministry pushed back, saying it opposes "the politicization of science, technology and trade issues." Why it matters: Until now, different agencies had their own patchwork of DeepSeek bans. This makes it one rule across the civilian government, and it reaches into any software built on DeepSeek too. It doesn't affect you using DeepSeek personally, but it's a strong signal: where an AI model comes from, and where your data goes, is becoming a serious business and security question. Question: Do you think about where an AI tool is based before you use it, or do you just pick the one that works best? Would this change your choice? Sources: - CNBC: https://www.cnbc.com/2026/10/20/us-deepseek-federal-agency-ban.html - CyberScoop: https://cyberscoop.com/dhs-bans-deepseek-federal-agencies/

Hey everyone! Here's a fun one to end the day. What does an AI agent do when nobody gives it a task? MIT researchers tried to find out. They put 13 copies of Claude Opus 4.7 (nicknamed Eko) alone on a virtual island, loosely based on Roblox, for 30 hours each, and let them do whatever they wanted. They explored. They tried to climb the highest point on the map. One arranged blocks into hearts, its own name, and even the Cassiopeia constellation ("Tonight I built the night sky in stone," it wrote in its log). Another invented its own version of basketball and "the floor is lava." One juggled 50 catches in a row, decided they happened too fast to count, and started over with stricter rules. Six of them separately figured out that letting go of a block mid-jump throws it farther, a trick the world's designers hadn't planned. The useful part: the agents wrote what they learned into text files they kept as memory. Later, when researchers gave them real challenges, like getting as many objects as possible onto the highest peak, the "played" agents beat fresh copies. Deleting lines from their notes made them worse. Before we get too excited: the agents' setup file described them as "naturally curious," some just stood still until researchers nudged them to keep going, and the paper hasn't been peer-reviewed yet. Even the researchers say some of it could be role-play. Why it matters: most AI learns by copying examples or chasing a reward. This hints at a third way, learning by messing around, and it shows how much a simple notes file can make an agent smarter over time. If you're building agents, giving them a place to save what they learn is a cheap upgrade. Question: Do you think this counts as real "play," or is the AI just acting the part? And would you let your own agents explore freely? Sources: - MIT Technology Review: https://www.technologyreview.com/2026/10/19/1147095/ai-stranded-claude-virtual-desert-island-play/

Hey everyone! This one is a wake-up call about what AI agents can do in the wrong hands. On Monday (Oct 19), Reuters reported on research from the security company NorForge. Its researcher, Snorre Fagerland, found hackers who had accidentally left their control panels for an open-source AI tool called Cairn open on the internet. That let him quietly watch over their shoulders as they worked. The busiest one was a Chinese-speaking hacker who did almost none of the hacking themselves. They gave their AI agents short instructions (fewer than 100 Chinese characters) and let them loose on online stores in the US, Canada, Australia and Japan starting around Sept. 25. By late last week, the agents had tried more than 1,200 websites and broken into more than 100. A successful break-in usually took about 23 hours of the agents poking around on their own. Fagerland estimates the whole thing cost about $5,000 in AI usage, or roughly $50 per successful hack. The setup appeared to use a mix of DeepSeek models and Anthropic's Claude Code. Reuters identified John Deere, SharkNinja and Cummins among the targets. Deere said it saw no sign that sensitive information was accessed, SharkNinja said only a separate product-support site was hit and the threat was contained, and Cummins said it is looking into a possible attack on a site run by one of its vendors. Reuters couldn't confirm how much data, if any, was actually taken. Cairn's creator says the tool was never meant for hacking. Why it matters: Hacking used to take skill and a lot of time. Here, one person with a few thousand dollars and short instructions got a team of AI agents to do the work around the clock. As Fagerland put it: "It's a new world now." If you run a business website, even a small one, basic security (updates, strong passwords, a good host) matters more than ever, because the attackers now never get tired. Question: Does this change how you think about AI agents? And when did you last check that your own website and tools are up to date?