Happy Monday!
Before this week's stories, I need 3 minutes.
For three years I've ended every issue the same way. Play with AI or get passed by someone who did.
Here's my confession. I've never given you a place to actually do that.
I'm fixing it. I'm building something new this fall. That's all I'm saying for now.
But before I lock anything in, I want your fingerprints on it. Ten quick questions, almost all one tap, about three minutes total. What you'd actually show up for. What's really blocking you. What it's worth.
Your answers decide what gets built. The people who answer hear about it first.
Alright, now this week in AI...
Enjoy
PROMPT: Before you answer, solve, create, recommend, or take any action, interview me first.
Ask me one question at a time, choosing each next question strategically based on my previous answer.
Your goal is to eliminate assumptions and reach at least 95% confidence that you understand:
What I actually want
The outcome Iām trying to achieve
My priorities and preferences
Any important constraints or requirements
What a successful result should look like
Do not ask a long list of questions at once. Ask only the single most useful next question.
Continue until you believe you have at least 95% confidence that you understand the assignment.
Once you reach that threshold:
Stop asking questions.
Briefly summarize your understanding of what I want.
Explain what information gave you enough confidence to proceed.
Tell me what you intend to do next.
Do not actually do the work yet.
Wait for me to explicitly tell you to proceed.
Only after I give you that signal should you produce the final answer or take the requested action.
In Todayās Edition:
š¤Æ ChatGPT can now send texts for you with new Apple Messages plug-in
š ļø 1 new AI tool that you should make time to try this week
š® Google is buying all of Spirit Airlinesā data to feed its AI models
š¤ Connect Gmail, Calendar, & Drive to Claude.
Read time: 4 minutes
šØ AI ART
Back to school
š° AI INDUSTRY NEWS
Setup asks for a permission far broader than messages, and OpenAI tells you not to use one of its own settings.
What lands on the disk afterward is readable by anything else running under the same login.
A fifth of layoff announcements now name AI, and EY thinks most of them are posturing.
The product now interrupts a teen mid-shortcut, which assumes the model can tell cheating from confusion.
Posters will now be told when readers suspect them, which is a new kind of performance review.
Researchers counted Oxford commas as evidence, which should make a few careful writers nervous.
Free on every plan, though you still need to bring your own subscription to the agents.
Deidentified, but the contract requires the links between records survive, and the flight attendants' union noticed.
The new AI chief does not report to the CEO, and the role arrived paired with a second promotion.
Connecting it to your ad accounts and Workspace hands Meta context it has never had before.
š AI TOOL TO PLAY WITH THIS WEEK
š Claude Google Workspace Connectors ā Connect Gmail, Calendar, & Drive to Claude.
š LEARN
Copilot Chat, Copilot Notebooks or Copilot Agents? Hereās What to Use & Whenā¦
ChatGPT Plugins Finally Work!
Claude Cowork Tutorial for Beginners
š¹ TIKTOK OF THE WEEK
š¤Æ What Can Claude Do That ChatGPT Canāt?
@willfrancis24 What can Claude do that ChatGPT can't? Here are 5 features that are worth knowing about when comparing the two leading AI tools: Usage lim... See more
āļø THE EXTRAS
One year ago today, I wrote MIT Report Reveals AI Failures
The most clicked link in last weekās newsletter: Anthropic says it will watermark text generated by its AI models
Have a burning AI question?
I read every single question submitted and use it to improve this newsletter. Submit your question below!
Thatās it for this time! See you next week.
ā Cliff Worley